id
stringlengths 33
33
| doc_url
stringlengths 33
33
| title
stringlengths 13
211
| publication_date
timestamp[s] | update_date
timestamp[s] | authors
stringlengths 6
16.6k
| category_primary
stringclasses 60
values | category_all
stringlengths 5
132
| abstract
stringlengths 123
1.99k
| text
stringlengths 190
2.08k
| embeddings
listlengths 768
768
|
|---|---|---|---|---|---|---|---|---|---|---|
http://arxiv.org/abs/2405.16153v3
|
http://arxiv.org/pdf/2405.16153v3
|
DefSent+: Improving sentence embeddings of language models by projecting definition sentences into a quasi-isotropic or isotropic vector space of unlimited dictionary entries
| 2024-05-25T09:43:38 | 2024-06-19T13:26:53 |
Xiaodong Liu
|
cs.CL
|
cs.CL, cs.AI, cs.LG
|
This paper presents a significant improvement on the previous conference
paper known as DefSent. The prior study seeks to improve sentence embeddings of
language models by projecting definition sentences into the vector space of
dictionary entries. We discover that this approach is not fully explored due to
the methodological limitation of using word embeddings of language models to
represent dictionary entries. This leads to two hindrances. First, dictionary
entries are constrained by the single-word vocabulary, and thus cannot be fully
exploited. Second, semantic representations of language models are known to be
anisotropic, but pre-processing word embeddings for DefSent is not allowed
because its weight is frozen during training and tied to the prediction layer.
In this paper, we propose a novel method to progressively build entry
embeddings not subject to the limitations. As a result, definition sentences
can be projected into a quasi-isotropic or isotropic vector space of unlimited
dictionary entries, so that sentence embeddings of noticeably better quality
are attainable. We abbreviate our approach as DefSent+ (a plus version of
DefSent), involving the following strengths: 1) the task performance on
measuring sentence similarities is significantly improved compared to DefSent;
2) when DefSent+ is used to further train data-augmented models like SIMCSE,
SNCSE, and SynCSE, state-of-the-art performance on measuring sentence
similarities can be achieved among the approaches without using manually
labeled datasets; 3) DefSent+ is also competitive in feature-based transfer for
NLP downstream tasks.
|
DefSent+: Improving sentence embeddings of language models by projecting definition sentences into a quasi-isotropic or isotropic vector space of unlimited dictionary entries: This paper presents a significant improvement on the previous conference
paper known as DefSent. The prior study seeks to improve sentence embeddings of
language models by projecting definition sentences into the vector space of
dictionary entries. We discover that this approach is not fully explored due to
the methodological limitation of using word embeddings of language models to
represent dictionary entries. This leads to two hindrances. First, dictionary
entries are constrained by the single-word vocabulary, and thus cannot be fully
exploited. Second, semantic representations of language models are known to be
anisotropic, but pre-processing word embeddings for DefSent is not allowed
because its weight is frozen during training and tied to the prediction layer.
In this paper, we propose a novel method to progressively build entry
embeddings not subject to the limitations. As a result, definition sentences
can be projected into a quasi-isotropic or isotropic vector space of unlimited
dictionary entries, so that sentence embeddings of noticeably better quality
are attainable. We abbreviate our approach as DefSent+ (a plus version of
DefSent), involving the following strengths: 1) the task performance on
measuring sentence similarities is significantly improved compared to DefSent;
2) when DefSent+ is used to further train data-augmented models like SIMCSE,
SNCSE, and SynCSE, state-of-the-art performance on measuring sentence
similarities can be achieved among the approaches without using manually
labeled datasets; 3) DefSent+ is also competitive in feature-based transfer for
NLP downstream tasks.
|
[
0.0000887383,
0.0707748383,
-0.0262543224,
0.0708708167,
-0.0622146986,
0.0043068361,
-0.0051897736,
-0.0391549319,
-0.0177006312,
-0.072623454,
0.0006134974,
-0.0187924225,
-0.0399480164,
0.0419883505,
0.0389121585,
-0.0457504988,
0.0469016545,
0.022465799,
-0.0356672779,
0.0154408347,
0.0255731419,
-0.0049913623,
0.0354004875,
0.0203659013,
-0.034681011,
-0.0458575226,
0.0003573468,
-0.0077519952,
0.0083637126,
-0.0379311815,
0.0353631377,
0.0291988719,
0.0561363883,
0.0027374967,
0.0000018795,
-0.0194475912,
-0.0233902223,
-0.0037218924,
0.0018191796,
-0.0026197792,
0.0275614727,
-0.0228838362,
0.0459774658,
-0.0157545339,
-0.0322727598,
-0.037859343,
0.0452896245,
0.0452926457,
0.0509705096,
0.0381896794,
-0.0348838121,
-0.0455825925,
0.0124562467,
-0.0183745082,
-0.0480228998,
0.0018005499,
0.0161722787,
0.0080860462,
-0.0302833132,
0.0358085483,
-0.0307101384,
0.0479559526,
0.0115409903,
0.0186145343,
0.0493358411,
0.0193498377,
0.0025681152,
0.0001004687,
-0.010282225,
0.0849053264,
-0.0143573098,
-0.0350282118,
-0.0189682804,
-0.0012102482,
-0.0316292681,
0.009823666,
-0.0489784963,
-0.0222750548,
0.0565399081,
0.0546336584,
0.0260074753,
0.0273651648,
0.0377859101,
-0.0012260029,
-0.0232033711,
0.0034191976,
0.0527451113,
-0.0617004521,
-0.0020596669,
0.0249332096,
-0.0454762429,
0.0078751193,
-0.0058910013,
-0.0206367057,
-0.0004665378,
0.0237179175,
-0.0399298966,
-0.0603117272,
-0.0036229359,
-0.0454125032,
0.07400015,
0.0325815529,
-0.006584167,
0.0270048231,
-0.043388918,
0.0475072712,
0.0060194922,
0.0327323452,
-0.0857327133,
-0.0274142828,
-0.0663516894,
-0.0304628592,
-0.0429474637,
0.0284068082,
-0.0175246876,
-0.0608566515,
-0.0088577475,
0.0213245135,
-0.0250408985,
0.0138948867,
-0.0125147589,
0.0041109445,
-0.0128039168,
0.043834541,
-0.0019418579,
0.0114585515,
-0.0315329097,
0.0121600451,
-0.0033509701,
-0.0259949937,
-0.0249415878,
0.0257276595,
0.0125364196,
-0.0286926851,
-0.0161285698,
0.0379172862,
0.0149846254,
-0.0633247569,
0.0087297326,
0.0086631831,
-0.0256083217,
-0.0327733345,
0.008189938,
-0.0680389255,
0.0334439613,
0.008627452,
-0.009804952,
-0.0728832558,
0.0014653599,
0.0210656524,
-0.0173470378,
0.1064379737,
-0.0059584309,
0.0124074919,
-0.0400017798,
0.0100309076,
0.0388016663,
0.0793718994,
-0.0028636821,
-0.0024823532,
0.055321712,
0.0164583791,
-0.0352876186,
0.0182103962,
0.0140709672,
0.0264778528,
-0.0915181413,
0.0289254598,
0.0291160326,
0.0382050239,
0.0236716475,
0.0520861,
-0.0193294547,
0.018134715,
0.0582479313,
0.060784366,
0.0118677542,
-0.0148483599,
0.0550836846,
-0.0001000607,
-0.0399206541,
-0.0089809988,
-0.0185756497,
-0.0241476446,
0.0682677701,
0.018876439,
0.0368393213,
0.0237823296,
-0.0402865745,
-0.0139822019,
-0.0459822379,
-0.0292476974,
0.0013325424,
0.0009013194,
0.0175560955,
0.0302462112,
-0.004307522,
0.1059442386,
0.0533888377,
-0.0525810458,
-0.0289502162,
0.0078689167,
-0.0031864394,
0.0370535664,
0.0533127636,
-0.0201185159,
-0.0249719173,
0.0051891068,
-0.0032320907,
0.0676082075,
0.0660805553,
-0.0187673606,
0.0067250668,
-0.0455418937,
-0.0406454094,
-0.033322826,
0.0418515876,
0.0035922343,
0.0213027783,
0.0088640973,
0.043477606,
-0.0486297011,
-0.0327347256,
0.0452867895,
-0.0389079079,
-0.0203019083,
-0.0092223855,
0.0254678819,
0.0099562509,
0.0167998783,
-0.0083315205,
0.0211878084,
0.0217718016,
-0.083521992,
-0.0254138596,
-0.024066668,
-0.0688269436,
-0.0788918734,
0.0425855145,
-0.0459867343,
0.0125399819,
0.0792155042,
-0.0395241156,
-0.0189944524,
-0.002924202,
-0.0280121807,
0.0402090624,
-0.0404834673,
-0.0138697457,
-0.0038522547,
0.0437718779,
-0.0413462333,
0.070881322,
0.0006590192,
0.0438489616,
0.0086583262,
0.0048681125,
0.0250750184,
0.0233388394,
-0.0375836045,
0.0112060057,
0.0026416681,
0.0311988592,
0.0076799365,
-0.0213465821,
-0.0154611683,
-0.0169314444,
-0.0449239016,
0.0156724304,
-0.0091429697,
-0.001085526,
0.0375504792,
0.0335934125,
0.0250342935,
0.0007559649,
-0.0290944893,
0.0131044714,
0.013605671,
0.0469262265,
0.0617474802,
-0.0315119401,
-0.038032759,
-0.0376232639,
-0.0112465415,
-0.0238452237,
0.077207081,
-0.0874998197,
-0.0038740924,
-0.0308579169,
0.0204693582,
-0.0330911316,
-0.0001632316,
0.0051864833,
-0.032844495,
-0.014313315,
-0.0065835183,
0.0355234593,
0.0228367839,
0.0029478413,
0.0228280965,
0.0554184876,
0.0066119055,
-0.011880382,
-0.0014154438,
-0.0337409452,
0.053543698,
-0.0116756288,
0.0094636874,
0.0174864791,
0.0461932234,
0.0079790466,
-0.0273276288,
-0.0352295116,
0.0236469693,
-0.0223656278,
0.0195294023,
-0.0287361126,
-0.0173908137,
-0.0255625136,
-0.011245125,
-0.0298554692,
-0.0266848318,
0.0479150563,
-0.0564050823,
0.0153450519,
-0.0398914777,
-0.0369957462,
-0.0350539722,
0.0768771768,
0.0352021456,
0.0180597454,
-0.0059796828,
0.0453224033,
-0.0508404896,
-0.0315960161,
0.0221027602,
-0.0039953073,
0.0540670529,
0.0182125159,
0.0000162002,
-0.005097975,
-0.0289819837,
-0.0468594432,
-0.0792901963,
-0.0388293266,
0.0409044921,
-0.0063165156,
0.0088770036,
0.0174076147,
0.0163749643,
-0.0167389382,
0.016656138,
-0.0012668651,
0.0059354752,
-0.0477010496,
0.047075633,
0.087884292,
0.0403692052,
0.0024238147,
-0.0101055941,
0.0542556942,
0.0032156566,
0.0135776335,
0.0202761032,
-0.0386915803,
0.0018826609,
0.0021533223,
-0.027747836,
-0.0295896512,
-0.001613216,
0.052307535,
0.0036016677,
-0.0482652672,
0.0992094353,
0.009439745,
-0.0170653425,
0.0248682071,
0.0105350632,
0.0551943108,
0.0343142152,
-0.01005546,
-0.0368855149,
-0.0688139871,
-0.0091096144,
-0.0102537321,
0.0368926451,
0.0112194261,
-0.0313367769,
-0.0830308944,
-0.0054321815,
-0.0045708977,
0.009763387,
-0.0424757376,
-0.0333790556,
0.0105462587,
-0.004903194,
0.0098299412,
-0.0533088483,
-0.0058343695,
0.0036073343,
-0.0143877082,
-0.0304391868,
-0.0410352834,
0.0144412294,
-0.0418683067,
0.0174478274,
0.0490820706,
0.0506597497,
0.0130798798,
-0.010101581,
0.0135229677,
-0.0882665813,
-0.0318030491,
0.0250987001,
0.0633249283,
-0.0245248713,
-0.0125838257,
0.0174251888,
-0.0148664471,
-0.045787815,
-0.0427461006,
-0.0230745319,
0.1175743192,
0.0171294492,
-0.0138963955,
-0.0350448154,
-0.0117155155,
0.0214172024,
0.0886259377,
0.0768473595,
-0.0039971853,
-0.0450567156,
0.0192841645,
0.0296846014,
-0.0020219535,
-0.0241865162,
0.0177931003,
-0.0017692985,
0.0068254601,
-0.067174837,
-0.0463369228,
-0.0170714222,
0.028569987,
0.0022384792,
0.0428300537,
-0.024001196,
-0.0285603944,
-0.053219974,
0.0069302614,
-0.0074587888,
0.0167582817,
0.0361188948,
0.0065712724,
0.0275621209,
0.0207940526,
-0.0422012135,
-0.0105398195,
-0.0747287124,
-0.0502959974,
-0.0102061341,
0.0010014619,
0.0001854782,
0.0112774651,
-0.0266039949,
0.0106597999,
0.0746759921,
0.0355230123,
-0.0116305863,
-0.0168579258,
0.0507153347,
-0.0338474102,
0.0199118443,
0.0763165206,
-0.0377377495,
0.0235032365,
0.0212230962,
0.0557545908,
0.0086798808,
-0.0836604983,
0.0168451481,
0.0084689166,
0.0713055134,
0.019033704,
-0.0182293672,
-0.0276085231,
-0.0136653055,
-0.0283678118,
-0.0326025076,
-0.0626837239,
0.0105081294,
0.0069977562,
0.0254428424,
-0.0233707651,
0.0216128398,
-0.005999113,
0.0238191448,
-0.0186594948,
0.009032649,
0.0250220541,
-0.0117038786,
-0.0242406633,
0.0146222226,
-0.0392132141,
-0.0181965232,
-0.0451297127,
-0.0271241274,
-0.0630635098,
-0.0028932521,
0.0256618559,
0.0080586653,
-0.0397977196,
0.0411327593,
-0.0803847462,
0.0042048036,
-0.0270534679,
-0.0221183896,
0.0628525093,
-0.0625011772,
0.046319183,
-0.0403915495,
-0.0141303958,
-0.0082649402,
0.059915591,
-0.0142009137,
-0.0290253758,
0.0387611724,
0.0194668528,
-0.0525337607,
-0.0417309515,
0.042991627,
-0.0079917936,
0.0427135117,
0.0096574835,
-0.0159832947,
-0.0340259708,
-0.0132824695,
0.0430684276,
0.025053706,
-0.0086546456,
0.0419086963,
-0.0169878956,
-0.0857284814,
0.0230692141,
-0.0103883017,
-0.054593619,
0.0120831495,
0.0299598388,
0.0291643236,
0.0238314699,
-0.0277234782,
-0.0207525231,
0.0233539473,
-0.0057003978,
-0.0517696142,
-0.0237888675,
-6.193551899e-33,
-0.0052004326,
0.0250514932,
-0.0274627302,
-0.0187842101,
-0.0772660598,
0.0197533015,
-0.0409538895,
-0.0333659314,
-0.0221077949,
0.0386795513,
-0.0096464446,
0.0164909326,
-0.0060267518,
0.0338065252,
-0.0099990964,
-0.0181275625,
0.0326553471,
0.001302281,
0.0043720254,
0.006967695,
0.0390386693,
0.0494323485,
0.0656260625,
-0.0249492936,
-0.0254507326,
-0.0130525995,
0.0429466255,
0.0256819408,
0.0645032749,
0.0407888219,
-0.0384850688,
0.0406769067,
0.0208859481,
-0.0404196307,
-0.0371869951,
0.0168491881,
-0.0328266844,
-0.0720235705,
-0.0016946504,
-0.0174947809,
-0.0085335486,
-0.0046438775,
0.0287698247,
-0.023269359,
0.0195027553,
-0.029101897,
0.0437710807,
-0.0189082325,
-0.0239191744,
0.0357879102,
-0.0769290775,
0.031105049,
0.0205802396,
-0.0050813435,
0.0377739295,
0.0357704423,
0.0128353378,
-0.0184146501,
-0.0041054529,
0.0039223293,
-0.0157394409,
0.0801182538,
0.0227903109,
-0.0067930217,
0.0135110253,
0.0766562521,
0.0114682019,
-0.0498637222,
-0.0118969008,
-0.0160934702,
0.0190640632,
0.0408689342,
0.0295749288,
0.084901765,
0.0410067625,
-0.0209559463,
-0.0429440252,
0.00623935,
-0.0191063061,
0.0123667158,
0.0444115512,
0.0010527745,
-0.0339029506,
-0.0455249697,
-0.0325382762,
0.0000742138,
-0.0072274674,
-0.0312431194,
0.0612938851,
-0.0114995763,
0.0081777452,
-0.0034715482,
-0.0013218408,
-0.0660225451,
-0.0048464802,
0.0477241315,
-0.0166859366,
0.0039453693,
0.0215926412,
0.0119945565,
-0.0462007597,
-0.0431020744,
-0.0332502089,
0.0186627563,
0.003633234,
0.008140265,
0.0619844049,
0.0094549451,
-0.0337033458,
-0.0069254604,
-0.0228703078,
-0.0676724017,
-0.074561514,
0.0173767488,
0.0286791231,
0.0361274369,
-0.0147875333,
0.0355635434,
0.0602616221,
-0.0137997949,
-0.0170557965,
-0.0336852595,
0.0280538835,
0.0564182177,
-0.0021860597,
-0.0144201722,
-0.0905256346,
0.0160328262,
0.0200803764,
-0.067295216,
-0.0026867236,
-0.0195364449,
2.606e-7,
0.0250252094,
0.0082093021,
-0.0093289493,
0.0314622521,
-0.0003337277,
0.0457513668,
-0.0152221955,
0.0642321259,
-0.0541195683,
-0.0032496275,
0.0324262194,
0.0312257856,
0.0562518537,
0.0329092592,
-0.0785169825,
0.0410958529,
-0.0736586601,
-0.0068356609,
-0.0426609293,
-0.0030023863,
0.0952040777,
0.0949673653,
0.0208387394,
-0.014527454,
-0.0516493246,
-0.0154683227,
0.0390885174,
-0.0262637958,
-0.0110019781,
-0.0063123861,
0.0126238428,
0.00330397,
0.0165560581,
0.0019786388,
-0.0276921019,
0.0615803599,
0.0324588306,
0.019190006,
0.0039175255,
-0.0137242908,
-0.0253599156,
-0.0044503664,
-0.0090956194,
-0.0316434093,
0.0693895146,
-0.0409646593,
-0.0105418721,
-0.0699509904,
-0.0234221928,
0.0296172742,
0.0413351879,
0.02395216,
-0.0282178912,
0.0066042626,
0.0523159876,
-0.0001613804,
-0.0326216593,
-0.0834772885,
-0.0290270783,
0.0289760344,
-0.0391212255,
-0.0096944589,
-0.0578925312,
-0.0021819919,
0.0829389021,
-0.039874211,
0.0167313088,
2.86502733e-34,
0.0016223028,
0.0180256963,
0.0261924583,
0.0086247185,
0.009993922,
-0.0496565886,
0.0432342067,
-0.0188845061,
0.0438435301,
-0.081734024,
0.0024275307
] |
http://arxiv.org/abs/2405.16150v1
|
http://arxiv.org/pdf/2405.16150v1
|
5W1H Extraction With Large Language Models
| 2024-05-25T09:42:58 | 2024-05-25T09:42:58 |
Yang Cao, Yangsong Lan, Feiyan Zhai, Piji Li
|
cs.CL
|
cs.CL
|
The extraction of essential news elements through the 5W1H framework
(\textit{What}, \textit{When}, \textit{Where}, \textit{Why}, \textit{Who}, and
\textit{How}) is critical for event extraction and text summarization. The
advent of Large language models (LLMs) such as ChatGPT presents an opportunity
to address language-related tasks through simple prompts without fine-tuning
models with much time. While ChatGPT has encountered challenges in processing
longer news texts and analyzing specific attributes in context, especially
answering questions about \textit{What}, \textit{Why}, and \textit{How}. The
effectiveness of extraction tasks is notably dependent on high-quality
human-annotated datasets. However, the absence of such datasets for the 5W1H
extraction increases the difficulty of fine-tuning strategies based on
open-source LLMs. To address these limitations, first, we annotate a
high-quality 5W1H dataset based on four typical news corpora
(\textit{CNN/DailyMail}, \textit{XSum}, \textit{NYT}, \textit{RA-MDS}); second,
we design several strategies from zero-shot/few-shot prompting to efficient
fine-tuning to conduct 5W1H aspects extraction from the original news
documents. The experimental results demonstrate that the performance of the
fine-tuned models on our labelled dataset is superior to the performance of
ChatGPT. Furthermore, we also explore the domain adaptation capability by
testing the source-domain (e.g. NYT) models on the target domain corpus (e.g.
CNN/DailyMail) for the task of 5W1H extraction.
|
5W1H Extraction With Large Language Models: The extraction of essential news elements through the 5W1H framework
(\textit{What}, \textit{When}, \textit{Where}, \textit{Why}, \textit{Who}, and
\textit{How}) is critical for event extraction and text summarization. The
advent of Large language models (LLMs) such as ChatGPT presents an opportunity
to address language-related tasks through simple prompts without fine-tuning
models with much time. While ChatGPT has encountered challenges in processing
longer news texts and analyzing specific attributes in context, especially
answering questions about \textit{What}, \textit{Why}, and \textit{How}. The
effectiveness of extraction tasks is notably dependent on high-quality
human-annotated datasets. However, the absence of such datasets for the 5W1H
extraction increases the difficulty of fine-tuning strategies based on
open-source LLMs. To address these limitations, first, we annotate a
high-quality 5W1H dataset based on four typical news corpora
(\textit{CNN/DailyMail}, \textit{XSum}, \textit{NYT}, \textit{RA-MDS}); second,
we design several strategies from zero-shot/few-shot prompting to efficient
fine-tuning to conduct 5W1H aspects extraction from the original news
documents. The experimental results demonstrate that the performance of the
fine-tuned models on our labelled dataset is superior to the performance of
ChatGPT. Furthermore, we also explore the domain adaptation capability by
testing the source-domain (e.g. NYT) models on the target domain corpus (e.g.
CNN/DailyMail) for the task of 5W1H extraction.
|
[
0.0637108237,
0.0247268025,
0.0340801291,
0.0363104604,
-0.0628276169,
0.0433380902,
-0.00048966,
-0.0038831022,
0.0297462679,
-0.0594472177,
-0.0106625631,
-0.0239448734,
-0.0305018369,
0.0142945908,
-0.0141126523,
-0.0678765997,
0.035017021,
-0.0242715292,
-0.0113155618,
0.0054595917,
0.0257572234,
0.0164745171,
0.0168501269,
-0.0011598801,
0.0204661749,
-0.0051986491,
-0.009491086,
-0.015461104,
0.0298291761,
0.0027352395,
0.0091658365,
0.057747066,
0.0351356007,
-0.011933431,
0.0000019501,
-0.0536643788,
-0.0328632034,
0.000719325,
-0.0032761213,
-0.0156242317,
0.0543696173,
-0.0402295776,
-0.012452499,
0.0067396369,
-0.0023134735,
-0.0250166971,
0.007036733,
0.0801847354,
0.0553539805,
0.1127822101,
-0.027785318,
-0.021358924,
0.0211022217,
-0.0272087101,
0.0082940096,
-0.0724995807,
0.0248814374,
-0.0108430376,
0.0236311052,
0.0584207512,
0.0328778513,
0.0184446946,
0.0395571962,
-0.0081260614,
-0.0586834624,
0.0142964665,
-0.0479176939,
0.0030652641,
-0.0056402329,
0.0556560419,
0.0166769587,
0.0026525152,
0.0464464538,
0.0166375544,
0.0264658853,
-0.001510585,
-0.0848988295,
0.0373449363,
-0.0315725915,
-0.0026229953,
0.0275744237,
0.0127751557,
-0.0391377434,
0.0555731319,
0.0190839265,
0.020926442,
0.0157155637,
0.0188633334,
-0.0190016404,
0.0455500074,
-0.0051483582,
-0.0553128123,
0.0154478168,
0.0310314279,
0.0243063997,
0.0010613016,
-0.0648690388,
0.0081568072,
0.0104903085,
-0.0761594996,
-0.0657718778,
0.0143144475,
-0.0442603566,
0.0823013857,
0.0020633354,
-0.005823256,
0.0035836133,
-0.007975515,
-0.0419515409,
-0.0131404884,
-0.054041516,
0.0229127221,
-0.0265713315,
0.0301431678,
-0.0760232732,
-0.0275203511,
-0.0105551174,
0.0217888337,
0.0479108319,
-0.0189929903,
0.0081899976,
-0.0262878444,
-0.032017231,
-0.0073833079,
0.0367007889,
0.0417921282,
-0.0530940443,
-0.0220734291,
0.0391327068,
-0.0471576825,
-0.042293679,
0.0288150106,
0.027013544,
-0.027998114,
-0.0013594858,
0.0295479298,
-0.0427974947,
-0.039570339,
-0.0361879915,
-0.017278621,
-0.0274698455,
-0.0376264527,
0.029999584,
-0.0540787913,
0.0227882471,
0.0044993185,
-0.0315357968,
0.0060115624,
-0.0013491978,
0.0263386015,
-0.0560114197,
0.0480170026,
0.0143871158,
-0.0033874915,
0.0600586832,
0.0126695381,
0.0203829929,
0.1046700627,
0.0083221653,
0.0673289597,
0.0630507395,
0.0149123902,
-0.0274322089,
0.0440991223,
0.0000787045,
0.0142324846,
0.0085486798,
0.0094792191,
-0.0328022577,
-0.010808846,
0.0590154827,
0.0367783457,
-0.0300331954,
0.0016894066,
0.0786727965,
0.0363910235,
0.0032105562,
-0.0038072234,
-0.0012165378,
0.022006141,
0.0412960947,
0.039368432,
-0.0312764794,
0.041879788,
-0.0502354056,
0.0209075995,
0.0095623257,
0.0329514965,
-0.0294173267,
-0.0213210005,
-0.008970052,
0.0041300361,
0.0079075629,
0.0223272499,
0.0719276816,
-0.0016236078,
-0.0225404557,
0.0276104845,
0.0476863943,
-0.0365634635,
0.0156106893,
-0.0464008562,
0.0265114475,
0.0145021286,
0.0358981304,
-0.017941311,
-0.0149205746,
-0.01160035,
0.0674690455,
0.0431087762,
0.0657597855,
-0.0289093982,
-0.010764068,
-0.0196540058,
-0.0288489331,
-0.0084216278,
0.00031971,
-0.0032035629,
0.0277041327,
0.0042645382,
0.0116916206,
0.0261524208,
-0.0367149711,
0.0302551202,
-0.0113701522,
-0.0405838266,
-0.0177678447,
-0.0075682453,
0.0041237264,
0.0378735289,
-0.0100078927,
-0.0145429503,
0.003468225,
-0.0344322287,
-0.0138491495,
-0.0151225757,
-0.0455084257,
-0.0457391031,
0.0189327542,
0.0163614992,
0.0128544914,
0.018508831,
-0.0431795679,
-0.0360975862,
-0.0220742822,
0.0060801622,
-0.0152675984,
-0.0630377829,
-0.0122216437,
0.0305380039,
0.0121712703,
-0.0539809391,
0.0877023041,
0.0232107136,
-0.0196863934,
-0.0088368086,
-0.0146587323,
-0.0281673409,
-0.02304543,
-0.0263022371,
0.0300102234,
-0.0153064346,
0.0087316614,
0.034108825,
0.0125337318,
-0.050937362,
-0.0253802333,
-0.1027261466,
0.0374675468,
-0.0420146286,
0.0379391275,
0.029134253,
0.0425352305,
-0.0158382468,
-0.0073313499,
-0.0360148922,
-0.0440606698,
-0.0179033708,
0.0495525263,
0.0242566671,
0.0476259477,
0.046380125,
-0.040581964,
-0.0044480362,
-0.0130303316,
0.038036678,
0.0092495279,
-0.048854202,
-0.0203496702,
0.0404218286,
0.0005828401,
-0.0059874495,
-0.0091334879,
-0.0210370161,
-0.0353489034,
-0.0132172853,
0.0296729151,
0.0684353933,
0.0344608203,
0.063226074,
-0.0220771171,
-0.0128002763,
-0.0359369591,
0.021297602,
-0.0533873886,
-0.0140640475,
0.0205522664,
-0.0018412793,
0.0287525859,
0.0292764883,
0.0082955994,
0.0422165692,
-0.0218400899,
0.0425632,
-0.0177048072,
0.020915024,
-0.0296144653,
0.0019633963,
-0.025538655,
0.0373916663,
-0.0418668948,
0.0223218892,
0.073647812,
-0.0073002181,
0.0079356516,
-0.0672496706,
-0.0230784286,
-0.0593604632,
0.0526836775,
-0.0249029398,
0.0103578083,
0.0326214507,
-0.0355536826,
-0.0179667808,
-0.0269523636,
0.0463611633,
-0.0318716317,
0.0536425374,
-0.0182992425,
0.0261569247,
-0.0370669886,
-0.0348844565,
-0.0239454117,
-0.0631777123,
-0.0115780504,
-0.0113025717,
-0.0244053788,
0.0072348523,
-0.0161552709,
0.0501961894,
0.0312561728,
0.074706085,
-0.0794280544,
0.0099156136,
0.0066366768,
0.0975771695,
0.0984759331,
0.0081372662,
-0.029455414,
-0.0272427276,
0.0406899415,
0.0386325605,
0.0117862271,
0.015452357,
0.0176809598,
0.0095153283,
0.0226557683,
-0.0583553724,
0.0026641565,
-0.0272302236,
0.0744271129,
-0.0646979138,
-0.0080635706,
0.1105249673,
0.0231869593,
-0.039987646,
-0.016159175,
-0.0170712434,
0.0062906807,
-0.009699625,
0.0330703855,
-0.067294091,
-0.0458277762,
0.0160397422,
0.0107746152,
-0.0020212277,
-0.0239507575,
-0.0023782451,
-0.09240105,
0.046629902,
0.002451241,
-0.0485463515,
-0.0084323008,
-0.0545820184,
0.0266244747,
-0.0382003374,
0.0155818686,
0.0172474664,
-0.0105854394,
-0.0054245065,
0.0163438376,
-0.0352275781,
-0.0615529828,
0.0242450424,
0.0065329326,
-0.0007917628,
0.0398723669,
-0.0070497743,
-0.0171532482,
0.0091866991,
0.0551071987,
-0.0244664475,
-0.0102828844,
-0.0212359298,
0.1160274893,
-0.0267269406,
-0.0175308436,
-0.0509612635,
0.0611722805,
0.0238321088,
-0.0336249806,
-0.011486032,
0.0321779288,
0.0067928866,
-0.0045936895,
-0.0224050395,
0.0105356406,
0.0025379229,
0.0989889279,
0.0362120159,
-0.0463818535,
0.0049002692,
-0.0021971031,
-0.0223440342,
-0.0506141484,
-0.0590356737,
-0.0043266509,
-0.0425356105,
-0.0730263293,
-0.0779205561,
-0.0487843491,
0.0307337046,
0.0229387097,
-0.0281737093,
-0.0187648684,
0.0081881825,
0.0444884151,
-0.0385533869,
0.0448024981,
-0.0250000954,
-0.0255365372,
-0.0249780733,
0.0123546692,
0.076375328,
-0.0142950891,
-0.0340268165,
0.0586791225,
-0.0747783631,
-0.0275040641,
-0.0079510268,
-0.0139521537,
0.0219891537,
0.0275979396,
-0.0150961215,
-0.0331246257,
-0.0027011354,
0.0421208926,
-0.0061668432,
-0.0249959771,
-0.0024446603,
0.0012341016,
-0.0161162522,
0.0199624337,
-0.0627958253,
0.0551490225,
0.0012072267,
0.03599957,
-0.0317560583,
-0.032836251,
-0.0037257315,
0.0368065499,
0.0259296838,
-0.0006846454,
-0.0142114498,
-0.031045353,
-0.0440742895,
-0.0261050817,
-0.0126081184,
-0.0400384404,
0.0398568474,
-0.0145268943,
-0.0131066963,
-0.014989607,
0.0396797918,
-0.0009242592,
0.039155446,
-0.0860807672,
-0.0079208789,
-0.0265632998,
0.0098848036,
-0.046030473,
-0.0055131018,
-0.0441212058,
-0.013102321,
0.0129755661,
0.0086068744,
-0.0378848426,
0.0539321676,
0.0337500013,
-0.003587076,
0.0137564195,
0.0207070746,
-0.0850834399,
0.0032521766,
-0.067751497,
0.0587635413,
0.0459354222,
-0.054361999,
0.0540409982,
-0.0130763939,
-0.0479471348,
-0.0125623355,
0.0100403531,
0.0304141697,
0.0056591248,
0.0210191831,
0.0393280312,
-0.0315968655,
0.0113984067,
0.068817094,
-0.0314820409,
0.035819523,
0.0676392317,
-0.0351631157,
-0.0498681068,
-0.0194625892,
0.0291636866,
0.001799181,
-0.0140222544,
-0.0080954144,
-0.0524655581,
-0.0157672483,
0.0224524084,
-0.017424006,
-0.0410661548,
-0.0007973082,
0.0195900425,
0.0589658916,
0.0195055082,
-0.0186608601,
-0.0459344871,
0.0257113632,
0.0064783371,
-0.0356902517,
0.035588704,
-6.650967608e-33,
-0.0513272323,
0.0230002888,
-0.0135363461,
0.0560308285,
-0.028636381,
0.017271731,
-0.045998279,
-0.0039808699,
-0.0253496785,
-0.0044427738,
-0.0203000009,
-0.0237774216,
0.0316941328,
0.0186109506,
-0.0109630069,
-0.0169862416,
0.0210845731,
-0.0334207155,
-0.0370507762,
0.0349462405,
0.0445519537,
0.0209624749,
0.0677003786,
-0.0279153194,
0.0386940539,
-0.0054983064,
-0.0168951061,
0.0188861173,
0.0976828858,
0.035889443,
-0.0093447296,
0.0134855863,
0.0446379744,
-0.089886114,
-0.0162375495,
0.0406943746,
-0.1268600672,
-0.0325316973,
0.0564189255,
-0.0067228912,
-0.011171679,
-0.0327888355,
0.0266842358,
-0.0068309479,
0.0319029428,
-0.0380222574,
0.0328106023,
-0.0498036779,
-0.0073673497,
-0.0370160267,
-0.0508827381,
0.0447867103,
0.0018567854,
-0.0339939408,
0.0238315109,
0.0273539089,
0.0005240236,
0.0603538901,
-0.0319944248,
-0.0024165597,
-0.0066091232,
0.046516601,
0.0447603352,
-0.0124729751,
-0.0202801358,
0.0692921802,
-0.0028119071,
-0.0057995184,
-0.0343302786,
-0.0057464913,
0.0261900108,
0.0331505463,
0.0055216895,
0.0544444993,
0.0206724945,
-0.0073656333,
-0.0690887943,
0.0614301264,
0.030922344,
0.0010503205,
-0.0151682245,
-0.0008590943,
-0.0408692546,
0.0048098196,
0.0256906245,
-0.004103078,
-0.0333000198,
-0.0228361208,
0.0138252005,
0.0043269186,
-0.0201546904,
0.0247012433,
0.0600432381,
-0.0042482298,
0.0396259427,
0.0382801108,
-0.0376860686,
0.0319113135,
-0.0132252583,
0.0086864829,
-0.0428489037,
0.0030362534,
0.007687421,
0.0159813277,
-0.0203336291,
0.0058495724,
0.0586729459,
0.0057328274,
-0.0530179106,
-0.0042490484,
-0.0068553858,
0.0306191556,
-0.0450823195,
0.0556915142,
-0.0032937592,
-0.009554035,
-0.0070847524,
0.047467526,
0.0250868201,
-0.018428592,
-0.0058876788,
-0.0188423954,
0.0216605365,
0.0114856623,
-0.0024797684,
0.013486878,
0.0185002759,
0.0342758857,
0.0066123325,
-0.0215778463,
-0.0099821799,
0.0143316956,
2.696e-7,
0.0256395191,
0.0333310552,
0.0319658145,
0.0234438386,
-0.0180588774,
0.0380692147,
0.006436687,
0.0345021002,
0.0139058949,
-0.0134919425,
0.0069376733,
0.029916307,
0.0602691807,
0.0126464143,
-0.083162792,
-0.0283876173,
-0.0712703317,
-0.0242169406,
-0.0407468751,
-0.0175712276,
0.0557243824,
0.0674858838,
0.0131698512,
-0.0088661341,
-0.0371452682,
-0.0480576679,
0.0365176164,
-0.0147120645,
-0.0970874354,
0.038473092,
0.0187927634,
0.0010140826,
-0.0193441454,
-0.0154899051,
-0.0077453237,
0.0028670768,
0.053205017,
0.081320703,
0.0119485008,
-0.0092493389,
-0.0188410748,
-0.0240343828,
0.0321579576,
-0.0307270121,
0.0853734836,
-0.0704080909,
-0.0275515039,
-0.0468103886,
-0.0396665223,
-0.0064876247,
-0.0404394791,
-0.0096892659,
-0.0389550067,
0.0590208359,
0.0040574456,
-0.0126958825,
-0.014525095,
-0.0773882493,
-0.0236483514,
0.0296340603,
-0.0221918542,
-0.0447926559,
-0.0167646278,
-0.0261756703,
0.0048594014,
-0.0695421323,
0.0033355481,
2.710304494e-34,
0.0431213193,
0.022864284,
-0.044716768,
0.0425640084,
-0.0041497424,
0.0159794446,
-0.0148113761,
-0.0140915001,
-0.0105705671,
-0.0649888217,
0.0090161627
] |
http://arxiv.org/abs/2405.16136v1
|
http://arxiv.org/pdf/2405.16136v1
|
C3LLM: Conditional Multimodal Content Generation Using Large Language Models
| 2024-05-25T09:10:12 | 2024-05-25T09:10:12 |
Zixuan Wang, Qinkai Duan, Yu-Wing Tai, Chi-Keung Tang
|
cs.AI
|
cs.AI, cs.CL, cs.LG, cs.SD, eess.AS
|
We introduce C3LLM (Conditioned-on-Three-Modalities Large Language Models), a
novel framework combining three tasks of video-to-audio, audio-to-text, and
text-to-audio together. C3LLM adapts the Large Language Model (LLM) structure
as a bridge for aligning different modalities, synthesizing the given
conditional information, and making multimodal generation in a discrete manner.
Our contributions are as follows. First, we adapt a hierarchical structure for
audio generation tasks with pre-trained audio codebooks. Specifically, we train
the LLM to generate audio semantic tokens from the given conditions, and
further use a non-autoregressive transformer to generate different levels of
acoustic tokens in layers to better enhance the fidelity of the generated
audio. Second, based on the intuition that LLMs were originally designed for
discrete tasks with the next-word prediction method, we use the discrete
representation for audio generation and compress their semantic meanings into
acoustic tokens, similar to adding "acoustic vocabulary" to LLM. Third, our
method combines the previous tasks of audio understanding, video-to-audio
generation, and text-to-audio generation together into one unified model,
providing more versatility in an end-to-end fashion. Our C3LLM achieves
improved results through various automated evaluation metrics, providing better
semantic alignment compared to previous methods.
|
C3LLM: Conditional Multimodal Content Generation Using Large Language Models: We introduce C3LLM (Conditioned-on-Three-Modalities Large Language Models), a
novel framework combining three tasks of video-to-audio, audio-to-text, and
text-to-audio together. C3LLM adapts the Large Language Model (LLM) structure
as a bridge for aligning different modalities, synthesizing the given
conditional information, and making multimodal generation in a discrete manner.
Our contributions are as follows. First, we adapt a hierarchical structure for
audio generation tasks with pre-trained audio codebooks. Specifically, we train
the LLM to generate audio semantic tokens from the given conditions, and
further use a non-autoregressive transformer to generate different levels of
acoustic tokens in layers to better enhance the fidelity of the generated
audio. Second, based on the intuition that LLMs were originally designed for
discrete tasks with the next-word prediction method, we use the discrete
representation for audio generation and compress their semantic meanings into
acoustic tokens, similar to adding "acoustic vocabulary" to LLM. Third, our
method combines the previous tasks of audio understanding, video-to-audio
generation, and text-to-audio generation together into one unified model,
providing more versatility in an end-to-end fashion. Our C3LLM achieves
improved results through various automated evaluation metrics, providing better
semantic alignment compared to previous methods.
|
[
-0.0038135392,
0.0417484306,
0.0192774944,
0.0019496101,
-0.0136510888,
-0.0273765754,
0.0371893495,
-0.0138400346,
-0.0818322748,
-0.0557917953,
-0.0421831422,
-0.0332967974,
0.005641873,
0.0387183353,
0.035600055,
-0.0853415951,
0.0049481993,
-0.0464296825,
-0.0257329997,
0.003735794,
0.0054161125,
0.0281201415,
-0.0201227814,
0.0172203165,
-0.0653975606,
-0.0140675632,
-0.0252185278,
0.0294120852,
0.0186489299,
-0.0481672212,
-0.0418074988,
0.0100924466,
0.0678371713,
-0.0185334105,
0.0000017998,
-0.0440465957,
-0.0218470097,
0.013589099,
0.0544757843,
-0.0112469317,
0.0121656265,
-0.0375268832,
-0.0170845482,
0.019575445,
-0.0144130094,
0.0089404918,
0.0391646922,
0.0225911848,
0.0576642416,
0.0471374094,
-0.0364960358,
-0.0677097961,
-0.0280634686,
-0.0255709514,
0.0458504483,
-0.004899241,
-0.0255932361,
0.0309784971,
0.0231609289,
0.0049901395,
-0.027428396,
0.0346300192,
0.0096029481,
-0.0073443926,
0.0326865315,
0.0587672144,
0.0118180821,
-0.0609747954,
-0.0273298565,
0.0575659946,
-0.0059594191,
0.018847663,
-0.0006938933,
0.0201177876,
-0.0005532571,
-0.032041274,
-0.0354133286,
0.0565039478,
0.003455132,
0.0176780745,
-0.0316669866,
0.0353100486,
0.033009287,
-0.028507648,
-0.0005297759,
-0.0164822098,
0.0432127491,
-0.0046396521,
-0.0327350199,
-0.0286304485,
-0.0154340323,
-0.0352563486,
0.0446428657,
0.0410843268,
-0.0154278111,
-0.0090095596,
-0.0822524503,
-0.0050896481,
-0.0192403235,
-0.0006013248,
0.0445242301,
0.0424525365,
-0.0536110103,
0.0521493591,
-0.0511281639,
0.0150858322,
-0.0256994255,
-0.0279850755,
-0.0358811617,
-0.0137061952,
-0.0122757619,
-0.0082282163,
-0.0395673178,
0.0116722081,
-0.0525756218,
-0.0575922504,
-0.040872518,
0.0163977798,
-0.0155248921,
0.0353365429,
-0.0019904098,
0.0200288147,
-0.0537347905,
-0.0267354678,
0.0209510252,
0.0104579227,
-0.0030164155,
0.0781022906,
0.0140295755,
-0.0103215957,
-0.0018121918,
0.036248669,
0.0222809557,
-0.0126527883,
0.0320977755,
0.1037318483,
-0.0401156209,
0.0159834698,
-0.0195306558,
-0.0588493682,
-0.0340091735,
-0.0372007787,
-0.0003009026,
-0.0592614599,
0.0301859565,
-0.0156827811,
-0.0281976275,
-0.0004461612,
0.0173712038,
-0.0180051643,
0.0270226784,
0.05498413,
-0.0043609794,
0.0642699078,
0.0426786765,
0.0109404605,
-0.0000392732,
0.0305977799,
0.0829404444,
0.0362020917,
-0.00222859,
-0.0066913804,
-0.0185610242,
0.0644995719,
-0.0055198823,
0.0047615641,
-0.0624281093,
-0.0369116925,
0.0104837827,
0.0686949864,
0.0535398386,
0.0828431323,
-0.0098159388,
-0.0040967627,
0.0724886805,
-0.0108534573,
0.0241267662,
0.0014887573,
0.0907917544,
-0.0037037923,
-0.0403798558,
0.0281930454,
0.0186430998,
0.0633443594,
0.0005174605,
0.0324479714,
0.0674241185,
0.0422301441,
-0.0288316719,
-0.0190104507,
0.0112808561,
-0.0160059985,
-0.0015221359,
-0.0165781192,
0.0243526362,
0.0438762568,
-0.0761728138,
0.0402581766,
-0.0310472026,
-0.043149896,
0.0752309337,
-0.003350609,
0.0395183116,
0.0675467625,
0.0124102086,
0.0080896942,
0.0095084589,
-0.0024978656,
0.0081961816,
0.0768264607,
-0.0130041521,
-0.0182331055,
0.0313867182,
-0.070317857,
-0.0180406272,
-0.0097513609,
-0.0083362414,
0.0321879759,
-0.0060260356,
0.0245816484,
0.0198443495,
-0.014150002,
0.0457492135,
0.020692097,
0.0091246301,
-0.0011862522,
-0.0002702905,
0.0071438751,
0.0371497869,
0.0139153805,
-0.0360676236,
0.0153780449,
0.0091811782,
-0.0483260415,
0.0172852445,
0.030450426,
-0.0781948492,
-0.0141729778,
0.0288842078,
-0.0476496182,
0.0374759436,
-0.0033433172,
-0.0059125326,
-0.0494294651,
0.025938401,
-0.0099109281,
0.0462264903,
-0.0275218915,
0.0058676968,
0.0664871335,
-0.0000579981,
0.0129935015,
-0.0040912721,
0.001746416,
-0.0309525728,
0.0098904641,
0.0556459464,
-0.010463004,
-0.0162975676,
-0.0123696262,
0.0221124869,
-0.0124301296,
-0.0167960599,
-0.0063298829,
0.005282836,
-0.049267441,
-0.0556047633,
-0.0913314223,
-0.0007070569,
0.0068802973,
-0.023398282,
-0.0043598707,
0.007493278,
0.0157821383,
0.0046050735,
-0.0385064259,
-0.0044756075,
0.0416329205,
-0.0105809392,
-0.0313856639,
-0.0098129753,
-0.0323324539,
-0.0143496897,
0.006431004,
0.0273017045,
0.0285279136,
0.0231625866,
0.0080506457,
-0.0304248352,
-0.0137460083,
-0.0444046594,
0.0210574456,
-0.0117951976,
-0.0151716415,
0.0041457717,
-0.0037266407,
0.0829397738,
-0.0029731728,
0.0314057395,
0.0119714616,
0.0417126194,
-0.0243671648,
-0.0193135943,
0.0055858968,
0.0124281915,
-0.0014036725,
-0.0471104719,
-0.0144083938,
-0.0208329,
0.0170362256,
0.0873611271,
0.0463191047,
-0.0485643446,
-0.0076119187,
0.0171997845,
-0.0489671975,
-0.0038152374,
-0.0455530547,
-0.0313854292,
0.0235034674,
-0.0173829459,
-0.0394904427,
0.0280609373,
-0.0882971883,
-0.0358279757,
-0.0137813669,
-0.0156184593,
-0.0134539651,
0.0599842146,
-0.0139230005,
0.0572516136,
-0.0078520579,
0.0019830971,
-0.0608270839,
0.0055550495,
0.0187585205,
-0.0468269847,
0.0117410189,
-0.0303678364,
-0.008496739,
0.0018696926,
-0.0079955896,
-0.0346426852,
-0.0202968102,
0.000492863,
0.0075768381,
0.001904624,
-0.0013989882,
-0.0373649709,
0.0406918786,
0.0355756283,
0.0353686213,
0.000142118,
-0.0057626115,
-0.027012039,
0.0863592476,
0.0470513403,
0.0263565462,
-0.0576938279,
-0.0644518957,
0.0436145626,
0.0121652633,
0.0225035194,
0.0073177279,
-0.0541788004,
0.0358352065,
0.0306959879,
0.0356658101,
-0.0192721114,
0.0506969765,
0.0283828266,
-0.0556671359,
0.0246192645,
0.1005063802,
-0.0032419516,
-0.0211382397,
0.0184559654,
0.0161441136,
0.0117221233,
0.0058802017,
-0.018353723,
-0.0805502683,
-0.0361777507,
0.0159427896,
-0.0032405015,
0.0413257554,
0.0407747105,
-0.0347765759,
-0.0677873269,
0.083446078,
0.041031573,
-0.0298174284,
-0.0166392308,
-0.0051706359,
-0.0081042331,
0.0464801379,
-0.0018507598,
0.0217823274,
-0.0206948388,
0.003966901,
0.025475651,
0.0186373517,
-0.049367018,
-0.0123041151,
-0.045464851,
0.067728892,
0.0694053546,
0.0406933688,
-0.0277986713,
-0.0335053392,
0.0057617077,
0.0327271074,
0.0131911067,
0.0082948403,
-0.0122438734,
-0.0167212822,
-0.0033555119,
0.0127688795,
0.0652533174,
0.0412388444,
-0.0178429354,
0.0212385654,
0.038182687,
-0.0390679762,
-0.0042376737,
-0.0527419858,
0.0860829949,
0.0090071652,
0.0721518621,
-0.001870875,
-0.0456539914,
0.0262245853,
0.0315598845,
-0.0002254422,
-0.0739615485,
0.022231739,
0.0060219574,
-0.0195420943,
-0.0028050027,
-0.04635977,
0.012423302,
0.0049774842,
0.0514409803,
0.0158675592,
0.0170799028,
-0.0044233412,
0.0037736217,
-0.0759997889,
0.0264048111,
0.0208845288,
0.0176000725,
-0.0387151986,
-0.0303081032,
0.05502747,
-0.080088146,
0.0010819252,
0.0187379606,
-0.0645609871,
-0.0358954966,
-0.0374257267,
0.0110367574,
0.0153917475,
0.0756044909,
0.0399536267,
-0.0022960736,
0.0406474769,
0.008066711,
-0.0342177041,
0.0519943796,
0.0001897304,
0.0282526184,
-0.0143350363,
-0.0008919086,
-0.0693389103,
0.0363337621,
-0.0200924445,
0.0666733533,
0.0262832772,
-0.0267778728,
0.0147614414,
-0.0382009447,
-0.0073878146,
-0.0741894171,
0.0381629467,
-0.0624317825,
-0.0318236463,
-0.0166252796,
-0.0125657376,
-0.0801882669,
0.0167652648,
0.0476082601,
0.0373694673,
-0.0037260132,
-0.0080940034,
0.0064213858,
0.0058535025,
-0.0072053871,
0.0275559183,
-0.0219013412,
-0.0614840984,
0.0341298245,
-0.0146553861,
0.017420942,
0.0397724099,
-0.0378588028,
0.0001130352,
-0.0000029752,
0.0805871487,
-0.0044103139,
-0.0126726031,
-0.0132051976,
0.0253798906,
0.0202423744,
0.0292635206,
-0.0394568928,
0.0331751704,
0.0381970331,
0.0103034079,
0.0450389497,
-0.0254488252,
0.0022860009,
-0.0232941564,
0.0620231964,
-0.0248270333,
-0.0640440807,
0.0385292061,
-0.0150837591,
-0.0075504915,
-0.0030397894,
0.047323551,
-0.0244765859,
-0.000956413,
-0.0310942922,
0.0240002349,
-0.0339425169,
-0.0334989838,
0.0063737389,
0.0112813059,
-0.0085069565,
0.0063739917,
-0.0152022475,
-0.0573926903,
0.0106231133,
0.0110127795,
-0.005216328,
-0.0017745805,
0.0508621,
0.0431700051,
-0.0194908325,
0.0104329661,
-0.0082743876,
-0.0460652336,
-0.0159575026,
-0.0405667983,
-0.0094145965,
-5.275733257e-33,
-0.0570276044,
-0.047712367,
0.0347903073,
-0.0297013875,
-0.0477380231,
-0.0111449128,
-0.0317366421,
-0.0009639392,
0.040073432,
-0.0123123741,
0.0454219654,
-0.0176988039,
0.0012385061,
0.0598064177,
0.007420558,
0.0229803324,
0.0739767104,
-0.0345395356,
-0.0343644023,
0.0320642069,
0.026597837,
0.0224904865,
-0.0074267765,
-0.0392263159,
-0.0372893326,
-0.0113524925,
-0.0686655864,
-0.0125133274,
0.0110940142,
0.0456454642,
-0.0097753191,
-0.0033351697,
-0.0041334918,
-0.0049643097,
-0.0006668686,
0.0105127348,
-0.033299055,
0.0339258164,
0.0286430046,
0.0252310596,
-0.0522201285,
-0.0222731903,
0.0381173417,
-0.0313131735,
-0.0240509566,
-0.0318279862,
0.0480845273,
-0.0301692951,
0.0041126078,
-0.0957067758,
-0.067883119,
0.0169838127,
-0.0291816443,
-0.0618955866,
0.0094080241,
0.0650463328,
0.0144415563,
-0.0169384461,
-0.033688996,
0.0593534894,
-0.0293250214,
0.0138590317,
0.07784722,
0.0523530692,
-0.0195905231,
0.0561084263,
0.0299978517,
-0.0914300904,
-0.0210583396,
0.0300585888,
0.0017083747,
0.0492106564,
-0.0087251244,
0.0335584879,
-0.0044239908,
0.008725646,
-0.0147859408,
0.0377532579,
-0.0620833226,
-0.0141024031,
0.0458698049,
0.041472733,
0.0089847641,
0.0232152287,
-0.0555229113,
0.0039926944,
-0.0267107207,
-0.0416943282,
0.0361262299,
0.0020911412,
0.0181909818,
-0.029467538,
0.0884428322,
0.0182090066,
0.0170702673,
0.0162260663,
-0.0555656701,
0.0377047658,
-0.0196448918,
0.0170539431,
0.0104389172,
-0.0181971733,
0.1160782278,
-0.0281834304,
0.015514628,
0.0373907313,
-0.0263192691,
0.0867916793,
-0.0319115892,
-0.0087593263,
-0.009943,
0.0056003388,
0.0085072694,
0.0129189612,
0.0287637096,
-0.0058162888,
-0.0682595372,
0.0005148441,
-0.0166043006,
0.0095851701,
-0.0307241362,
-0.0456933863,
-0.0415470079,
0.0257479642,
-0.0552865304,
0.0615772679,
-0.0381030999,
-0.0621378087,
0.0390474796,
-0.0382397026,
-0.0349344909,
0.0556450114,
2.673e-7,
-0.0369136035,
-0.0115919942,
-0.0181649942,
-0.0139122168,
0.0379374139,
0.0174378157,
-0.0326673016,
0.0279012993,
0.0140355434,
-0.0129668852,
0.0312880389,
0.0136166112,
0.0575106889,
0.0163088702,
-0.0469487645,
0.0287060868,
-0.0655523762,
-0.0007926484,
-0.0493883379,
0.0232664235,
0.0618205182,
0.0791515708,
-0.0159853715,
-0.0433389395,
0.0413689874,
-0.0232481007,
0.0714888126,
-0.0070317481,
-0.0355970338,
-0.0342222676,
0.0179030523,
-0.0037473126,
-0.030198928,
-0.0126098385,
-0.0174805447,
-0.0142330704,
0.0526196584,
0.0596445799,
0.0131225092,
-0.042900648,
-0.018823361,
0.0543316081,
-0.0354491174,
-0.0144150602,
0.033606723,
-0.0015651956,
-0.027595818,
-0.0256914254,
-0.041982621,
0.012485452,
-0.0604794957,
-0.0507682003,
-0.0270917118,
-0.0114338491,
0.0409571864,
0.0196416229,
-0.0355761088,
-0.0398155414,
-0.0362101756,
0.0050935862,
-0.0779882446,
-0.0274133515,
-0.0364121683,
-0.0243818052,
0.0087789726,
-0.0454263687,
-0.044381842,
2.035470176e-34,
0.0244429912,
-0.0569942035,
0.000006398,
-0.0520687625,
0.0179350618,
0.0293814559,
0.0013848314,
0.0291445851,
-0.0179710314,
-0.0531926751,
-0.0467023514
] |
http://arxiv.org/abs/2405.16129v1
|
http://arxiv.org/pdf/2405.16129v1
|
iREL at SemEval-2024 Task 9: Improving Conventional Prompting Methods for Brain Teasers
| 2024-05-25T08:50:51 | 2024-05-25T08:50:51 |
Harshit Gupta, Manav Chaudhary, Tathagata Raha, Shivansh Subramanian, Vasudeva Varma
|
cs.CL
|
cs.CL
|
This paper describes our approach for SemEval-2024 Task 9: BRAINTEASER: A
Novel Task Defying Common Sense. The BRAINTEASER task comprises multiple-choice
Question Answering designed to evaluate the models' lateral thinking
capabilities. It consists of Sentence Puzzle and Word Puzzle subtasks that
require models to defy default common-sense associations and exhibit
unconventional thinking. We propose a unique strategy to improve the
performance of pre-trained language models, notably the Gemini 1.0 Pro Model,
in both subtasks. We employ static and dynamic few-shot prompting techniques
and introduce a model-generated reasoning strategy that utilizes the LLM's
reasoning capabilities to improve performance. Our approach demonstrated
significant improvements, showing that it performed better than the baseline
models by a considerable margin but fell short of performing as well as the
human annotators, thus highlighting the efficacy of the proposed strategies.
|
iREL at SemEval-2024 Task 9: Improving Conventional Prompting Methods for Brain Teasers: This paper describes our approach for SemEval-2024 Task 9: BRAINTEASER: A
Novel Task Defying Common Sense. The BRAINTEASER task comprises multiple-choice
Question Answering designed to evaluate the models' lateral thinking
capabilities. It consists of Sentence Puzzle and Word Puzzle subtasks that
require models to defy default common-sense associations and exhibit
unconventional thinking. We propose a unique strategy to improve the
performance of pre-trained language models, notably the Gemini 1.0 Pro Model,
in both subtasks. We employ static and dynamic few-shot prompting techniques
and introduce a model-generated reasoning strategy that utilizes the LLM's
reasoning capabilities to improve performance. Our approach demonstrated
significant improvements, showing that it performed better than the baseline
models by a considerable margin but fell short of performing as well as the
human annotators, thus highlighting the efficacy of the proposed strategies.
|
[
0.04232857,
0.030540837,
-0.0030431531,
0.0342414901,
-0.0675622746,
0.0363836214,
0.0236265175,
0.0014618789,
-0.0174020436,
-0.0557657592,
-0.0205589812,
-0.0128243305,
-0.0304006301,
0.0334287919,
0.0248292703,
-0.0248454101,
0.0595841631,
-0.0725193173,
0.0241495315,
-0.0163470581,
-0.029816065,
-0.002514642,
-0.0392205566,
0.0145339742,
0.0160881616,
-0.0496913195,
-0.0291046463,
-0.0003835153,
-0.0012145925,
-0.0486240312,
-0.0124618299,
0.0751892254,
0.0155785391,
0.0170286596,
0.0000021048,
-0.0584180057,
0.0030497464,
-0.0199577082,
-0.0613932088,
-0.0262525026,
-0.012585504,
0.0483370721,
0.0016226351,
0.0150980651,
-0.0551155098,
0.0161300842,
0.0145371929,
0.0704066381,
0.0677912161,
0.0704023242,
0.0215541534,
-0.0150189418,
0.0120424274,
-0.0160240084,
0.0743225813,
-0.0409538932,
0.0150730927,
-0.0083743883,
0.0126755461,
0.0276826173,
-0.0346905552,
0.0140913446,
0.0666734278,
-0.0083043706,
-0.0268953517,
0.030682873,
-0.046869088,
-0.0225241594,
-0.0413430482,
0.0777156651,
0.0334330313,
-0.0178117435,
-0.0370072126,
0.0186829139,
-0.0157159455,
0.0170725174,
-0.0519139655,
0.0303043146,
0.0253922176,
-0.0009507976,
0.0331568271,
0.0180308726,
0.0132202897,
0.0216029845,
-0.017994795,
0.1180025712,
0.0185648203,
-0.0331902616,
0.0298764333,
0.0015827671,
-0.0277251881,
-0.0455155671,
-0.0145019433,
0.0511320904,
-0.0338711329,
0.0203786306,
-0.0037534037,
-0.0236575492,
0.0111632533,
0.0062338156,
0.0271678828,
0.0548810735,
-0.0364343859,
0.073681578,
-0.0327350125,
0.0493774563,
-0.015743209,
0.0575030558,
-0.0300185923,
-0.0021808383,
-0.016849976,
-0.0141183818,
-0.0351897143,
0.0255698673,
-0.0373701379,
-0.0332010984,
-0.0286298748,
0.0199197996,
0.0046787192,
0.0037642487,
0.0196339674,
0.0444022119,
-0.0001038552,
-0.0140623646,
-0.0565802082,
0.0012014742,
-0.0503693521,
0.0004334866,
0.0299388766,
-0.0464969166,
-0.0523293763,
0.0251203701,
-0.016125001,
-0.0600834936,
0.0292953085,
0.0563349053,
-0.0188809149,
-0.028189702,
-0.0563160628,
-0.0318399854,
0.0020612162,
-0.0790830478,
0.0418981761,
-0.0577984229,
0.0125591233,
-0.0054653962,
-0.0118153989,
-0.018068986,
0.011208334,
0.0336307511,
-0.0309994631,
0.0868711248,
-0.0363887548,
-0.0065207668,
0.0123494146,
0.0246006995,
-0.0200008545,
0.0860324949,
0.0040970352,
0.005507939,
0.0302353632,
0.0009509507,
0.0099839913,
0.0340508223,
0.0289227068,
-0.0182813052,
-0.0403128453,
0.0350262858,
0.0349693373,
0.0302023888,
0.0334354751,
0.0224074069,
-0.0145450588,
-0.0018737679,
0.0350079127,
0.0800108314,
-0.052831687,
-0.0220043454,
0.0596197434,
-0.0443534851,
0.0112171993,
0.0284218043,
-0.0682901591,
0.0664911419,
-0.0565665774,
-0.00088063,
0.0104173571,
0.0073720771,
-0.0389846675,
-0.0300920494,
-0.0122868773,
0.0240039118,
0.0224408191,
0.0529874004,
0.0422015712,
0.0348279402,
-0.0458921082,
0.0314735994,
0.0081593543,
-0.0269995369,
-0.0226561688,
-0.0668264776,
0.0462961718,
0.051705569,
0.0574567728,
0.0066260798,
-0.0563351586,
-0.0051526246,
0.0485750809,
0.0666120201,
0.024125237,
0.0024439576,
0.014293164,
-0.0340388082,
-0.0271285828,
-0.0534750111,
0.0045681028,
0.0057820273,
0.0187255237,
0.0025647192,
-0.0089317132,
-0.041378174,
0.0249473695,
0.05043412,
-0.0362893306,
-0.0204629749,
-0.0146841574,
0.0113147628,
-0.006001472,
-0.0125497309,
0.0154123418,
0.0198259763,
0.0400695242,
-0.0271438695,
0.0081073297,
0.0046522887,
-0.0481891818,
-0.083703883,
0.0289855544,
-0.0180058945,
-0.0231275056,
0.0433959253,
-0.0393582098,
-0.0384917296,
0.0282512065,
-0.0139144752,
0.0214955546,
-0.0559302941,
0.0183779076,
0.0150412172,
-0.0023841099,
-0.039749749,
0.0839835778,
-0.0252190921,
0.0149049051,
0.0143857589,
0.0211234801,
0.0139659867,
-0.0718757883,
-0.0187028628,
0.0035839279,
0.0043208273,
-0.0393754654,
-0.0064671673,
0.0404966064,
-0.0726330951,
-0.0191706866,
-0.0499254167,
-0.0063171368,
-0.0324773043,
-0.0110865673,
0.0058168448,
0.0047624912,
-0.0166458301,
-0.0430102572,
-0.042642206,
-0.0021460569,
0.0277303755,
0.0269423947,
0.0432427116,
-0.0039806343,
-0.0122272093,
-0.0003764437,
-0.0039762151,
-0.024464028,
0.0530030094,
0.0006797221,
0.0371100269,
-0.0217656381,
0.044236578,
-0.0323530696,
0.0020740465,
-0.0068112272,
-0.0099144038,
0.048524376,
-0.0016459125,
0.0188666638,
0.0948741287,
0.0028168648,
0.0135792522,
0.0317834988,
-0.0146540124,
-0.0449628532,
0.045420967,
0.0012563707,
0.0041725994,
-0.0145738861,
-0.0300496481,
-0.0073640938,
0.0561889373,
-0.0009785744,
-0.0020391636,
-0.0501990654,
0.0856109113,
0.0315152556,
-0.0564320683,
-0.0526475534,
-0.0015305246,
-0.0160950143,
0.0667228922,
-0.0414186157,
-0.0241798088,
0.0643179342,
-0.0333340913,
-0.014783822,
-0.0477250032,
-0.0049338541,
-0.0490075946,
0.0686969757,
0.0020002343,
-0.0030352329,
0.0245721247,
-0.0055012018,
-0.069473438,
-0.0121745253,
-0.006474657,
-0.0043526865,
0.0261577591,
-0.0148341954,
0.0011933154,
0.0208397545,
-0.0209462475,
-0.0877596661,
-0.025431646,
-0.0373510197,
0.0189322177,
0.0241543353,
0.011920874,
-0.0343139507,
-0.0088431314,
0.0143555515,
0.0107760718,
-0.016244527,
-0.0154963201,
-0.01550953,
0.0362760499,
0.0652270168,
0.0595343634,
0.0360382013,
-0.0165544339,
0.03884352,
0.0495208949,
0.0153371561,
-0.0355196185,
-0.0142195355,
0.062744379,
0.0195400957,
-0.0366040468,
0.001496535,
0.025777109,
0.0484725833,
-0.0593105778,
-0.0409222022,
0.1121046394,
0.0589640066,
-0.0260546561,
-0.0075466451,
-0.0194704439,
0.0337421,
0.0392770655,
0.0387439728,
-0.0527630448,
-0.0645951405,
0.0107531706,
-0.0076571046,
-0.0158828106,
-0.0337629095,
-0.0520495288,
-0.1270044148,
0.0358547047,
-0.0245463904,
-0.0444316752,
0.0501135662,
-0.0183561202,
-0.0154589359,
-0.0052742222,
0.0217853207,
0.0159023982,
-0.0006764977,
0.0066938256,
0.0057995329,
0.0078455107,
-0.0329923481,
0.0298331399,
-0.0331556872,
0.0562221408,
0.0786756128,
-0.0060603572,
-0.0060744644,
-0.0794632584,
0.0191433169,
0.004922518,
0.0511754714,
-0.0369587652,
0.0451742075,
-0.0140659958,
-0.0123515148,
-0.0200489275,
0.0200208984,
-0.0328120552,
-0.0736338496,
0.0135538392,
0.082330443,
-0.0498805083,
-0.0068049855,
-0.0322319046,
0.0465826914,
-0.0229998939,
0.1015928835,
0.0311556403,
-0.0225855317,
-0.0449830629,
-0.0371197909,
-0.0157054532,
-0.0570304766,
-0.0074962517,
-0.007701532,
0.0392865539,
0.0518653691,
-0.0480793826,
-0.0337408856,
0.0538570248,
0.0612290502,
-0.0175607521,
0.0128020132,
-0.0800050721,
0.0356997512,
-0.0390987843,
0.03660376,
-0.0186363291,
-0.0031343652,
-0.045331303,
-0.003610851,
0.0025099716,
-0.0443809181,
-0.027536789,
0.0289911684,
-0.0897023752,
-0.0431565642,
-0.0058765952,
0.0534090512,
0.0279745497,
0.0255003087,
-0.0075036497,
-0.0380248465,
0.024365766,
0.0183675457,
-0.0207295828,
0.0157015529,
-0.0006748352,
0.0174017604,
0.0041894973,
0.031510368,
-0.0214527939,
-0.007994988,
0.0601343066,
0.0783042088,
-0.0230722334,
-0.0009785007,
0.0499088392,
0.0311950166,
0.0388201103,
-0.0124796964,
0.012911451,
-0.0137906885,
0.0326968394,
-0.0154133989,
-0.0291516371,
-0.0348255299,
-0.0019701824,
-0.0169980582,
-0.0249735005,
-0.0139370346,
0.00291853,
-0.0315730087,
0.0339447446,
-0.0072037238,
-0.0081748692,
0.0297433902,
-0.0019885604,
-0.0330753364,
0.0174708944,
-0.0079953019,
0.0391139239,
-0.0064355568,
0.0675370544,
0.0130104907,
0.0900905281,
0.024959743,
-0.0186507516,
-0.0301468596,
0.0219453964,
-0.0852855742,
-0.0511181504,
0.0113356747,
0.0728389472,
0.0485521182,
0.0203164276,
-0.0134163825,
-0.0128122699,
-0.0032627105,
0.0241338052,
-0.003040886,
0.0264576804,
-0.0362166241,
0.0696435571,
0.0284847468,
0.0007145788,
-0.0229102559,
0.0476138443,
-0.0107794311,
0.0060993065,
0.0335067026,
-0.0015027894,
-0.0280504841,
-0.0207623169,
0.0319451503,
0.0371085294,
0.0193655286,
0.0291672871,
-0.0532465316,
-0.063720949,
0.02487606,
0.0247033946,
-0.0186725743,
0.0207298137,
0.0236442126,
0.060072396,
0.0031085573,
-0.0587767325,
0.0054719709,
-0.0042310609,
-0.0316625424,
-0.0317677408,
-0.0361219533,
-6.640959743e-33,
-0.0389576629,
-0.0475995429,
0.0545860119,
-0.023627555,
-0.0174537729,
-0.0233218372,
-0.0241316427,
-0.0050478391,
-0.0224034712,
0.0033945302,
-0.0321395434,
0.0228335354,
0.0063665244,
0.0034637926,
0.0030890694,
-0.0386845358,
0.0119392751,
-0.0099216215,
-0.0377018675,
-0.0243094079,
0.0159503408,
0.0228514746,
0.0528542548,
0.0187158715,
0.030780578,
0.0044193012,
-0.0276025906,
0.0684639663,
-0.0451083705,
0.0483083352,
-0.0333966464,
-0.0109144151,
0.0157478843,
-0.0463654995,
-0.0046890727,
0.0077441386,
-0.0473835729,
-0.0700930357,
0.0492513217,
-0.0024277677,
-0.0553201661,
-0.0530394837,
0.00633374,
-0.0388022922,
-0.0047417805,
0.0205532536,
0.0229000822,
-0.013037446,
0.0103805065,
0.0125733381,
-0.03164801,
-0.0077849668,
-0.0190615933,
-0.0266947299,
-0.0227766298,
0.0264585149,
-0.0340633988,
0.0079001496,
-0.0501809455,
0.0421515927,
0.0311208852,
0.0046755904,
0.0734000877,
-0.0078268107,
-0.0133846039,
0.0676485896,
0.0338974595,
-0.0239198245,
-0.037757121,
-0.0176763479,
0.03403062,
0.0560042709,
0.0591668673,
0.0379227884,
0.0613309667,
0.0379370973,
-0.0551131219,
0.0360889174,
0.0268889628,
0.0341683477,
0.0098193986,
-0.0281329267,
-0.0250348859,
-0.0167592391,
-0.046082329,
0.0282469187,
-0.0584736131,
-0.0511165299,
0.0470636189,
-0.0183401462,
-0.0184690319,
-0.0458152816,
0.0289657805,
-0.0346655995,
0.0384535901,
0.0065257321,
-0.0000951833,
0.0084822206,
0.0115385223,
0.0339239836,
-0.0314811207,
0.01095175,
0.0381001271,
0.0073988959,
0.0115444325,
-0.0108116772,
0.0229749419,
0.0181133151,
-0.0534905232,
-0.0330999345,
-0.0361403413,
0.0316931456,
0.0016236594,
-0.0285387989,
0.0128980828,
0.0235578921,
-0.0269530173,
0.0012079701,
0.0042462051,
-0.0400511399,
-0.026629338,
-0.0015122475,
0.0082306378,
0.0322420113,
-0.0142025771,
0.0100297304,
0.0147789903,
-0.0215415675,
-0.028922094,
-0.057881061,
0.0028227482,
0.0692328736,
2.757e-7,
0.0501122959,
0.0298652444,
0.0050056903,
0.035887856,
0.0022360939,
0.0055283508,
-0.046736896,
-0.0237851124,
0.0229846481,
-0.0146106081,
-0.0048052375,
0.0700069815,
0.0434878096,
0.029282691,
-0.0671950877,
0.0118497508,
-0.0248104911,
-0.0347735547,
-0.0525674298,
-0.0061532017,
0.0630764663,
0.0574427694,
-0.0069984007,
-0.0332148597,
-0.0176517162,
-0.0430953354,
0.0534222871,
-0.0344117619,
-0.0174287818,
-0.0109911393,
-0.0026810032,
0.0060968986,
0.0115373768,
0.06165912,
0.0041232896,
0.003559822,
-0.0069265901,
0.0334594734,
-0.010274983,
-0.0797953978,
-0.0366618223,
0.003242559,
0.0003837255,
-0.045587562,
0.0961245894,
-0.0669956878,
-0.0407813378,
0.0104995249,
-0.0206551664,
-0.021708684,
0.0333796851,
-0.0314584561,
-0.0156923663,
0.0006921776,
0.0332902148,
-0.0068083187,
-0.0152785713,
-0.0553636327,
-0.013553936,
-0.0036427537,
-0.02634985,
-0.053839203,
-0.0472819582,
-0.0106774671,
0.0272518229,
-0.0390182622,
-0.0273684375,
2.63879908e-34,
0.0428481996,
-0.0019742306,
0.010579058,
0.0526459515,
0.0006870838,
0.0077858712,
0.0131865246,
-0.021981461,
0.0259955488,
-0.0880408287,
-0.0150117502
] |
http://arxiv.org/abs/2405.16128v1
|
http://arxiv.org/pdf/2405.16128v1
|
How Well Do Deep Learning Models Capture Human Concepts? The Case of the Typicality Effect
| 2024-05-25T08:38:30 | 2024-05-25T08:38:30 |
Siddhartha K. Vemuri, Raj Sanjay Shah, Sashank Varma
|
cs.AI
|
cs.AI, cs.CL
|
How well do representations learned by ML models align with those of humans?
Here, we consider concept representations learned by deep learning models and
evaluate whether they show a fundamental behavioral signature of human
concepts, the typicality effect. This is the finding that people judge some
instances (e.g., robin) of a category (e.g., Bird) to be more typical than
others (e.g., penguin). Recent research looking for human-like typicality
effects in language and vision models has focused on models of a single
modality, tested only a small number of concepts, and found only modest
correlations with human typicality ratings. The current study expands this
behavioral evaluation of models by considering a broader range of language (N =
8) and vision (N = 10) model architectures. It also evaluates whether the
combined typicality predictions of vision + language model pairs, as well as a
multimodal CLIP-based model, are better aligned with human typicality judgments
than those of models of either modality alone. Finally, it evaluates the models
across a broader range of concepts (N = 27) than prior studies. There were
three important findings. First, language models better align with human
typicality judgments than vision models. Second, combined language and vision
models (e.g., AlexNet + MiniLM) better predict the human typicality data than
the best-performing language model (i.e., MiniLM) or vision model (i.e.,
ViT-Huge) alone. Third, multimodal models (i.e., CLIP ViT) show promise for
explaining human typicality judgments. These results advance the
state-of-the-art in aligning the conceptual representations of ML models and
humans. A methodological contribution is the creation of a new image set for
testing the conceptual alignment of vision models.
|
How Well Do Deep Learning Models Capture Human Concepts? The Case of the Typicality Effect: How well do representations learned by ML models align with those of humans?
Here, we consider concept representations learned by deep learning models and
evaluate whether they show a fundamental behavioral signature of human
concepts, the typicality effect. This is the finding that people judge some
instances (e.g., robin) of a category (e.g., Bird) to be more typical than
others (e.g., penguin). Recent research looking for human-like typicality
effects in language and vision models has focused on models of a single
modality, tested only a small number of concepts, and found only modest
correlations with human typicality ratings. The current study expands this
behavioral evaluation of models by considering a broader range of language (N =
8) and vision (N = 10) model architectures. It also evaluates whether the
combined typicality predictions of vision + language model pairs, as well as a
multimodal CLIP-based model, are better aligned with human typicality judgments
than those of models of either modality alone. Finally, it evaluates the models
across a broader range of concepts (N = 27) than prior studies. There were
three important findings. First, language models better align with human
typicality judgments than vision models. Second, combined language and vision
models (e.g., AlexNet + MiniLM) better predict the human typicality data than
the best-performing language model (i.e., MiniLM) or vision model (i.e.,
ViT-Huge) alone. Third, multimodal models (i.e., CLIP ViT) show promise for
explaining human typicality judgments. These results advance the
state-of-the-art in aligning the conceptual representations of ML models and
humans. A methodological contribution is the creation of a new image set for
testing the conceptual alignment of vision models.
|
[
0.0085655777,
0.0473330729,
-0.0500811115,
0.027086243,
-0.0230305698,
0.0202280413,
-0.017125804,
-0.0476212092,
-0.064109996,
-0.040610794,
-0.0554851331,
-0.0434870198,
0.0163540561,
0.0675963387,
0.0316649005,
-0.0855120495,
0.0051192832,
0.0267984141,
0.0032639296,
-0.020130815,
0.0024160882,
-0.0069333185,
0.0115943057,
0.0609737448,
0.0001687959,
-0.0172335804,
0.0040530311,
-0.0133627262,
0.0323808305,
0.0123983705,
0.0090888385,
0.0369258523,
0.0484751761,
0.0102158813,
0.0000018241,
-0.0493771955,
0.0533562154,
-0.0260682516,
0.0214056261,
0.0101812063,
0.0256513506,
-0.010078053,
0.0068789218,
-0.0228693485,
-0.0186315347,
-0.0406478234,
0.0686543807,
0.0639697611,
-0.0033974564,
0.0029038887,
0.0049636215,
-0.0697250739,
0.0121308425,
-0.0093255136,
0.01796606,
-0.0078563634,
0.0150660714,
-0.0026937625,
-0.0408601686,
-0.0187921263,
-0.03972077,
-0.0086895861,
0.0299191289,
0.0509275049,
0.0356919691,
0.0509764552,
-0.0368402861,
-0.0616492964,
0.0078955898,
0.0895211548,
-0.0181918498,
-0.023868436,
-0.0445486717,
0.008893325,
0.0249679014,
0.0283695981,
-0.0202048328,
0.0395311713,
0.0133817345,
0.0159834828,
-0.0445128866,
0.0237694364,
0.0181125235,
0.0147768725,
-0.0252639223,
-0.0145388898,
0.0084397141,
0.0006468201,
-0.0012640482,
0.0077516721,
-0.0312811658,
-0.0163873266,
-0.0083016623,
-0.0146813374,
0.0269026104,
-0.0105224391,
-0.0142846527,
-0.0019647798,
0.0155144492,
-0.0252613649,
0.0518454164,
0.0176041033,
-0.0692101941,
0.0066340347,
-0.0372824781,
0.0699293166,
-0.0238345135,
0.0148609644,
-0.0069889459,
-0.019283263,
-0.0550072901,
-0.0097044129,
0.0326087922,
0.0244160816,
0.0311326999,
-0.021033084,
-0.0349460542,
-0.0183212869,
-0.0542362034,
0.0022625721,
-0.0481342636,
0.0693304762,
-0.0273150913,
0.0089433752,
-0.0627537891,
0.0006590991,
-0.0398078933,
0.0397047475,
-0.0191425551,
-0.0730631799,
0.0611397959,
-0.0083823614,
0.0243299529,
-0.005407284,
0.0018066707,
0.059619315,
0.003814484,
-0.0132589601,
0.0192458611,
-0.0133331865,
-0.010491997,
-0.0430120006,
0.0025757994,
-0.0729493201,
-0.0039633866,
-0.0041638399,
0.0160973296,
-0.0223567244,
-0.0203421917,
-0.0009378761,
-0.0077452776,
0.0822594017,
0.0275817644,
0.0083718291,
0.0224258546,
0.0278326049,
0.0069852103,
0.0458124466,
0.0119096944,
0.0214119032,
0.0162690785,
-0.0315004997,
0.018553244,
0.0138575444,
-0.0080082845,
-0.046808593,
-0.0556855351,
-0.0243383944,
-0.0126891797,
0.0569933914,
0.0816459805,
0.0553381778,
-0.0146295093,
-0.0639546812,
0.0854906887,
0.051226683,
-0.0121802036,
0.0569603741,
0.0411418267,
0.0106530758,
0.0105580939,
0.050662782,
-0.0703406557,
0.042810902,
0.0263770297,
-0.0106662577,
-0.0431514606,
0.0493003055,
-0.0411618464,
-0.0183064006,
-0.0064355433,
-0.0102352202,
0.0091576921,
-0.0111113777,
0.0287371948,
-0.0208918601,
0.0229290128,
0.0079714619,
0.0160164572,
-0.0064040944,
-0.0157114193,
-0.0677797496,
-0.0023503851,
-0.0021002407,
-0.0138494801,
0.0263971109,
-0.0057949075,
0.00481991,
-0.0747047663,
0.021981433,
0.0569185354,
0.0095720356,
0.0151480241,
-0.0065391944,
-0.023320375,
-0.0047529759,
0.0250394605,
0.0350664593,
-0.0138020059,
0.0218966864,
0.0352610238,
-0.0586512201,
-0.0072255693,
0.0306602437,
0.0148108844,
-0.0013358544,
-0.0279979873,
0.0255690981,
0.0177727006,
-0.0098221432,
-0.0001282768,
0.0215101242,
-0.0014780855,
-0.0049654832,
-0.0008180485,
0.0593335517,
-0.0601969361,
0.0203535128,
0.0107579548,
-0.1137793884,
-0.0071145664,
0.0952166021,
-0.0165061038,
-0.0046577565,
0.026642574,
-0.0347858034,
0.0326454267,
-0.0288021993,
-0.015371765,
-0.0404785722,
0.0457792953,
0.0288071018,
0.0445632674,
0.0013659229,
-0.004813116,
-0.0027978974,
-0.0471937433,
-0.0020121981,
-0.0507949702,
-0.0338032283,
-0.00554734,
-0.0148475524,
0.0199767202,
0.0159850176,
-0.0276137106,
-0.0529896058,
0.0693732277,
-0.0909600705,
-0.0209828485,
0.0069901026,
-0.050837066,
-0.0155967129,
-0.0342368484,
-0.0000153371,
-0.0238150712,
-0.0518504828,
0.019012779,
-0.0439846963,
0.0623241849,
0.0187904127,
-0.004768061,
-0.0421817824,
-0.0122746062,
-0.0146293594,
0.0006841485,
0.0596088655,
-0.0120711792,
-0.0173280537,
0.0106192296,
0.0227630623,
-0.0587950945,
0.0249082949,
0.0024922574,
-0.0505955033,
0.0463183485,
0.0033426997,
0.0208340827,
0.0712607056,
0.0428387709,
0.0020742849,
0.0549127534,
0.0463628769,
-0.0559712872,
-0.042208869,
0.0004997574,
-0.0166147482,
0.0277388953,
-0.036794547,
-0.0242297761,
0.0522853434,
0.0100849904,
-0.021326419,
-0.0449022241,
0.0669367984,
0.040469721,
-0.0688495263,
-0.0552170947,
0.0187338162,
-0.0123398528,
0.065479517,
-0.0289049298,
-0.0684904531,
0.0558834411,
0.0327983499,
-0.0036461384,
0.0295942482,
0.0138969403,
-0.0063067176,
0.0807590634,
0.0557361171,
0.0293637812,
-0.0249891877,
-0.0137239676,
-0.0871007591,
-0.0103112683,
0.0127918338,
-0.0115849664,
0.0815893337,
-0.012074532,
-0.0416444764,
-0.0036867128,
-0.0440999232,
-0.0644110218,
0.0196373742,
-0.0501521789,
0.019995559,
-0.0132169202,
-0.0002543727,
-0.0222655851,
-0.0306743383,
0.0083024073,
-0.0173877496,
0.008139018,
-0.0119728073,
0.0231389441,
0.0335291885,
0.0591568463,
0.0397113413,
0.0098506007,
-0.0133233145,
0.0232444052,
0.0387359932,
-0.0166429523,
0.035802111,
-0.0850445777,
0.0298989657,
-0.0070285876,
0.0136245312,
-0.0143996049,
0.0343031213,
0.0335224345,
-0.0302284472,
0.0200713556,
0.0614372119,
-0.0003438945,
-0.050469093,
0.0342569537,
-0.0073776194,
0.0190112013,
0.0200391896,
-0.0029293115,
-0.0346375071,
-0.0341628306,
-0.0068372986,
-0.0226044804,
0.0063892496,
-0.0303954445,
-0.0054816292,
-0.0657493696,
0.0051251096,
0.0020863281,
-0.0687856674,
-0.0497827232,
0.0082658995,
0.0235712118,
-0.0388293527,
0.0351228826,
-0.0232338179,
-0.0021649082,
-0.0027593086,
-0.0005105679,
-0.0222282019,
-0.0495209135,
0.0512723476,
-0.0766215548,
0.0415537991,
0.0410290696,
0.0234918687,
-0.0417110883,
-0.0411281995,
0.0079228384,
0.0106117828,
0.0350835323,
0.0415905789,
0.0323619954,
-0.0317195132,
-0.0039998298,
-0.0408750065,
0.0010254177,
0.0189539194,
-0.0175108854,
-0.0450271554,
0.0519355945,
-0.0094094099,
-0.0628253073,
-0.0883891284,
0.0089511583,
-0.0312147494,
0.0920234323,
0.0523475334,
-0.0291373637,
-0.0100468779,
0.0090468042,
-0.0046906779,
-0.0195302553,
-0.053703092,
0.0371392071,
0.0476527438,
-0.0347937942,
-0.0328599475,
-0.037825726,
-0.0158945713,
-0.009774548,
-0.002816577,
-0.0502861887,
-0.0235072989,
0.0168562997,
-0.0552971028,
0.0743908435,
0.007210983,
-0.005008339,
-0.0106170848,
-0.0308889244,
0.0490018986,
-0.0533719882,
-0.0337694474,
-0.0135996472,
-0.0348682255,
-0.0267358106,
0.0220673401,
0.0198626742,
0.0243427735,
0.0523939244,
0.0487069935,
-0.0591789931,
0.0503351986,
0.0612328053,
0.0440774783,
-0.0157230143,
-0.0026584771,
-0.0184711944,
-0.0110979965,
0.0298806746,
-0.0371161178,
0.0692973882,
-0.0173338,
-0.0320026837,
-0.002039707,
-0.053460069,
0.0230944958,
-0.0156429466,
0.0399097577,
-0.0198612232,
-0.012935997,
-0.0188420024,
-0.0608003624,
0.0517343394,
0.0059650624,
-0.0208724476,
-0.0458519794,
0.0315729938,
0.0520857349,
-0.024703024,
-0.0346884876,
-0.0143076666,
0.0265900195,
-0.0257079918,
-0.0115023833,
0.0362905636,
-0.0158343762,
-0.0246686209,
0.0413502455,
-0.0313886814,
0.0279295351,
-0.0546261929,
0.0270112064,
0.0188908409,
0.0247522797,
0.0223343801,
0.0294726882,
-0.0231845435,
0.0189351961,
-0.0215730779,
-0.0587696843,
-0.0242526662,
-0.002977865,
0.0420545042,
-0.0420731008,
-0.0420787074,
-0.0198060814,
0.0085659819,
-0.0255468618,
0.0134111689,
-0.0062060072,
0.0117630921,
0.0670135617,
0.0166964978,
0.0096187182,
-0.0352183878,
0.0357968062,
-0.025477428,
0.0227211174,
-0.0440784432,
-0.0291703902,
-0.002504104,
0.0401291959,
0.066760987,
0.0205328669,
0.0121172844,
0.0004069025,
-0.0441006906,
-0.079946205,
0.0091108242,
0.0117000733,
0.0265006758,
0.0212310068,
0.0379968472,
0.0695799887,
0.0025794103,
-0.0284201484,
0.0205468237,
-0.0063793818,
-0.0234907549,
0.0283847861,
-0.0167543925,
-5.670934091e-33,
-0.0071245651,
0.0108968038,
-0.027698243,
-0.0170900002,
0.0091826655,
0.0282828622,
-0.0070914757,
-0.0001535009,
-0.0592544898,
0.04829951,
0.0366580263,
0.0035679964,
0.0014147247,
-0.0164732765,
0.0250080694,
-0.0061029522,
0.0459157117,
0.0076361969,
0.0040038568,
-0.0314785801,
0.0077998801,
0.0067784945,
0.0660059452,
-0.0689568743,
0.0029491563,
-0.0058599263,
0.0066222455,
-0.0153603563,
0.0230503548,
0.0607315078,
-0.0478702299,
0.0700510368,
0.0653212592,
-0.0098854881,
-0.0263062268,
0.0580235235,
-0.0383470319,
-0.0381853282,
0.0109008392,
0.0079406193,
-0.0386378802,
-0.0096356999,
0.0897248387,
-0.0288575068,
-0.0286679994,
-0.0178237129,
0.0905111432,
-0.0251726676,
0.0373327658,
0.0094809085,
-0.0485412255,
-0.0110037113,
0.0105054323,
-0.006152574,
0.0160426572,
0.0053131254,
-0.0152788647,
0.0252446048,
-0.0647279844,
0.0587759092,
-0.0041881599,
-0.020650465,
0.0819570646,
0.0497922376,
0.0125629203,
0.0850689337,
0.0851687267,
-0.0013134751,
-0.053869389,
-0.0323913246,
0.0651847944,
0.052303534,
0.0202386286,
0.0936121121,
0.0820324123,
-0.0081444317,
-0.0027302199,
0.021634547,
0.0325769223,
0.036448881,
0.0140061099,
0.0145925144,
-0.0597269982,
0.0017995796,
-0.0262291748,
-0.0322521962,
-0.0476522557,
-0.0280188266,
0.0153441057,
-0.0347577408,
0.0271725096,
0.0281287674,
0.043119207,
-0.0482537001,
-0.0057798927,
-0.0219430104,
-0.0234551281,
0.007531011,
0.0223360956,
-0.0085400231,
-0.0406702124,
-0.0296474397,
0.0247694198,
0.0042160926,
0.0302590802,
0.0401699506,
0.0188421737,
0.0028377918,
-0.0173675716,
-0.0285762269,
-0.0310486909,
-0.0815252066,
0.0354233868,
0.0171344951,
0.0412523821,
0.0119656567,
-0.042741131,
-0.0147196492,
-0.0015323105,
0.0019244859,
-0.0602101795,
-0.0249646883,
-0.0272822119,
0.0414101444,
-0.0172449108,
-0.0296489131,
-0.0455683023,
-0.069957979,
0.0250607729,
-0.0388352722,
-0.0182881989,
0.0037595914,
2.473e-7,
0.0064244368,
0.0036970957,
-0.0173754282,
-0.0159976371,
0.0543612055,
0.0736462995,
0.0089228768,
0.0242976453,
0.0075842794,
-0.0616037138,
-0.0024244201,
0.0483607464,
0.0002187462,
0.0001222692,
-0.0231533498,
-0.0108834673,
-0.0348920375,
0.0427533723,
-0.0075393012,
0.0218918025,
0.0427153893,
0.0835272968,
-0.0104407463,
-0.0424470454,
0.0353231989,
-0.0229724422,
0.0523977615,
-0.0667110682,
0.0139105124,
-0.0336938351,
-0.0142863458,
0.0510673299,
0.010729162,
0.0581228063,
0.0241791438,
-0.0100584505,
0.0561000779,
-0.007997686,
-0.0448544212,
0.010305021,
-0.0084447218,
0.0222223774,
-0.0082692914,
-0.0709632114,
0.0689079091,
-0.0596989244,
-0.0444508903,
-0.0347904786,
-0.0248428918,
0.0172619,
0.0409376137,
0.010604687,
-0.0182414819,
-0.009014261,
-0.0022633439,
-0.0084373979,
-0.0365660936,
-0.078250654,
-0.0173965059,
0.0159418024,
-0.0705796406,
-0.0267426837,
-0.0330591612,
-0.0568350255,
0.057141576,
-0.0292259119,
-0.0258318428,
1.535979897e-34,
0.0087332847,
0.0581117459,
0.0020576739,
-0.0295467824,
-0.0155935613,
-0.0408121012,
0.0087189162,
-0.0030118709,
0.0307305343,
-0.014774967,
-0.012213218
] |
http://arxiv.org/abs/2405.16122v1
|
http://arxiv.org/pdf/2405.16122v1
|
Prompt Optimization with EASE? Efficient Ordering-aware Automated Selection of Exemplars
| 2024-05-25T08:23:05 | 2024-05-25T08:23:05 |
Zhaoxuan Wu, Xiaoqiang Lin, Zhongxiang Dai, Wenyang Hu, Yao Shu, See-Kiong Ng, Patrick Jaillet, Bryan Kian Hsiang Low
|
cs.AI
|
cs.AI, cs.CL, cs.LG, stat.ML
|
Large language models (LLMs) have shown impressive capabilities in real-world
applications. The capability of in-context learning (ICL) allows us to adapt an
LLM to downstream tasks by including input-label exemplars in the prompt
without model fine-tuning. However, the quality of these exemplars in the
prompt greatly impacts performance, highlighting the need for an effective
automated exemplar selection method. Recent studies have explored
retrieval-based approaches to select exemplars tailored to individual test
queries, which can be undesirable due to extra test-time computation and an
increased risk of data exposure. Moreover, existing methods fail to adequately
account for the impact of exemplar ordering on the performance. On the other
hand, the impact of the instruction, another essential component in the prompt
given to the LLM, is often overlooked in existing exemplar selection methods.
To address these challenges, we propose a novel method named EASE, which
leverages the hidden embedding from a pre-trained language model to represent
ordered sets of exemplars and uses a neural bandit algorithm to optimize the
sets of exemplars while accounting for exemplar ordering. Our EASE can
efficiently find an ordered set of exemplars that performs well for all test
queries from a given task, thereby eliminating test-time computation.
Importantly, EASE can be readily extended to jointly optimize both the
exemplars and the instruction. Through extensive empirical evaluations
(including novel tasks), we demonstrate the superiority of EASE over existing
methods, and reveal practical insights about the impact of exemplar selection
on ICL, which may be of independent interest. Our code is available at
https://github.com/ZhaoxuanWu/EASE-Prompt-Optimization.
|
Prompt Optimization with EASE? Efficient Ordering-aware Automated Selection of Exemplars: Large language models (LLMs) have shown impressive capabilities in real-world
applications. The capability of in-context learning (ICL) allows us to adapt an
LLM to downstream tasks by including input-label exemplars in the prompt
without model fine-tuning. However, the quality of these exemplars in the
prompt greatly impacts performance, highlighting the need for an effective
automated exemplar selection method. Recent studies have explored
retrieval-based approaches to select exemplars tailored to individual test
queries, which can be undesirable due to extra test-time computation and an
increased risk of data exposure. Moreover, existing methods fail to adequately
account for the impact of exemplar ordering on the performance. On the other
hand, the impact of the instruction, another essential component in the prompt
given to the LLM, is often overlooked in existing exemplar selection methods.
To address these challenges, we propose a novel method named EASE, which
leverages the hidden embedding from a pre-trained language model to represent
ordered sets of exemplars and uses a neural bandit algorithm to optimize the
sets of exemplars while accounting for exemplar ordering. Our EASE can
efficiently find an ordered set of exemplars that performs well for all test
queries from a given task, thereby eliminating test-time computation.
Importantly, EASE can be readily extended to jointly optimize both the
exemplars and the instruction. Through extensive empirical evaluations
(including novel tasks), we demonstrate the superiority of EASE over existing
methods, and reveal practical insights about the impact of exemplar selection
on ICL, which may be of independent interest. Our code is available at
https://github.com/ZhaoxuanWu/EASE-Prompt-Optimization.
|
[
0.0156240091,
0.0668184385,
-0.0259938296,
0.0403352343,
-0.0233860221,
0.0067882049,
0.0151906926,
0.0355974659,
0.0409772992,
-0.0533774421,
-0.0050493861,
-0.0126940412,
-0.04579686,
0.0092524458,
0.0305874087,
0.0013484316,
0.030342333,
-0.0457415655,
0.0008334307,
-0.0092297243,
-0.044642821,
-0.0207690783,
0.0056769508,
0.0351101942,
0.0242512859,
0.0216248799,
-0.0149196396,
0.042862501,
-0.0069134389,
-0.038014695,
0.0255155619,
0.0469472036,
0.0214662403,
0.0108426195,
0.0000023714,
-0.0541966558,
-0.0512802489,
-0.0133225089,
-0.0406962186,
0.0194690637,
0.0112150554,
0.0432530455,
0.0004481409,
0.0457290225,
-0.0288215764,
0.0001984052,
0.038749069,
0.0009710621,
0.0350864157,
0.0846602544,
0.0080432501,
0.0441743508,
-0.038354069,
-0.0180189405,
0.0407370999,
-0.0445362777,
0.0242611896,
-0.0846802518,
0.0402056575,
0.0454972573,
0.0041039386,
0.0145835606,
0.0108792139,
-0.0003202871,
-0.0188465379,
0.0129912319,
-0.0348909013,
0.0139214015,
-0.026125174,
0.0584358983,
-0.0519963838,
-0.0189935341,
0.0155299902,
-0.047491461,
0.0262310877,
0.0088828141,
-0.0129390787,
0.000572164,
0.0061319689,
0.0153287742,
0.0094941715,
0.0356651843,
0.0171852559,
0.0072436989,
-0.0319987759,
0.0681580082,
0.0002991471,
-0.0366414599,
0.0203910582,
-0.0237679537,
-0.0452713817,
-0.0135644488,
0.0058529009,
0.0105278231,
-0.0643818825,
0.0115643591,
-0.0263732933,
-0.0177392513,
0.0164951142,
-0.0037156737,
0.0608449057,
0.0435283072,
-0.0026135449,
0.0560500324,
-0.0309725553,
0.0857606605,
0.0373489484,
0.0318604,
-0.0346124135,
0.0114791477,
-0.0198802128,
-0.0023837464,
-0.0644358099,
0.0582061782,
-0.0234700553,
-0.0212205295,
-0.0263673775,
0.0276738387,
-0.0301883817,
-0.0051918039,
-0.0048690052,
0.0144150648,
-0.025799457,
0.0246239882,
-0.0357202813,
-0.0518138073,
-0.0190876611,
-0.0067839078,
0.0030245595,
-0.0413262546,
-0.0259512421,
0.0007078091,
-0.0201475844,
-0.0252388958,
0.0174910259,
0.0784507468,
0.018474631,
-0.0687285662,
-0.0019551173,
-0.0147108696,
0.0353133082,
0.0053626364,
0.0146034015,
-0.0650939494,
0.0173894186,
-0.0011732569,
-0.0191374552,
-0.0419588573,
0.0013545497,
0.0368157141,
-0.01120135,
0.0468711033,
0.0473822355,
0.0167437363,
0.0074529001,
0.0007422552,
-0.0081883892,
0.0745297,
-0.0078248717,
0.0335359983,
0.0454698429,
0.001177522,
0.0239672717,
0.0675950646,
0.0210863203,
0.0069685173,
-0.0265103895,
0.0291615389,
0.0294381063,
0.0340581574,
-0.0085526695,
0.0089954687,
-0.001712485,
-0.0239362661,
0.0262635928,
0.0361162387,
0.036595203,
-0.058824908,
0.0026691773,
0.0062438971,
-0.0115470719,
-0.0162964389,
-0.0506006218,
0.0077314638,
-0.0140855815,
-0.0038165227,
0.0315148644,
0.0214582887,
-0.0112375114,
-0.0291448589,
-0.0139862029,
-0.0002832064,
-0.0197329931,
0.051418297,
0.0209379066,
-0.00528165,
-0.0890074298,
0.0377115496,
0.0226194635,
0.0036642572,
-0.048857566,
-0.0556623824,
-0.0079840757,
0.0384515673,
0.041508276,
-0.076601319,
0.0152069172,
0.011123755,
0.0091220401,
-0.0076373727,
0.0345701873,
-0.0284668524,
0.0050646472,
-0.0383442827,
-0.0431055799,
0.0203745905,
0.0193962809,
0.0118331863,
0.0430912636,
-0.0118085612,
0.0251760874,
-0.0260804407,
-0.0337238945,
0.0434028544,
-0.0265529752,
-0.0993542746,
0.0469233617,
0.0211006012,
0.0182978678,
-0.0282485466,
0.0130349938,
0.0245870668,
0.0575363375,
-0.0118078729,
0.0027658264,
-0.0408725031,
-0.0604051948,
-0.0433681458,
0.0059636342,
-0.014090444,
-0.0129477158,
0.0238361694,
-0.0717405975,
-0.0228434354,
-0.0359681733,
-0.0415783785,
0.0261618849,
-0.0211648345,
0.0246940739,
-0.0083629163,
-0.0154949781,
-0.0820028558,
0.0604249388,
-0.0107086329,
0.0352270268,
0.0297092441,
0.01534604,
-0.0109595032,
-0.0538196713,
-0.0180917419,
-0.0136299981,
-0.0263469741,
0.0376876742,
0.0087476131,
0.0265014116,
-0.0242184866,
-0.0124831265,
-0.1074314341,
0.0172499828,
-0.0163700841,
0.0093196109,
0.0206687637,
0.0102225272,
-0.0144263739,
-0.0379747488,
-0.0208104141,
0.0067445603,
0.0125784241,
0.0235588066,
0.0126859322,
0.0117902588,
-0.0178922918,
-0.0244809315,
-0.02983482,
-0.0440652445,
0.0808611512,
-0.0284119342,
0.0233829599,
-0.0444430709,
0.0261614136,
0.0094215572,
0.0236359164,
-0.025156159,
-0.0036801135,
-0.0255781468,
-0.0283993278,
-0.0258881077,
0.0697673187,
0.0323772207,
0.0084990738,
0.0308346264,
0.0225943178,
-0.0846202075,
-0.0016044877,
-0.0185242388,
-0.0158281028,
0.026601959,
-0.0382066891,
0.0267346166,
0.0324492455,
-0.0100954026,
-0.0075457734,
-0.0244340207,
0.0555839539,
0.0020429627,
-0.0012298208,
-0.0860987008,
-0.0025195859,
-0.0331537947,
0.027021423,
-0.0086247465,
-0.0499323271,
0.0846580416,
-0.0010553611,
0.0191510264,
-0.0572915412,
-0.0102264732,
-0.0351231955,
0.0173147097,
0.0262008607,
0.0189310946,
0.0149734793,
-0.0338889956,
-0.0635089427,
0.0121967094,
-0.0032109455,
-0.0449335165,
0.0533040985,
-0.0223084744,
0.0113483397,
0.0212020427,
-0.0843976215,
-0.0589323752,
-0.0670018941,
0.0366168432,
0.0258600414,
0.0261988845,
0.0192680415,
-0.0419892855,
-0.0055166902,
0.0063545825,
0.0814727619,
-0.0196691975,
0.000366473,
-0.0241290722,
0.0065251621,
0.033229325,
-0.001787069,
0.0230510384,
-0.0389471911,
0.0050171786,
0.0145319076,
0.0797974467,
-0.0335816592,
0.0002522892,
0.0121590123,
0.0348227844,
-0.0687618256,
-0.0027442041,
0.0774958506,
-0.0099430941,
-0.0488625504,
-0.0445261076,
0.0945175365,
0.0384782106,
-0.0329016931,
0.0178375524,
-0.0033810826,
0.0465378426,
0.0717799067,
0.0774316937,
-0.0839840919,
-0.0310950913,
0.0161583852,
0.0471377485,
-0.0318921581,
-0.0329976156,
0.017944755,
-0.1669077128,
0.0436404124,
-0.0063233036,
-0.0505562685,
0.0164735038,
-0.0500732139,
-0.0122084124,
-0.0011992849,
0.0088167759,
0.0508734323,
-0.0318837985,
-0.0017464648,
0.0103779109,
0.0365091413,
-0.0669836923,
0.0372598581,
-0.041761972,
0.0033737805,
0.0489476472,
-0.0152978199,
-0.0115369055,
-0.0042711487,
0.0532147735,
-0.0197463986,
0.0060379161,
-0.0174812265,
0.0451062322,
0.0012875814,
-0.0060556242,
0.0317078307,
0.0052917632,
-0.0320141613,
0.017037794,
0.0211226847,
0.0659136847,
-0.0115674986,
0.0137250498,
-0.0186479036,
-0.0501873307,
0.0485842638,
0.0642880127,
0.0148941586,
-0.0341397822,
-0.0182935689,
-0.0540201552,
-0.0004494671,
-0.0008582074,
-0.0274303649,
0.0189677067,
0.0201547593,
0.006048901,
-0.056372799,
-0.0286810473,
0.0261577833,
0.0613819957,
0.0128408503,
0.0061422,
-0.0274339635,
-0.0066143321,
-0.0514153317,
0.0372560844,
-0.017168358,
0.0034435433,
-0.062880151,
0.0235063918,
0.053992983,
-0.0288531091,
-0.0458747186,
0.0244805645,
-0.0422668532,
-0.0698234662,
0.0072144032,
0.0816098228,
0.0149682714,
0.0067851115,
-0.0151159232,
-0.0605186448,
0.0959546715,
0.0585304461,
-0.02014447,
-0.0007757978,
0.0168268178,
-0.0779293329,
-0.0263686553,
0.0116389645,
-0.0716523677,
0.0216052271,
-0.0007821353,
0.0270093717,
0.0013530038,
-0.0207865331,
0.0266080219,
0.0242092684,
0.0531751402,
0.0028707441,
-0.0465535,
-0.0327833518,
0.0756979808,
0.0016453044,
-0.0383509845,
-0.0142182019,
0.044444859,
-0.0219138358,
0.0147080962,
0.0070812418,
-0.0273087826,
0.0161231719,
0.0299095083,
-0.0342328213,
0.0259173475,
0.0264180992,
0.0229178406,
0.0341594964,
-0.0014353734,
0.0023609318,
0.0505994,
0.0174988732,
0.0613334775,
-0.0507892668,
0.0369720906,
0.0327584893,
0.0167219378,
-0.0267224424,
0.0289775543,
-0.0488990732,
-0.0374352448,
-0.0529920235,
-0.0109885093,
0.0349135399,
-0.0645827353,
0.0017646125,
-0.0074145161,
-0.0022504011,
-0.0347394496,
-0.0127664544,
-0.0060347384,
-0.037318781,
0.0314387642,
0.0520938151,
0.0043700426,
0.02833209,
0.0674833059,
0.0173459649,
0.019976493,
0.0156166991,
-0.002746423,
-0.040979553,
0.0044807354,
0.0728860199,
0.0423302613,
-0.0449311212,
-0.0336171836,
-0.0216785353,
-0.0321226791,
0.0156213287,
0.0122979023,
-0.0077214502,
0.0012539981,
0.0720291063,
0.0816899985,
0.0116392951,
0.0167524889,
-0.0481821932,
-0.0196687076,
-0.0347782113,
-0.052642934,
-0.0210068915,
-6.536325317e-33,
-0.0217134953,
-0.0443428271,
0.014404336,
-0.0608551204,
0.0681881383,
0.0257701632,
-0.0343818776,
-0.0076326523,
0.0141428299,
-0.0343662202,
-0.0380942188,
-0.027013937,
0.02101117,
0.0148333739,
0.0110567249,
-0.0084662568,
0.0018224113,
-0.0175394863,
-0.053846851,
0.0063474886,
0.0105235139,
0.0511365235,
0.0219803639,
-0.0658285171,
-0.0340623111,
0.0103493258,
0.0203627273,
0.0142141208,
-0.0339451954,
0.0740573183,
0.0203963351,
0.0349260531,
0.0052977572,
-0.0640089884,
0.0003381693,
-0.0142518496,
-0.0324015133,
-0.0599784292,
-0.0218480509,
-0.0113516711,
-0.0135566769,
-0.0184471887,
0.0605151467,
-0.038580887,
-0.015613663,
0.0250569042,
0.0361532122,
0.0001461563,
-0.0463891365,
-0.0169154238,
-0.0224113166,
0.0064814966,
0.0106702112,
0.002374358,
0.0258174203,
0.0507352762,
0.0045297141,
0.0266948957,
-0.0310002025,
0.0369782411,
0.0142997224,
0.0426236093,
0.0470882282,
0.025376739,
-0.0009186292,
0.0497683734,
-0.0146881985,
0.0012212528,
0.0543193109,
-0.0555788428,
0.0541621856,
0.0277073532,
0.01471595,
0.02845506,
0.0588344634,
0.0015816633,
-0.079422839,
0.0405803099,
0.037675254,
-0.0601693094,
0.0497343615,
0.0086039258,
-0.085801594,
-0.0387783796,
-0.0549277924,
0.0149453273,
-0.0518910177,
-0.0202118456,
0.0650939345,
0.021779988,
-0.0411992744,
-0.042417977,
0.0410552844,
-0.0345990285,
-0.0582834259,
0.017406797,
-0.0379680693,
-0.0155318156,
-0.0055926917,
0.0437327959,
-0.0424373113,
-0.0471059866,
0.0041060699,
-0.0254362989,
0.0356382579,
0.0147727951,
0.0051303999,
0.0127948467,
-0.0815078542,
0.0106044346,
-0.0222372431,
0.0135076856,
-0.0122637656,
0.0000559951,
0.0372362696,
0.0090077044,
0.0220956393,
0.039576143,
0.0420694165,
-0.0716415793,
-0.0529878065,
-0.0109342365,
-0.004214915,
0.0003736663,
-0.0035818061,
0.0154867545,
0.0125364335,
0.0637194365,
0.0188292228,
-0.0330249034,
-0.0165394414,
0.0237337705,
3.045e-7,
0.0232016966,
0.0655567497,
-0.0207150001,
0.0503270738,
-0.0096356785,
0.041356761,
-0.0183034986,
0.0213209242,
0.0168105587,
-0.0164872631,
0.0204005856,
0.001252554,
0.0160963666,
0.0127472114,
-0.0823228508,
0.0149355112,
-0.0352767929,
0.0218985863,
-0.0500635691,
-0.0278881807,
0.0060628895,
0.1131719798,
-0.0192765687,
-0.0048815426,
-0.0343497545,
-0.0486782752,
0.0246474929,
-0.0442427322,
0.0058759572,
-0.0016813433,
-0.0113779875,
0.0236013774,
-0.0065643531,
0.036579784,
-0.024748994,
0.0402751267,
0.017250767,
0.0456832871,
0.0145422351,
-0.0683422014,
0.006615778,
0.0054721162,
0.0111551192,
-0.0621253811,
0.0632827803,
-0.0487868413,
0.0072834305,
0.0016558132,
-0.0004833115,
-0.0505657084,
0.016613204,
-0.0295790397,
0.0057122041,
0.0169349387,
0.0045482484,
-0.0008752036,
0.0064462917,
-0.0640024841,
-0.0195381921,
0.0263254121,
-0.0355280228,
-0.0258997008,
-0.0431296565,
0.048925139,
-0.0213070139,
-0.0582463108,
0.0116065303,
3.607630734e-34,
0.0187450331,
-0.0023645789,
0.0172219537,
0.0511496514,
0.0428313613,
-0.0077692303,
-0.0146781579,
0.0320428573,
0.0309957564,
-0.0657164976,
-0.0297269262
] |
http://arxiv.org/abs/2405.16115v1
|
http://arxiv.org/pdf/2405.16115v1
|
SNOBERT: A Benchmark for clinical notes entity linking in the SNOMED CT clinical terminology
| 2024-05-25T08:00:44 | 2024-05-25T08:00:44 |
Mikhail Kulyabin, Gleb Sokolov, Aleksandr Galaida, Andreas Maier, Tomas Arias-Vergara
|
cs.CL
|
cs.CL, cs.LG
|
The extraction and analysis of insights from medical data, primarily stored
in free-text formats by healthcare workers, presents significant challenges due
to its unstructured nature. Medical coding, a crucial process in healthcare,
remains minimally automated due to the complexity of medical ontologies and
restricted access to medical texts for training Natural Language Processing
models. In this paper, we proposed a method, "SNOBERT," of linking text spans
in clinical notes to specific concepts in the SNOMED CT using BERT-based
models. The method consists of two stages: candidate selection and candidate
matching. The models were trained on one of the largest publicly available
dataset of labeled clinical notes. SNOBERT outperforms other classical methods
based on deep learning, as confirmed by the results of a challenge in which it
was applied.
|
SNOBERT: A Benchmark for clinical notes entity linking in the SNOMED CT clinical terminology: The extraction and analysis of insights from medical data, primarily stored
in free-text formats by healthcare workers, presents significant challenges due
to its unstructured nature. Medical coding, a crucial process in healthcare,
remains minimally automated due to the complexity of medical ontologies and
restricted access to medical texts for training Natural Language Processing
models. In this paper, we proposed a method, "SNOBERT," of linking text spans
in clinical notes to specific concepts in the SNOMED CT using BERT-based
models. The method consists of two stages: candidate selection and candidate
matching. The models were trained on one of the largest publicly available
dataset of labeled clinical notes. SNOBERT outperforms other classical methods
based on deep learning, as confirmed by the results of a challenge in which it
was applied.
|
[
0.0178393926,
0.0458689407,
0.0127397804,
-0.0226054378,
-0.0284075812,
0.0370814539,
-0.0385655351,
0.0233779922,
-0.0049463916,
-0.011562055,
-0.0260429513,
-0.0359145701,
0.0179748684,
0.0207580719,
-0.0089877686,
-0.0587731972,
0.0231658258,
0.0521222241,
-0.0006504104,
0.0154861752,
-0.0238083079,
0.0293188002,
-0.0297886524,
0.0313419998,
-0.0202656891,
0.0261930339,
0.0322833695,
0.0255422611,
0.0403979309,
-0.0647861287,
-0.0159381274,
0.0308106318,
-0.020722039,
0.0402792841,
0.0000017732,
-0.0802059546,
-0.0629757568,
0.0187043753,
-0.0580159798,
-0.0308303125,
0.0130116763,
-0.0591486506,
-0.0099752406,
-0.0068510855,
-0.0166424084,
-0.0368119851,
0.0355599821,
0.0544379018,
0.0454062596,
0.0813119635,
-0.0226221625,
0.0130701177,
0.0236428697,
-0.0257359799,
0.0201478023,
-0.0618401505,
0.0579497702,
-0.0349720456,
0.0255256854,
0.0160853732,
0.0049939551,
0.0373002104,
-0.0000274914,
0.0350784771,
0.0149448384,
0.0575855933,
-0.0660242587,
0.0128127169,
0.024377875,
0.0843184069,
-0.0054159388,
-0.0011959794,
0.0231186487,
-0.0084594171,
-0.0320982635,
-0.0004043982,
-0.0041291052,
-0.0235765893,
-0.0271533318,
0.0026404471,
0.0201974753,
-0.0205959585,
0.0300142076,
0.0140569042,
-0.0431539156,
-0.0307936277,
0.0046728239,
-0.0271916892,
0.0420904979,
0.0055241776,
-0.0228424072,
-0.0462262034,
0.0583208948,
0.0397625379,
-0.0295235775,
0.0192370862,
-0.0418469571,
-0.0620330423,
0.0531677045,
-0.0628873929,
-0.0508350022,
-0.0032206962,
-0.0259567499,
0.0442724526,
-0.0520124063,
-0.0313927829,
-0.055069603,
0.0456621908,
-0.0305177867,
-0.0331584588,
-0.0398143977,
-0.0086157471,
-0.0745330676,
0.0087840119,
0.0138211418,
-0.0185236596,
-0.0562794469,
-0.0434895158,
-0.0007445436,
0.0609569624,
0.0251515266,
0.0345352888,
-0.0322717391,
0.019249633,
-0.024778448,
0.0083192699,
-0.0096609285,
-0.0347296745,
-0.0468816608,
-0.0142117571,
0.0306364093,
0.0363136716,
0.0304943416,
-0.025908282,
0.0518869273,
0.0252787303,
0.0318211429,
-0.0236781761,
-0.0170811545,
-0.0493467823,
0.0048526851,
0.0355292782,
0.0085131302,
-0.061618533,
0.0290908702,
-0.0035627224,
0.0008297418,
-0.026841322,
0.0132071162,
0.019250283,
-0.0800717548,
0.0762818381,
-0.0903898031,
-0.0346552171,
0.0435975902,
0.0355150886,
0.0179183371,
-0.0161809642,
0.0564466529,
-0.0331177935,
-0.0124277119,
0.0103406366,
-0.0081782462,
0.0221711658,
0.0315198898,
-0.0613889955,
0.0077709211,
0.0236604363,
0.0175272711,
0.0952777267,
-0.0210126359,
0.0799667388,
-0.013759803,
-0.0513981208,
-0.0003947812,
0.0691842064,
0.0484784991,
0.0585402064,
0.0067474949,
-0.029636072,
0.0050796168,
-0.0377334543,
-0.0091709364,
0.0136698307,
0.0409125611,
-0.0129082454,
-0.0220022462,
-0.0115191294,
-0.0422137491,
-0.0025678989,
-0.0348977298,
0.0022820518,
-0.0374765769,
-0.0208725743,
0.0630127341,
-0.0030617064,
-0.0444975197,
0.0869348124,
0.023603335,
0.012483079,
-0.0173426978,
-0.0219358243,
-0.0562615059,
0.058919508,
0.0512387566,
-0.0560059026,
-0.0754473284,
-0.0283076316,
-0.0151583999,
0.045511052,
0.0315116309,
-0.019097317,
-0.0102590984,
-0.0425660349,
-0.0021374454,
0.0571807548,
-0.002178929,
0.0022170031,
-0.0037880139,
0.0380480886,
0.0012683655,
-0.0289533362,
0.0129978377,
0.0560310185,
-0.0138143348,
-0.0186028145,
-0.0212337933,
0.0348512605,
0.0036182073,
0.0245644767,
-0.0261016265,
0.0187224466,
0.0191208497,
-0.0285877343,
-0.0470404997,
0.0013603395,
-0.0412686355,
-0.0196893252,
0.0073856469,
-0.0563188493,
-0.007772835,
-0.0142802903,
-0.0251754392,
-0.0245471504,
-0.0325510167,
-0.0358545557,
0.0233674254,
0.0130109396,
0.010661603,
0.0530554354,
-0.0007365894,
-0.0164396409,
0.040900562,
-0.0234011095,
0.0212031454,
0.039048586,
-0.0238519665,
0.0432405062,
-0.0266516581,
-0.0423254408,
0.0212667771,
-0.0177207645,
0.0012249069,
-0.0217509829,
0.0280633066,
-0.0228827298,
-0.0212449897,
-0.1012514979,
0.0335512049,
0.0106575638,
0.031295307,
0.0453799814,
0.0206102058,
-0.0019250452,
-0.0164091438,
-0.0533281937,
-0.0344816521,
0.0142055182,
-0.0048712292,
0.0659990683,
0.0597309023,
-0.0149977831,
-0.0293383151,
-0.0057397885,
-0.0649580285,
0.0176260751,
0.0227538832,
-0.051761765,
0.0279995184,
-0.0157985166,
-0.0276933443,
0.0149321137,
0.0102062402,
-0.0127393426,
0.0165059138,
-0.0123788938,
-0.0206913762,
0.0482072122,
-0.0221783482,
0.0262995251,
0.0049479739,
0.0162333306,
-0.0297827851,
0.0321184807,
-0.0150743155,
0.0002183481,
0.0123678762,
-0.0568871237,
0.0161806419,
0.0874593556,
-0.0286361091,
0.0040074172,
-0.0430662334,
-0.0555165522,
0.0007312528,
0.0560219809,
-0.0414012261,
0.018219009,
-0.0149710122,
0.0230887663,
0.0066973548,
-0.004611141,
0.0803921223,
-0.0190529134,
0.0393690541,
-0.0808959305,
0.0039054935,
-0.0441726185,
0.0156134041,
-0.0069744415,
0.0394331627,
-0.0066687982,
-0.0338577256,
0.0063384948,
0.0279384609,
0.0081090555,
-0.053929694,
0.0287112631,
0.0219494086,
0.048061315,
0.0001964628,
-0.0097095212,
-0.0450101905,
0.0079757301,
-0.0100539578,
0.0248247422,
0.0207957942,
-0.0129369134,
-0.0667809099,
0.0302362777,
-0.0165057369,
0.0870698914,
0.0641064122,
-0.0116475541,
-0.0057068616,
0.0455303267,
0.0291356631,
-0.0434155352,
-0.0006042531,
-0.0087711979,
-0.0092301527,
0.066812247,
0.0275968686,
0.0602033734,
-0.0173923206,
-0.0194285922,
0.001615801,
-0.0729257986,
-0.0178315938,
0.0054535577,
-0.0252704993,
-0.0353426635,
-0.0080489684,
0.1113752574,
0.0546654686,
-0.0409638844,
0.0055216141,
0.0133930482,
-0.030365169,
-0.0451956056,
0.0100821462,
-0.0854935423,
-0.0123128882,
-0.0126474956,
-0.0030029479,
-0.0029278214,
-0.0475403182,
0.0359446667,
-0.0218505487,
-0.0125246439,
-0.0334460363,
-0.03867824,
-0.0201045927,
-0.0664628372,
0.0110974954,
-0.0348298624,
-0.0366927162,
-0.0146556208,
0.0112890871,
-0.0215551741,
0.0152225411,
-0.0650820881,
-0.0585225299,
0.0274556205,
-0.0776137561,
0.0403808504,
0.0527783483,
0.0569114424,
-0.023546407,
-0.0273837037,
-0.0031611349,
0.0018410992,
-0.0357438475,
-0.0088017005,
0.0873689875,
-0.0382732376,
-0.0091025736,
-0.0159878619,
0.0304283053,
-0.0472461432,
-0.0106008388,
-0.0531377792,
0.0436704382,
-0.0064192326,
0.0075965142,
-0.0525654256,
0.0128653254,
0.0092883194,
0.0476863645,
-0.0285723414,
-0.0246332735,
0.0028652446,
0.0259187706,
0.0553750284,
-0.0018274985,
-0.0022747898,
0.0583604053,
-0.0683394074,
-0.0479177609,
-0.0638302937,
-0.0044529354,
0.0381317213,
0.0119550405,
0.0194650479,
0.0107011395,
-0.0166042875,
-0.0188681707,
-0.0804256871,
0.0773935169,
-0.0035690172,
0.0281279404,
-0.0420322455,
-0.0670156777,
0.0569210351,
0.0013481819,
-0.0345980786,
-0.007858295,
-0.0204783939,
0.028263405,
-0.0220976211,
-0.005809898,
0.0214812048,
0.004921115,
-0.0144816106,
0.0100567108,
0.0372761302,
0.034487769,
-0.0140309688,
-0.0512332544,
-0.0139560653,
0.0041005011,
-0.021219587,
0.0149585698,
-0.0172472913,
0.0105988737,
-0.0031636155,
0.0260964856,
-0.0109770494,
-0.0610129237,
0.0180865061,
0.0126327034,
0.0406037085,
0.0025317282,
-0.0127482014,
-0.0237090159,
0.0224725921,
0.0540420339,
0.0237545054,
-0.0095493849,
-0.0104133543,
0.0355410241,
0.013319904,
-0.0163056888,
0.0225550532,
-0.0282361675,
-0.0388295762,
-0.0259612203,
0.0510836355,
0.0106668212,
0.0380188376,
0.030833941,
-0.0280234087,
-0.0136562744,
-0.0126235709,
0.0251078103,
0.0320529416,
-0.0296849348,
0.0168674942,
0.0428525135,
0.0674408451,
0.0065150997,
0.0495908782,
-0.0091292681,
-0.0145466914,
-0.0405750088,
0.0007926702,
0.0696452036,
-0.0210018195,
0.0207622834,
-0.0356132016,
-0.0056600943,
0.0104398727,
0.011827088,
-0.0264034495,
0.0268977582,
0.0187081359,
0.0021463623,
0.0312269535,
-0.0129546318,
0.0648445636,
0.0001503474,
0.0792785361,
0.0372365713,
-0.0564736202,
-0.0504216924,
-0.0044318056,
0.046188008,
-0.0010717571,
-0.0535689481,
-0.0158481486,
-0.0195368603,
-0.0246570148,
0.0704222098,
0.0076166978,
-0.0265779048,
0.0279248785,
0.0895827413,
0.0306227654,
0.0218190104,
-0.0067396043,
0.0267199203,
0.0228806045,
-0.0197396223,
-0.0288382545,
0.0260591004,
-5.537102589e-33,
-0.0081386492,
-0.0150601501,
0.0150082698,
-0.0140093854,
0.0038937435,
0.0304261819,
-0.0209952369,
-0.0516749397,
-0.0268085934,
0.0081654768,
-0.0347144529,
-0.0054128831,
0.0111706657,
0.0394558944,
-0.022531176,
-0.0347087532,
0.0267908406,
0.0244720746,
-0.0175666008,
0.0125606386,
-0.0251484439,
0.0588669367,
0.0828716159,
-0.0941029787,
-0.0114368228,
0.0202028453,
0.0398848616,
0.012652697,
0.020815067,
0.0536041595,
-0.0331633687,
-0.0074063474,
0.0324469618,
-0.0291261561,
0.0236201379,
0.0324309245,
-0.1087210476,
-0.0352396332,
-0.0317586437,
-0.0027100954,
-0.0247667637,
-0.0275649801,
0.0424845219,
-0.0101943547,
0.0302496403,
0.0079178773,
0.0446209237,
-0.0070036785,
-0.0172728617,
-0.0556446798,
-0.0194748584,
0.0768154263,
0.0002672475,
-0.0212742928,
0.0020317037,
-0.0430691838,
-0.024838984,
0.0157428551,
-0.0334729329,
-0.028584199,
0.0499121919,
0.0411134027,
-0.0129565615,
-0.0347892866,
-0.015718855,
0.0645320863,
0.0179433282,
-0.0200596154,
-0.0268208534,
0.014240467,
-0.0118667493,
0.0337712094,
0.0225780103,
-0.0067626662,
0.0353996828,
-0.0142709957,
-0.1055689305,
0.0666841865,
-0.0361154452,
0.0468771197,
-0.0124574536,
-0.0210901145,
-0.0425097719,
0.0156884044,
-0.0182309691,
-0.0190728493,
-0.0054550068,
-0.0096553536,
0.0131121213,
0.0104354098,
-0.0135209002,
0.055926431,
0.0074923239,
0.0043687671,
0.0138832675,
-0.0344005525,
0.0550046042,
0.0234937388,
-0.049164854,
0.0117839491,
-0.0752018914,
-0.0528057404,
0.0270431936,
-0.0375469513,
0.0464314148,
0.0354933739,
0.0377973355,
-0.0213299356,
-0.0406624228,
0.0000711237,
0.029824812,
-0.0455016345,
-0.0614347011,
0.0402737223,
0.057442762,
0.0219267104,
0.0285251029,
0.0421714447,
0.0022067914,
0.0092784381,
-0.0411313884,
-0.0186191443,
0.0138600897,
-0.0171009358,
-0.0272287745,
-0.0127870105,
-0.0362767093,
0.0304750483,
0.0350219868,
-0.0831319839,
0.002804776,
0.0172287952,
2.691e-7,
0.0880434215,
0.03841215,
0.0299911574,
-0.0184263512,
0.0168718547,
-0.00593282,
0.0240364019,
0.0468018241,
-0.0343050882,
-0.0324531421,
0.0536343418,
-0.011310895,
0.0054257801,
0.0402510539,
-0.0739600286,
-0.0402961373,
-0.0665656403,
-0.0004179544,
-0.0048351935,
-0.0331850164,
0.0518793836,
0.054675106,
-0.0285566691,
-0.0385439843,
-0.0066364063,
-0.0048934394,
0.0290202051,
-0.0482377931,
0.0431513973,
0.054353375,
0.0088248942,
0.0412583575,
-0.0096490104,
0.0733995959,
0.0048898724,
0.0243528709,
0.0091070086,
0.0520576835,
-0.0511079058,
-0.0002294327,
0.0075421962,
-0.0354800597,
0.0072161616,
-0.0020642646,
0.0469973125,
-0.0451220125,
-0.0270513538,
0.0413985103,
-0.0187178962,
0.012461449,
0.0681932718,
-0.0120861912,
-0.0296062212,
0.0354601666,
0.0418756679,
0.0095221289,
-0.0553059056,
-0.1113194004,
0.0229307823,
0.0273363385,
-0.0135074537,
-0.0059525557,
0.0335678831,
-0.0277908836,
0.0124342339,
-0.0591886565,
0.0088968175,
2.320956887e-34,
0.0035467788,
-0.0132556912,
-0.0297864284,
0.0286100134,
0.0109182457,
-0.0330880396,
0.046290841,
-0.0053541921,
-0.0054400358,
-0.0438550562,
-0.0103597287
] |
http://arxiv.org/abs/2406.00032v1
|
http://arxiv.org/pdf/2406.00032v1
|
Paths of A Million People: Extracting Life Trajectories from Wikipedia
| 2024-05-25T06:57:33 | 2024-05-25T06:57:33 |
Ying Zhang, Xiaofeng Li, Zhaoyang Liu, Haipeng Zhang
|
cs.CL
|
cs.CL, cs.AI, cs.IR
|
Notable people's life trajectories have been a focus of study -- the
locations and times of various activities, such as birth, death, education,
marriage, competition, work, delivering a speech, making a scientific
discovery, finishing a masterpiece, and fighting a battle, and how these people
interact with others, carry important messages for the broad research related
to human dynamics. However, the scarcity of trajectory data in terms of volume,
density, and inter-person interactions, limits relevant studies from being
comprehensive and interactive. We mine millions of biography pages from
Wikipedia and tackle the generalization problem stemming from the variety and
heterogeneity of the trajectory descriptions. Our ensemble model COSMOS, which
combines the idea of semi-supervised learning and contrastive learning,
achieves an F1 score of 85.95%. For this task, we also create a hand-curated
dataset, WikiLifeTrajectory, consisting of 8,852 (person, time, location)
triplets as ground truth. Besides, we perform an empirical analysis on the
trajectories of 8,272 historians to demonstrate the validity of the extracted
results. To facilitate the research on trajectory extractions and help the
analytical studies to construct grand narratives, we make our code, the
million-level extracted trajectories, and the WikiLifeTrajectory dataset
publicly available.
|
Paths of A Million People: Extracting Life Trajectories from Wikipedia: Notable people's life trajectories have been a focus of study -- the
locations and times of various activities, such as birth, death, education,
marriage, competition, work, delivering a speech, making a scientific
discovery, finishing a masterpiece, and fighting a battle, and how these people
interact with others, carry important messages for the broad research related
to human dynamics. However, the scarcity of trajectory data in terms of volume,
density, and inter-person interactions, limits relevant studies from being
comprehensive and interactive. We mine millions of biography pages from
Wikipedia and tackle the generalization problem stemming from the variety and
heterogeneity of the trajectory descriptions. Our ensemble model COSMOS, which
combines the idea of semi-supervised learning and contrastive learning,
achieves an F1 score of 85.95%. For this task, we also create a hand-curated
dataset, WikiLifeTrajectory, consisting of 8,852 (person, time, location)
triplets as ground truth. Besides, we perform an empirical analysis on the
trajectories of 8,272 historians to demonstrate the validity of the extracted
results. To facilitate the research on trajectory extractions and help the
analytical studies to construct grand narratives, we make our code, the
million-level extracted trajectories, and the WikiLifeTrajectory dataset
publicly available.
|
[
0.0466138907,
0.1190300584,
-0.023388464,
0.0263634883,
0.0080483034,
0.0175704863,
-0.0051743248,
0.0118640307,
-0.0088002626,
0.0201608539,
0.0784465894,
-0.0297683235,
-0.0168778636,
-0.0013466706,
-0.0011499018,
-0.1074443087,
0.0080809118,
0.0027005759,
-0.0132048829,
0.0112391412,
-0.0457686484,
0.0285336822,
-0.0216740072,
0.0299839191,
-0.0219567958,
-0.010774482,
-0.0138642294,
0.0102474606,
0.0832965299,
-0.0226784013,
0.0195512194,
0.0389071591,
0.0608219951,
0.0180305447,
0.0000020585,
-0.0228711199,
-0.0008133491,
-0.0391753539,
-0.0335990787,
-0.016510997,
0.013899019,
0.043002665,
-0.0089667514,
-0.0308958907,
-0.0104252081,
0.0161128696,
0.000374659,
0.0156568177,
0.0152765783,
0.0465188846,
-0.0028822874,
-0.0386676788,
0.0156522747,
-0.0263688061,
0.0507980026,
0.0049268296,
0.0436238758,
0.0001018161,
-0.04582683,
-0.0385036692,
-0.0161593687,
0.0274164714,
-0.0233204179,
-0.0021502872,
0.063257955,
0.0495006964,
-0.0056192568,
0.0015587744,
0.0153886359,
0.0560708381,
0.0677455962,
0.0251966733,
0.0639429167,
0.0227310602,
-0.013413527,
0.0333381407,
-0.0470931716,
0.0306507293,
-0.0174356569,
-0.0183101669,
0.0008658271,
0.0302609242,
0.0080656111,
-0.0060253106,
-0.0113536557,
-0.005224389,
0.0115372082,
-0.0103826048,
-0.0185939111,
0.0014840058,
0.0616148151,
-0.0515475348,
0.0281650778,
-0.0024425145,
0.0255900528,
-0.0119269304,
0.0033365849,
0.0101172365,
-0.0068792249,
-0.022616297,
0.0049017021,
-0.020576138,
0.0242468119,
0.027219478,
0.0773421377,
-0.0262217503,
-0.0315701291,
-0.0527532361,
0.0062791035,
0.021178063,
-0.0709169209,
0.032769952,
0.0101872431,
0.0177486204,
-0.0290503241,
0.0143935457,
-0.02495596,
0.0037417193,
0.0089051519,
0.0135810813,
-0.0266464893,
0.0193542931,
-0.0277187135,
-0.0024802731,
-0.0451220497,
-0.0137985228,
-0.0659333095,
-0.009653586,
-0.0143404324,
0.0688002184,
0.0127419783,
0.0024998342,
0.0778260008,
-0.0358864181,
0.0237977672,
0.063543193,
0.0560986027,
-0.0267467201,
-0.0538136698,
-0.048462946,
0.0146161038,
-0.0375574604,
-0.0202318151,
-0.0716561228,
0.0477740131,
0.0114715332,
-0.0073909601,
-0.0609736443,
-0.0003334748,
-0.0095089031,
-0.0314920805,
0.0109086484,
0.000308397,
0.015073033,
0.0364327356,
0.0083310464,
0.0317363665,
-0.0121917715,
0.0612780936,
0.000888016,
0.0536396429,
0.0398915187,
0.0171282031,
0.0227510575,
0.0009077421,
0.0064688702,
0.0531264395,
0.0294227079,
0.0083782431,
0.0479231328,
0.0207017083,
0.008149717,
-0.0498591661,
0.0030279448,
-0.000770506,
0.0656055063,
-0.0252063125,
0.0591524318,
0.0517927483,
-0.0123096481,
0.0138477422,
0.0148137938,
-0.0231741015,
-0.0021181963,
0.0155722257,
-0.0133194178,
0.0262525398,
0.0460520014,
-0.0137919383,
-0.0068502808,
0.0451789871,
-0.0142509844,
0.0157197285,
0.0000533189,
0.0341976583,
-0.0373732224,
-0.024268331,
0.0314382166,
0.0585552566,
-0.1065239683,
0.0359608829,
0.0379730426,
-0.0107832048,
0.0998154581,
0.0149510298,
-0.0197662711,
-0.0039431574,
-0.0383023918,
0.0499504022,
-0.0044822507,
-0.008816638,
-0.0426223837,
0.0347080231,
-0.0002938962,
-0.049880404,
0.0123509048,
0.003844481,
-0.0097086774,
0.052083347,
0.0309129059,
0.0153964171,
-0.0253725406,
-0.0297296941,
0.0192523357,
-0.0888981596,
-0.0376678966,
-0.011799857,
0.0124344369,
-0.0152012939,
0.0445383042,
-0.02851028,
-0.0556322262,
0.0425359085,
-0.032857351,
0.0082296235,
0.0331091546,
0.0194068365,
-0.0329405628,
0.0161178447,
-0.0836554691,
-0.0636462495,
0.0320049115,
0.002569546,
-0.0493080504,
0.0150630288,
-0.009578797,
0.0172709711,
0.0419264659,
-0.0049793771,
0.0140694799,
0.0107803056,
-0.0368361846,
0.0781590268,
0.0120287705,
0.0103044165,
-0.0102812601,
-0.0278637279,
-0.0069139921,
-0.0661515743,
-0.0255599655,
-0.0351610556,
0.011138794,
0.0045478386,
0.0419460498,
-0.0500592217,
-0.0617157295,
-0.0028779078,
-0.0798604339,
0.0239983406,
-0.0221270993,
0.0312638693,
0.0000771462,
0.0081233913,
-0.0125234844,
0.009239289,
-0.0212728716,
0.0037581846,
-0.0462676138,
-0.0419657752,
0.0539075211,
0.029606089,
-0.0190536492,
-0.0431213044,
-0.0337648764,
-0.0490803868,
0.0276265536,
0.0658459365,
-0.0206492171,
0.028130319,
-0.0244757198,
-0.0410968512,
-0.0030085077,
-0.0389218405,
0.0301935039,
0.0141999125,
-0.0187801998,
0.0381544232,
0.0155844036,
0.0204334483,
0.0583744049,
-0.030992724,
-0.0679138675,
0.0069789849,
0.0188552737,
-0.0342628323,
0.0415701382,
-0.0503838733,
-0.0478684008,
-0.0070191137,
-0.0321022049,
-0.0020640166,
0.0163539108,
-0.0312081557,
-0.0184116792,
0.0056766807,
0.0510187745,
0.0075305365,
0.0357967131,
-0.0144786621,
0.0501805358,
0.0357812159,
0.0236018002,
0.0346523635,
0.0511438698,
0.0495199375,
0.0486904122,
-0.0373667143,
-0.0089501534,
0.0907296836,
0.0167623349,
0.0302544385,
0.0308240354,
-0.0229856633,
-0.0152376005,
-0.0038817432,
0.0248658601,
-0.0303748772,
0.0249317642,
-0.0176681373,
-0.0202646311,
0.0031367184,
-0.0606865697,
-0.0368195064,
-0.0746148229,
-0.0209864434,
-0.0392102934,
-0.0307863522,
-0.0742003322,
-0.0218796115,
-0.042085804,
0.0033008971,
0.0118944263,
0.0064454004,
-0.0347046405,
-0.0045098197,
0.0708928704,
0.0511721075,
-0.0056246743,
-0.0140976384,
-0.0382684991,
-0.0018416479,
0.0072575994,
0.0078406222,
0.0362614617,
0.0291942302,
-0.0107313506,
0.002159277,
-0.0721611604,
-0.0182940233,
0.0280892104,
0.0107111456,
-0.044652164,
-0.0092016915,
0.094985947,
0.0318909995,
-0.0296524744,
0.0165264755,
0.0219009984,
-0.0317139961,
-0.0046719816,
0.0138703249,
-0.0433668233,
0.0277095698,
-0.0305234436,
0.0322221108,
-0.0013670319,
-0.0726222694,
0.0386029258,
-0.0685174987,
0.04625443,
0.0178068243,
-0.0038266403,
0.0042213737,
-0.0491440631,
-0.0131429089,
-0.0717979297,
-0.0600394569,
0.014133703,
-0.0120962095,
-0.005333317,
0.0100435689,
-0.028546771,
-0.036734473,
0.0517182387,
-0.0705033019,
0.0301149506,
-0.0249526147,
0.0849490911,
-0.0202588942,
0.0007884084,
-0.0644514337,
0.0124732237,
0.0018114807,
0.021482436,
0.111576423,
-0.0155771561,
-0.0010197066,
-0.0718126073,
0.0619484633,
-0.0519857369,
-0.0210054051,
0.0417171679,
0.0508310236,
0.0092441728,
-0.0052065365,
-0.0182838552,
-0.0140247568,
0.0428877585,
0.0348987319,
-0.0221524239,
-0.0425186418,
-0.0178841017,
-0.0242022127,
0.0296876263,
-0.0613202341,
-0.0489108078,
-0.0487710834,
-0.0291436277,
0.0004264926,
-0.0536519662,
-0.0357839391,
-0.0194144137,
0.0634741858,
-0.0085165463,
-0.0042561651,
0.0275506787,
-0.0165753439,
-0.0869621933,
0.0439974964,
-0.0617862791,
0.0037347043,
-0.0169534367,
-0.0569451377,
0.0149615733,
0.0306000914,
-0.0041373563,
0.0068160747,
-0.0629114509,
-0.0443744101,
0.0091875829,
-0.0030459706,
0.0237079673,
0.0987935066,
0.0144946231,
-0.0170363188,
0.0401336923,
0.0542116128,
0.0500812642,
-0.017314652,
-0.0211128928,
0.0098058963,
-0.0237639695,
0.0092851287,
-0.0298427884,
-0.0265822578,
-0.031997934,
-0.0120184133,
-0.0787416548,
-0.0182363652,
-0.0437777601,
0.0063837846,
0.0542185046,
-0.0191343892,
-0.0580186993,
-0.0288246181,
-0.0451302975,
-0.0320920721,
0.004996229,
0.0219165403,
-0.0257861391,
0.0406458303,
0.0137608666,
-0.0267256573,
0.0313234814,
-0.0085901143,
-0.0385388732,
-0.1341770291,
-0.0276304614,
0.0243648365,
0.0355473012,
0.020215692,
-0.0244666003,
0.0203823075,
-0.0207595173,
0.0304613169,
0.0453635715,
-0.0041790525,
-0.0327819288,
-0.0031714726,
0.0404688902,
0.0097681452,
-0.0164773092,
-0.1058809459,
0.0011144638,
-0.0276159756,
0.0047328877,
0.0625001341,
-0.0203734059,
0.0631817803,
-0.0028322914,
-0.0237604193,
0.0123816216,
0.0079234494,
-0.0067091906,
0.0357303917,
-0.0016894516,
0.0288307238,
-0.0118342685,
0.0626005828,
0.0465447307,
-0.0030237052,
0.0343884379,
0.0516239367,
-0.0479352772,
-0.0332461447,
0.0011117972,
-0.0214700531,
0.0050322618,
-0.0026825273,
-0.0318766907,
-0.0476530716,
-0.0163315199,
0.0729351565,
0.0183449667,
-0.0186136942,
-0.0146800345,
0.0281209368,
-0.0192407183,
0.0107773365,
-0.0161365084,
-0.0619103312,
0.0225013942,
-0.0533423908,
-0.0220993496,
-0.0121163242,
-6.270952325e-33,
-0.0203963984,
-0.0286058541,
0.0096958559,
0.0275331959,
-0.0874378532,
-0.0195976049,
-0.0430527367,
-0.0012143312,
-0.0173697434,
0.0010662044,
-0.0145301316,
-0.0009199349,
0.0151881324,
0.019529622,
0.0456205197,
0.0197692327,
0.0296365712,
0.0150244348,
-0.0297274552,
0.0679577589,
0.0637294725,
0.0217541307,
0.013404794,
-0.0939139202,
0.0350474305,
-0.0263851937,
-0.0037939667,
0.0045079887,
0.0492284596,
0.0108343065,
-0.0157423057,
0.0186269451,
-0.0160707571,
-0.0157181211,
0.0145714851,
-0.0389357507,
-0.1091431156,
-0.0131344171,
-0.0275365207,
0.0030078865,
0.0269227419,
-0.0435904115,
0.0333717093,
0.0012942636,
-0.0038117131,
-0.0252378751,
0.0963581875,
0.0079814252,
0.0254796222,
0.0099272979,
-0.0277679935,
0.0208415128,
-0.0482158512,
-0.0211718194,
-0.0105381813,
-0.0131687457,
0.0231731925,
0.0207645725,
-0.0652516112,
0.0075061033,
-0.0233266801,
-0.0005044369,
-0.021619942,
-0.0052636871,
0.0295778383,
0.0527017266,
0.00603178,
0.034178365,
-0.0414046161,
0.0222800635,
-0.0032473078,
0.052064199,
-0.0229331516,
0.060632892,
-0.0271468088,
-0.0563382171,
-0.0886622593,
0.0454405621,
0.0353281312,
-0.04289639,
-0.0313742422,
0.0152046811,
0.002761649,
0.0014687184,
0.0083784023,
-0.0660972968,
0.0352270715,
0.01754467,
-0.0449503139,
-0.0204490237,
-0.0586311743,
0.0212248489,
-0.0145671647,
-0.0387910604,
-0.0584807172,
0.0517438352,
-0.0319545828,
0.0235618278,
0.0193899386,
-0.0190174896,
-0.0148879811,
-0.0308035668,
-0.0234890692,
0.0263722502,
0.0437111333,
0.0188261922,
0.0181551948,
0.0223588962,
-0.0194317959,
-0.0301933996,
0.0054254443,
-0.0423899032,
-0.0178049318,
-0.033183679,
0.0241045561,
0.0212843698,
0.017255554,
0.0162429046,
0.0574995428,
0.0841179714,
0.001851983,
-0.0304301642,
0.0266592018,
-0.0051482986,
0.0251007751,
-0.0150363129,
-0.031929899,
0.0560381785,
0.0399808772,
-0.0147229536,
-0.0269603301,
-0.0095707225,
2.765e-7,
0.0173187368,
0.0305609796,
0.0549418889,
0.0095269326,
0.0099495212,
0.014783809,
-0.0017683216,
0.0197679568,
-0.0169938169,
-0.0101724975,
0.0414852798,
0.0132925557,
0.0271098558,
0.0403455496,
-0.0567597523,
-0.0325609185,
-0.026155103,
-0.0402028114,
-0.0822471082,
0.0135048972,
0.0491718389,
0.035272155,
0.049661655,
0.0053937701,
0.0335324109,
0.0110639483,
0.0386955366,
-0.025633242,
-0.0412572958,
0.0025970694,
-0.0256208442,
0.0057401294,
-0.0453714095,
-0.0009086161,
-0.0404283106,
0.0317795724,
0.0236098915,
0.0817291588,
0.0117503535,
0.0294769686,
-0.0151516525,
-0.0087294793,
0.0092252837,
0.0170929059,
0.0446748994,
0.0393451378,
-0.0135741392,
-0.0074925977,
-0.0264435634,
0.0291405283,
0.0392103605,
0.0241549332,
-0.016654186,
0.0070356969,
-0.0150194373,
0.0190695226,
0.0070572989,
-0.0785865039,
0.0154445563,
0.0657957867,
-0.0237156842,
-0.0433254875,
0.0448279008,
-0.0182115156,
0.005593372,
-0.0048101828,
0.0182526316,
2.289729375e-34,
0.0087641384,
0.0566202663,
-0.0251470134,
-0.060246259,
0.0335249566,
-0.0155046582,
-0.0449173674,
-0.027153261,
0.0375086255,
-0.0460929722,
-0.0263838656
] |
http://arxiv.org/abs/2405.16089v1
|
http://arxiv.org/pdf/2405.16089v1
|
COLT: Towards Completeness-Oriented Tool Retrieval for Large Language Models
| 2024-05-25T06:41:23 | 2024-05-25T06:41:23 |
Changle Qu, Sunhao Dai, Xiaochi Wei, Hengyi Cai, Shuaiqiang Wang, Dawei Yin, Jun Xu, Ji-Rong Wen
|
cs.CL
|
cs.CL, cs.IR
|
Recently, the integration of external tools with Large Language Models (LLMs)
has emerged as a promising approach to overcome the inherent constraints of
their pre-training data. However, realworld applications often involve a
diverse range of tools, making it infeasible to incorporate all tools directly
into LLMs due to constraints on input length and response time. Therefore, to
fully exploit the potential of tool-augmented LLMs, it is crucial to develop an
effective tool retrieval system. Existing tool retrieval methods techniques
mainly rely on semantic matching between user queries and tool descriptions,
which often results in the selection of redundant tools. As a result, these
methods fail to provide a complete set of diverse tools necessary for
addressing the multifaceted problems encountered by LLMs. In this paper, we
propose a novel modelagnostic COllaborative Learning-based Tool Retrieval
approach, COLT, which captures not only the semantic similarities between user
queries and tool descriptions but also takes into account the collaborative
information of tools. Specifically, we first fine-tune the PLM-based retrieval
models to capture the semantic relationships between queries and tools in the
semantic learning stage. Subsequently, we construct three bipartite graphs
among queries, scenes, and tools and introduce a dual-view graph collaborative
learning framework to capture the intricate collaborative relationships among
tools during the collaborative learning stage. Extensive experiments on both
the open benchmark and the newly introduced ToolLens dataset show that COLT
achieves superior performance. Notably, the performance of BERT-mini (11M) with
our proposed model framework outperforms BERT-large (340M), which has 30 times
more parameters. Additionally, we plan to publicly release the ToolLens dataset
to support further research in tool retrieval.
|
COLT: Towards Completeness-Oriented Tool Retrieval for Large Language Models: Recently, the integration of external tools with Large Language Models (LLMs)
has emerged as a promising approach to overcome the inherent constraints of
their pre-training data. However, realworld applications often involve a
diverse range of tools, making it infeasible to incorporate all tools directly
into LLMs due to constraints on input length and response time. Therefore, to
fully exploit the potential of tool-augmented LLMs, it is crucial to develop an
effective tool retrieval system. Existing tool retrieval methods techniques
mainly rely on semantic matching between user queries and tool descriptions,
which often results in the selection of redundant tools. As a result, these
methods fail to provide a complete set of diverse tools necessary for
addressing the multifaceted problems encountered by LLMs. In this paper, we
propose a novel modelagnostic COllaborative Learning-based Tool Retrieval
approach, COLT, which captures not only the semantic similarities between user
queries and tool descriptions but also takes into account the collaborative
information of tools. Specifically, we first fine-tune the PLM-based retrieval
models to capture the semantic relationships between queries and tools in the
semantic learning stage. Subsequently, we construct three bipartite graphs
among queries, scenes, and tools and introduce a dual-view graph collaborative
learning framework to capture the intricate collaborative relationships among
tools during the collaborative learning stage. Extensive experiments on both
the open benchmark and the newly introduced ToolLens dataset show that COLT
achieves superior performance. Notably, the performance of BERT-mini (11M) with
our proposed model framework outperforms BERT-large (340M), which has 30 times
more parameters. Additionally, we plan to publicly release the ToolLens dataset
to support further research in tool retrieval.
|
[
0.0345367342,
0.0008401079,
-0.0111654568,
0.0477423966,
-0.0255028121,
-0.0008241654,
-0.025231503,
0.0213562995,
0.0144781219,
-0.0299431589,
-0.0476966128,
0.0027978166,
-0.0668710098,
-0.0136631439,
-0.0126350094,
-0.0427601486,
0.0319319107,
0.013353697,
-0.0420011021,
0.0079891635,
-0.0120861968,
0.0396502912,
0.0081539964,
0.0618573651,
-0.0044128639,
0.0164870396,
-0.0003452203,
0.0258583054,
-0.024060268,
0.0157480128,
0.0118248519,
0.032753408,
0.0389347263,
0.0550208949,
0.0000017637,
-0.0506735742,
-0.051336918,
-0.0124766007,
0.0133128315,
-0.0439202301,
0.0475346409,
0.0209052153,
0.0593447909,
0.0227333326,
-0.0212006923,
-0.0055947565,
0.0445942506,
0.0226838123,
0.0303391069,
0.0399137139,
0.004031152,
0.0032743914,
0.0045514656,
-0.0294521451,
-0.0207626522,
-0.0184294954,
0.0002467794,
0.0027912757,
0.0429958776,
0.023395028,
0.0407707952,
0.059531305,
0.0137007302,
0.0090648439,
0.0439361855,
0.0051488574,
-0.0307421982,
-0.0225094855,
0.0569979176,
0.0528291762,
-0.0140778096,
-0.0380186774,
-0.0002268386,
-0.0020950066,
0.0106719155,
0.0895897225,
-0.0532015637,
-0.0192253944,
-0.0380869508,
-0.0097774295,
-0.0456923395,
-0.0477249138,
0.02714568,
0.0017652393,
0.0033128858,
0.0458006337,
0.0178447962,
0.006812362,
-0.0249964949,
-0.0238658413,
-0.0220429972,
-0.0058261687,
0.0150059378,
0.0375565253,
-0.0342545547,
0.0171529148,
-0.0131021813,
0.0117515177,
0.0230789799,
-0.057802096,
0.0378428251,
0.0160027277,
-0.0386718996,
0.0636213422,
-0.0383740887,
-0.004795312,
0.0004782168,
0.0411831252,
-0.0674381629,
-0.0425934084,
-0.0822638571,
-0.0264015421,
-0.0533986799,
0.0257001873,
-0.0210773014,
-0.0528700687,
0.0281182788,
0.0364442579,
0.0049765059,
0.0288908817,
0.0526671894,
0.0106430296,
-0.0299226735,
-0.0221238043,
-0.0119801741,
0.0235372037,
-0.011520071,
0.0025403129,
-0.0235704873,
-0.0357692689,
0.0088023925,
0.0089521194,
-0.0184317864,
-0.0424897,
-0.0069655823,
0.0739610717,
-0.0130022019,
0.0006216372,
-0.0266702585,
-0.0527338348,
0.0366648436,
-0.0004490885,
-0.0060676122,
-0.0585221313,
0.0255131312,
0.0396812819,
0.0063516409,
-0.0313965902,
0.0083449688,
0.017689079,
-0.0105050243,
0.0679169148,
-0.0115444884,
0.0089730779,
0.0438314192,
0.0245300476,
-0.0100840516,
0.0420538709,
-0.0333676301,
0.0705956966,
0.0535845831,
-0.0177513976,
-0.0128894057,
0.0362422653,
-0.0038256927,
-0.0214700457,
-0.0726426616,
0.018480869,
0.0280842781,
0.0036039387,
-0.0289871786,
0.0147875305,
-0.0239543449,
-0.0216915272,
0.0294707771,
0.0154857179,
0.0932895616,
0.0108870817,
0.0268363692,
0.0403899662,
-0.0186269376,
-0.0152229713,
-0.0314374641,
0.0084881242,
-0.0526402704,
0.0034482062,
-0.0013182716,
-0.0254026037,
-0.0115794037,
-0.0128310891,
-0.0201793816,
-0.0023621693,
-0.0439211801,
0.0292837694,
0.0682144985,
-0.0356098711,
0.0020385098,
0.0669186413,
0.0341718905,
-0.0545344874,
-0.0043755015,
-0.0320693329,
0.0199015103,
0.0720518902,
0.0758692697,
-0.057171382,
-0.04503325,
0.0404997468,
-0.0408192165,
-0.028360717,
0.0237998273,
-0.0518327653,
-0.0107092084,
-0.0410066992,
-0.0708295703,
-0.0115875639,
-0.0173490588,
0.0198373236,
0.0456965081,
-0.0385384299,
0.0439902693,
-0.0969388559,
0.0007169476,
0.0159900654,
0.0092185577,
-0.0046029924,
-0.0228709579,
0.0292116683,
0.0077953255,
0.0020731068,
-0.0246874299,
0.0170895252,
0.0376107916,
-0.0323890224,
-0.0342593826,
-0.0022288458,
0.0120033752,
0.0148196509,
0.0422697589,
0.0231006797,
0.0443110391,
0.0427449718,
0.0068779667,
-0.0300988145,
0.0020054905,
-0.0091989012,
0.0704203025,
-0.0501991138,
0.02038344,
-0.0294357054,
-0.0245356467,
-0.0724364668,
0.0516459681,
0.0002788006,
0.0209628846,
-0.0098833852,
-0.010678824,
0.0077858516,
-0.0179702155,
-0.0500331298,
0.0493370183,
0.0170538463,
0.0423793159,
0.0409477204,
-0.0313441902,
-0.0456681475,
-0.0211881008,
-0.1059461683,
0.0161611475,
-0.0085232304,
-0.0176899694,
0.0001466131,
-0.0008677575,
-0.016844403,
-0.0565946214,
-0.0064372024,
0.018507611,
-0.0059709493,
0.0322974399,
-0.0053900029,
0.0249027815,
0.0002992052,
-0.0319817476,
-0.0376037098,
-0.0019305046,
0.0329051465,
-0.0037015735,
-0.0601845793,
0.0013978166,
0.0689202026,
-0.0412238203,
0.023118956,
-0.0128688198,
0.0061344891,
0.0159734115,
-0.0658637732,
-0.0297780614,
0.02930912,
-0.0152739324,
0.0088901725,
0.0637230575,
-0.0047130673,
-0.1248142272,
0.0055666822,
-0.0092766341,
-0.04595704,
0.0149748418,
0.0190681852,
0.0114577496,
-0.0024048649,
-0.0110459318,
0.0145191206,
-0.0123762283,
-0.0178124569,
0.0151329553,
-0.0265715942,
-0.037429858,
-0.024634311,
0.0181965139,
0.0054343361,
-0.0154888276,
-0.0506932028,
0.0505228043,
-0.027206976,
0.0067632329,
0.0089838738,
-0.0304514561,
-0.0389401168,
0.0883189887,
0.0307546481,
0.037705984,
-0.0376605988,
-0.024629131,
-0.0399467312,
0.0014944676,
0.0596879981,
-0.0533606745,
-0.003182638,
-0.0167788994,
0.0414182208,
-0.0204449296,
-0.0244770292,
-0.0217202529,
-0.0257923808,
-0.0192380641,
0.0286647901,
-0.0129341902,
-0.0286468323,
-0.0066923513,
0.0325551331,
0.0114304237,
0.0166873746,
0.0312954895,
-0.0052002217,
-0.0115310196,
0.0315751657,
0.0436767898,
-0.0279926658,
-0.0473740548,
-0.0188150518,
-0.0105820596,
0.0305864718,
0.0767867491,
-0.0468872748,
-0.0516446382,
-0.0304619521,
0.0597645342,
-0.0315009914,
-0.02758535,
0.0004106504,
0.0054299668,
-0.0483732857,
0.0356324837,
0.0810754076,
0.0186495464,
-0.0679935515,
-0.0459877066,
0.0126611423,
0.0095321629,
0.0591864027,
0.0466314144,
-0.0567725971,
-0.0254429709,
0.0114003159,
0.0505585447,
0.0621791482,
-0.0252594296,
0.0065917908,
-0.1312347353,
0.0244677179,
-0.0003521125,
0.0291567165,
0.0116065582,
-0.0407335274,
0.020871399,
-0.0018448875,
-0.0589533448,
0.0052701742,
-0.0223378614,
0.0079207169,
0.0156441331,
0.0155387949,
-0.0229940917,
0.005138027,
-0.047651913,
0.007680614,
0.0245241635,
-0.013375381,
0.0057001486,
0.0124839619,
0.0907294229,
0.0260164943,
0.0160458833,
-0.0272913165,
0.0383363441,
-0.0189557523,
-0.004956081,
0.0155588798,
-0.0213257279,
0.0029457274,
-0.0074565709,
-0.0252290536,
0.0675369129,
-0.0223469101,
0.0346261263,
-0.0303303059,
-0.0304477867,
-0.0169684254,
0.0541578755,
0.0249369852,
0.0167410355,
0.0185376965,
0.0021908227,
-0.007009333,
0.0067131664,
-0.0187478773,
-0.0219729859,
-0.0346794762,
-0.0011768979,
-0.0452276133,
-0.0138015077,
0.0557458214,
0.0464525111,
0.0112841735,
0.0151997032,
-0.0284835566,
0.0075528109,
-0.0955269039,
0.063034609,
-0.0390635021,
0.023308903,
-0.0536061339,
0.002665645,
0.030814141,
-0.048549775,
-0.0111984229,
-0.0001675455,
-0.0293404162,
-0.0549750254,
-0.0405796394,
0.0225388482,
0.0217793267,
0.0353334546,
-0.0278096721,
-0.0840670988,
0.0089041609,
0.0571822226,
-0.040649116,
-0.0302545764,
-0.0142358057,
-0.0242574662,
-0.0313912369,
0.0513280332,
-0.0672677532,
-0.0261620209,
-0.0176156517,
0.0562037304,
-0.0368404277,
-0.096704863,
0.0503905714,
0.0370678194,
0.0448796898,
0.0361650214,
0.0450317971,
-0.0806209594,
0.0096502956,
0.0365613736,
-0.0456233695,
-0.042375993,
0.0387778915,
0.0114148287,
-0.0521389022,
0.0341687016,
0.017855132,
0.0330317058,
0.0655625835,
-0.0594480634,
-0.0007998539,
0.0101906313,
-0.0174456108,
0.0252009854,
0.036794804,
-0.0152128972,
0.0007159974,
0.0145489192,
0.0410932116,
-0.0311043113,
0.0583830662,
-0.0051558944,
0.0013269786,
-0.0320690908,
0.0234143995,
-0.0380135924,
-0.1126475036,
-0.0331756659,
0.0256778374,
0.0555553325,
-0.0448400229,
0.0219904147,
-0.0257862024,
-0.0337973759,
-0.054227259,
0.0011586136,
-0.0012856594,
-0.0260309801,
0.0221546274,
0.0465484858,
-0.0102794943,
0.0126370247,
0.0513099469,
-0.0331879593,
0.044026155,
-0.0293483287,
-0.0602614395,
-0.0292237736,
-0.0383250788,
0.0771308988,
0.0300118923,
-0.0338723026,
0.0074169007,
-0.0077366713,
-0.0662845001,
-0.040322233,
-0.0059184963,
-0.0412424207,
-0.000400801,
0.0386304408,
0.1175382137,
0.02175349,
-0.0024728414,
-0.0458694622,
0.0100300843,
-0.0113449683,
0.0169237144,
0.0020721452,
-5.151222691e-33,
-0.0404954702,
-0.02060223,
0.017784873,
-0.0216033049,
0.0122967269,
0.0034557309,
-0.0094310688,
0.0114068035,
-0.0464144833,
-0.0184756797,
-0.0078289276,
0.001908849,
0.0146944067,
0.0028972579,
0.0386736654,
0.0188529454,
0.0467872843,
0.0240106396,
-0.0070850002,
0.0143077206,
0.0743675753,
0.0484136492,
0.0418931097,
-0.0820339546,
-0.0004955176,
0.013038219,
0.0338886231,
0.0437657647,
0.0334720127,
0.0775211379,
0.0096208826,
0.0420317091,
0.006164962,
-0.0092730122,
0.0002035369,
-0.0411420837,
-0.0616648793,
-0.0144204553,
0.0103666605,
0.003772587,
0.0027257693,
-0.0025262693,
0.0404853188,
-0.0297518466,
0.0231659506,
-0.0020585386,
0.0396032073,
-0.0104303639,
-0.0418312512,
-0.0274040345,
0.0059401426,
0.0192425288,
0.0122368168,
-0.0426188447,
0.0450045466,
0.0185793787,
0.0426101647,
0.0187966712,
-0.0307583567,
0.0137891583,
-0.0245974809,
0.0524203517,
0.0918993875,
0.0436511785,
-0.007867652,
0.0319874845,
0.0228398573,
-0.0272931792,
0.0154484082,
-0.0298910514,
-0.0074340887,
0.1076743826,
0.0171756539,
0.0103744175,
-0.0358282477,
0.0235097148,
0.0300592724,
0.0485823378,
0.0222183373,
0.0805485398,
0.0315355211,
0.0778457969,
-0.0435741842,
-0.013816678,
-0.0299832709,
-0.0299066231,
-0.0568814836,
-0.0278480537,
0.0303346124,
0.017890675,
0.0059573529,
-0.0138158081,
0.0331797861,
-0.0261833053,
0.0037302037,
0.0442248583,
-0.0116675515,
0.0108713768,
-0.0037108711,
0.0331561565,
-0.0028270003,
-0.0144845797,
0.0036723192,
-0.0562693514,
-0.0318186358,
0.0034518165,
-0.0033949919,
0.0123987803,
-0.0358997658,
0.0124765215,
-0.0609216839,
-0.0552181415,
-0.0339095257,
0.0078264149,
0.0586461648,
-0.0015578966,
-0.0032073045,
0.035254024,
0.0244818944,
0.0057280203,
-0.0339295268,
-0.0475860052,
0.0037582247,
-0.045649983,
0.0224446952,
0.0361014083,
-0.0420533717,
0.0172041655,
0.0417851992,
0.0129593862,
-0.0068880664,
-0.0115007805,
2.544e-7,
0.0462852493,
0.0138380546,
0.0151061099,
0.0447403044,
-0.0456451625,
0.0339834392,
-0.0345229432,
0.0520222299,
0.0366749056,
0.0143455099,
0.0027733033,
0.034296304,
0.015123358,
0.0370108932,
-0.0439569168,
-0.0057434901,
-0.0512779914,
-0.0214199778,
-0.068086274,
-0.0252782945,
0.0315756798,
0.0478475206,
0.0214542169,
-0.0170885921,
-0.0113577135,
-0.1137567163,
0.0460853875,
-0.0619872138,
-0.030973915,
0.0085542621,
0.0311607216,
0.0551148094,
0.0096481191,
-0.0075370814,
0.0074606664,
0.0476086922,
0.0583356507,
0.0296528488,
-0.0047116964,
-0.0399892554,
-0.0030765124,
-0.0595610477,
0.0116207954,
-0.0646302402,
0.0609198138,
-0.0111797685,
0.0247717518,
0.0375124589,
-0.0028661103,
-0.0207576863,
-0.0060267579,
-0.0115396949,
-0.0129858796,
0.0008195856,
0.0035953526,
-0.0181311872,
-0.025612127,
-0.0744445175,
-0.0291492045,
0.0199521016,
-0.0590067767,
-0.0219432358,
-0.0402664468,
0.0027508372,
-0.042922534,
-0.0140333185,
0.0138936592,
2.129994616e-34,
0.0172300115,
0.046644792,
-0.0091995671,
0.0181549825,
0.0533076674,
-0.024368884,
-0.0377183594,
0.0254052021,
-0.0048754211,
-0.0850600228,
-0.0299530737
] |
http://arxiv.org/abs/2405.16064v1
|
http://arxiv.org/pdf/2405.16064v1
|
Keypoint-based Progressive Chain-of-Thought Distillation for LLMs
| 2024-05-25T05:27:38 | 2024-05-25T05:27:38 |
Kaituo Feng, Changsheng Li, Xiaolu Zhang, Jun Zhou, Ye Yuan, Guoren Wang
|
cs.CL
|
cs.CL
|
Chain-of-thought distillation is a powerful technique for transferring
reasoning abilities from large language models (LLMs) to smaller student
models. Previous methods typically require the student to mimic the
step-by-step rationale produced by LLMs, often facing the following challenges:
(i) Tokens within a rationale vary in significance, and treating them equally
may fail to accurately mimic keypoint tokens, leading to reasoning errors. (ii)
They usually distill knowledge by consistently predicting all the steps in a
rationale, which falls short in distinguishing the learning order of step
generation. This diverges from the human cognitive progression of starting with
easy tasks and advancing to harder ones, resulting in sub-optimal outcomes. To
this end, we propose a unified framework, called KPOD, to address these issues.
Specifically, we propose a token weighting module utilizing mask learning to
encourage accurate mimicry of keypoint tokens by the student during
distillation. Besides, we develop an in-rationale progressive distillation
strategy, starting with training the student to generate the final reasoning
steps and gradually extending to cover the entire rationale. To accomplish
this, a weighted token generation loss is proposed to assess step reasoning
difficulty, and a value function is devised to schedule the progressive
distillation by considering both step difficulty and question diversity.
Extensive experiments on four reasoning benchmarks illustrate our KPOD
outperforms previous methods by a large margin.
|
Keypoint-based Progressive Chain-of-Thought Distillation for LLMs: Chain-of-thought distillation is a powerful technique for transferring
reasoning abilities from large language models (LLMs) to smaller student
models. Previous methods typically require the student to mimic the
step-by-step rationale produced by LLMs, often facing the following challenges:
(i) Tokens within a rationale vary in significance, and treating them equally
may fail to accurately mimic keypoint tokens, leading to reasoning errors. (ii)
They usually distill knowledge by consistently predicting all the steps in a
rationale, which falls short in distinguishing the learning order of step
generation. This diverges from the human cognitive progression of starting with
easy tasks and advancing to harder ones, resulting in sub-optimal outcomes. To
this end, we propose a unified framework, called KPOD, to address these issues.
Specifically, we propose a token weighting module utilizing mask learning to
encourage accurate mimicry of keypoint tokens by the student during
distillation. Besides, we develop an in-rationale progressive distillation
strategy, starting with training the student to generate the final reasoning
steps and gradually extending to cover the entire rationale. To accomplish
this, a weighted token generation loss is proposed to assess step reasoning
difficulty, and a value function is devised to schedule the progressive
distillation by considering both step difficulty and question diversity.
Extensive experiments on four reasoning benchmarks illustrate our KPOD
outperforms previous methods by a large margin.
|
[
0.0072573572,
0.0358476155,
0.00042282,
0.0113882832,
-0.0528560504,
0.033368513,
0.0065284558,
-0.0247657187,
-0.0002636686,
-0.0333657451,
0.0184665285,
0.0023770984,
-0.0206509382,
0.0410972312,
0.0287720207,
-0.0679537728,
0.0363618508,
-0.0717336908,
0.0114614414,
-0.0235838406,
-0.037855193,
0.0166211016,
-0.0462655276,
0.0300317276,
0.0248788781,
-0.0613101237,
-0.0406846553,
0.006962039,
0.0361020751,
-0.0383777283,
0.0071485857,
0.0456035659,
0.026739344,
-0.0052922368,
0.0000022861,
-0.0798445866,
-0.0136870211,
-0.004748831,
-0.0713324472,
-0.0174681656,
-0.0327135772,
0.0493407026,
0.0259456336,
0.0343757421,
-0.034377832,
0.019773135,
0.0235016905,
0.039650552,
0.0812041834,
0.0552186184,
0.016594734,
-0.0558716394,
0.0167690832,
-0.0007068468,
0.0185962059,
-0.0223856084,
0.0101625388,
0.028485328,
-0.0029837098,
-0.0035425369,
-0.0135282613,
0.0165887941,
0.0363735892,
-0.0035310285,
0.0097835083,
0.0255348217,
-0.0139769856,
-0.0310037825,
-0.0055210316,
0.0722425506,
-0.0457466803,
-0.0298343878,
0.0079420041,
-0.0124454387,
0.0226181857,
0.008397514,
-0.0783395246,
0.0363053158,
0.0020348839,
0.0315332562,
0.0260965433,
-0.0100751426,
0.0270923618,
0.0139359478,
-0.0574914068,
0.0508868955,
-0.0121058049,
-0.031931974,
0.035386458,
-0.0124186082,
-0.0450781211,
-0.0556784943,
0.0147182653,
0.0763076097,
-0.0090588303,
0.0303241871,
0.0000089617,
0.0029880626,
0.0059882049,
0.0026134751,
0.0241313055,
0.0344867967,
-0.00493393,
0.0339614116,
-0.0557704568,
0.0529502295,
0.0177617501,
0.0813461468,
-0.0359256491,
-0.0028259726,
-0.073531501,
-0.014978352,
-0.0140932258,
0.0949546099,
-0.0225667525,
-0.0008141294,
-0.0219594799,
0.0021762638,
-0.0383493192,
0.0084163593,
0.039885357,
0.0040348112,
0.0231168419,
-0.033824183,
-0.0329244733,
0.0665775836,
-0.0475046895,
0.0142239276,
0.0535363592,
0.0007231251,
-0.0419119969,
0.0128304865,
-0.001590085,
0.0068309568,
0.0202508327,
0.0436035059,
-0.0357632339,
0.0057558292,
-0.0557279252,
-0.0421948433,
-0.0019617635,
-0.0925051868,
0.024496872,
-0.0873108283,
0.0224438813,
0.0080057913,
-0.0106078414,
-0.0335747227,
0.0159792695,
0.0268446431,
-0.0031354704,
0.0521045886,
0.0253017098,
-0.0082693314,
0.0022050936,
0.0329056196,
-0.0828223005,
0.0776424482,
0.0267578848,
0.0324009657,
-0.0098172054,
-0.0008014945,
-0.0254325662,
-0.0011744346,
0.0329775661,
-0.0020132323,
-0.0162651893,
-0.00225072,
-0.0015234991,
0.0378736109,
0.0247371458,
0.0515648238,
-0.0532688051,
0.039288111,
0.0334782638,
0.0636174902,
0.0219408143,
-0.0041264119,
0.048804272,
0.0063252617,
0.0020060814,
0.0452600308,
-0.0574326552,
0.0159477536,
-0.004234571,
0.0119826542,
0.0121678626,
0.0288523398,
-0.0050541549,
-0.07018473,
0.0111953178,
0.0167395957,
-0.0020551565,
0.0359894261,
0.0374145657,
0.0221250989,
-0.0390429944,
0.0891545564,
0.0093571423,
-0.0414653979,
0.0046955855,
-0.0670279264,
0.023887936,
0.0201572571,
0.019363746,
-0.006173477,
-0.0261135176,
-0.0080006877,
0.0370819978,
0.0034538412,
0.0012414833,
-0.0368546546,
0.0070156404,
-0.0641697124,
-0.0328432918,
-0.0270656142,
0.0120542906,
-0.0192698687,
0.0061685345,
0.0026128655,
0.0119228773,
-0.0704333708,
0.0180110764,
0.0217394661,
-0.0404202975,
-0.0359648541,
-0.0194694623,
0.0081111137,
0.0077352296,
0.0115982825,
0.0347502381,
0.0205397364,
0.023564864,
-0.051977668,
-0.0142429126,
0.0049458067,
-0.0828330144,
-0.0770154521,
0.0577168129,
-0.0440658405,
-0.0237527099,
0.0330408402,
-0.0668039843,
-0.0633571446,
0.0232083127,
0.0023546931,
0.0112513294,
-0.0092495484,
0.0356535502,
0.0126938112,
-0.0256836973,
-0.0563602746,
0.0905054882,
-0.0314234234,
0.0123083657,
-0.0225232095,
0.0064359335,
0.001535538,
-0.0064757718,
0.0227968823,
0.016647594,
-0.0125182942,
0.0002891356,
-0.0158209726,
0.0116828671,
-0.0738794282,
-0.0283211693,
-0.0482015423,
0.0111306915,
-0.0136066647,
0.0263258889,
0.0132380826,
-0.0028573647,
0.0091866413,
-0.0053765601,
-0.0533623844,
-0.0277147181,
-0.0094976677,
0.000156579,
0.0174666326,
0.0096245082,
-0.0099527286,
0.0002687083,
-0.0157845654,
-0.0349453054,
0.0476966761,
-0.0329368934,
0.0036322409,
-0.0425852649,
0.0321419612,
-0.0178169068,
-0.0101120993,
0.0002523265,
0.0005457299,
0.0504258126,
-0.0260436628,
0.0544901267,
0.0692394003,
-0.0016653128,
0.0168169029,
0.0574297942,
-0.0209093727,
-0.079122588,
0.0466407798,
-0.0137811638,
0.0262287594,
-0.0270683877,
-0.0438403226,
-0.0068006385,
0.0623410195,
-0.0223244093,
-0.0034155576,
-0.0589815862,
0.0193718076,
0.0032010567,
-0.0217689313,
-0.0331135988,
-0.0033357479,
-0.0291641261,
0.0207184702,
-0.0279775187,
-0.0207316969,
0.0767762139,
-0.0129034128,
0.0216756575,
-0.0484205224,
0.015567882,
-0.0262558311,
0.0590918995,
-0.017568212,
-0.0149440337,
-0.0035119092,
-0.0082342569,
-0.0850415379,
-0.0053695659,
0.0220275056,
-0.037696857,
0.0848790035,
-0.0846310332,
-0.0388967209,
-0.0138253281,
-0.0250931177,
-0.0571643151,
0.0367563218,
-0.0235604681,
0.0242464542,
-0.0042232154,
-0.0002116304,
-0.0025978864,
-0.0409595221,
0.0252889041,
0.0284065753,
-0.0081437808,
-0.0075995321,
-0.013529825,
0.030645268,
0.0580954328,
0.0323139727,
0.0447681695,
-0.0395944528,
0.0747255534,
0.0630044043,
0.0917186961,
-0.0222204253,
-0.0004462016,
0.0291370992,
0.0254153181,
-0.0277304612,
-0.0014208768,
0.035456907,
0.0770668313,
-0.0681844801,
-0.0120914932,
0.1130175218,
0.030765906,
-0.0134234531,
0.0081631318,
-0.0480563454,
0.0845423266,
0.0279026311,
0.0059551178,
-0.0690877587,
-0.0462182611,
0.019406464,
0.049848292,
-0.0109167257,
-0.0488456562,
-0.0038137909,
-0.1079255044,
0.0360818915,
0.0177961458,
-0.0256641749,
-0.0066905445,
-0.0521642342,
-0.0233858451,
-0.0012854404,
0.0064827837,
0.0159366243,
-0.007704657,
0.0219684318,
0.0448949784,
-0.0421079285,
-0.021869136,
0.0322323591,
-0.01642875,
0.0125088608,
0.0144351283,
0.0245894566,
0.0202168468,
-0.0844248012,
0.0532506183,
-0.0139487078,
0.0004911809,
-0.0132495798,
0.0061686481,
0.0174086932,
0.011445703,
-0.008872251,
0.0780042633,
-0.0299252402,
-0.0541725457,
0.0246036984,
0.1109329462,
-0.0209963843,
-0.0351016186,
-0.0199662223,
-0.0062767835,
-0.013587364,
0.0688176379,
0.0554161668,
-0.0258627925,
-0.0569896996,
-0.0376093499,
-0.0190073345,
-0.065883711,
-0.0051192795,
0.034643244,
0.0171864741,
0.0485134684,
-0.0497501716,
-0.0152909094,
0.0344617069,
0.0660960823,
-0.0321213305,
0.0283181574,
-0.0149651794,
0.0133198407,
-0.032259874,
0.0322491303,
-0.0340232067,
-0.0018373905,
-0.0376566611,
0.0080126505,
-0.0192304309,
-0.018433867,
-0.024931591,
0.0128485123,
-0.0988381654,
-0.0670609549,
0.0030966122,
0.0624079742,
0.0271477439,
0.0768074915,
0.0137124406,
-0.0098279277,
0.0548921637,
-0.0156120034,
0.0339137241,
-0.0130730914,
0.0495879576,
0.0105958031,
0.0016767118,
0.0295915138,
-0.0364212878,
-0.0020928753,
0.0123635279,
0.0893384218,
-0.0341474153,
0.0196539611,
0.0358184539,
0.0190387946,
0.0353126712,
-0.000738234,
-0.0026882493,
-0.0278392117,
0.0485877916,
0.001609291,
-0.0015692257,
-0.0451004356,
0.0082365498,
0.0172837153,
-0.0141727524,
0.0225795899,
-0.035358388,
0.0023979158,
0.0214055479,
0.007468293,
0.0313265771,
-0.0071036988,
0.0359590463,
0.0455164388,
-0.0333773345,
-0.0285969265,
-0.0117702363,
-0.0061648344,
0.0341803059,
-0.0026384061,
0.0866386667,
0.0462448634,
0.0171019137,
-0.0005663449,
0.0040208935,
-0.0613931715,
-0.0745037571,
-0.0086649759,
-0.0250611603,
0.0280140117,
0.02117352,
0.0342467874,
-0.0296498239,
-0.0031572015,
0.0074492004,
0.0299360473,
-0.0124885151,
-0.0283954609,
0.0230093598,
0.0455468632,
0.0264116786,
-0.0046605594,
0.0561573133,
-0.0281294193,
0.0367628261,
-0.0159188174,
0.0322063118,
-0.0336272866,
-0.0629399344,
-0.0203702673,
0.0328663588,
0.0255769975,
0.0052157952,
-0.0642503425,
-0.0922235399,
-0.0059316973,
0.0086096833,
0.0184866488,
-0.0375523418,
0.0475295186,
0.0420850664,
-0.0427706614,
-0.0435615629,
-0.0092288442,
0.0054935645,
0.0246409252,
-0.035334859,
-0.0009544452,
-6.59969034e-33,
-0.0006008997,
-0.0341920033,
0.0530942827,
0.0130007993,
-0.0266417079,
-0.0248324946,
-0.0393284932,
-0.004950942,
-0.0348811895,
0.029949313,
0.0018219079,
-0.0067798258,
0.0112715829,
0.0410111733,
0.0008031338,
-0.0030666068,
0.0375351571,
0.0205357652,
-0.0146828722,
0.0253471434,
0.0246570893,
0.0007944647,
0.0151730869,
-0.0276800077,
0.0055562328,
-0.0075541646,
-0.0087812003,
0.0344082825,
-0.0652639121,
0.0254327804,
0.0125485659,
0.0240203775,
0.01673471,
-0.0604276657,
-0.0262650978,
-0.0097103948,
-0.0810115486,
-0.0606906451,
0.0213904101,
-0.0220530704,
-0.0521165393,
-0.0472388044,
0.0392124951,
0.0048377556,
0.016295474,
-0.0606014878,
0.0154416002,
-0.0186913945,
-0.0248460937,
-0.0134701366,
-0.0304971784,
0.0146274585,
-0.0314137079,
-0.0348500162,
-0.0153661231,
0.0328714438,
-0.0035849451,
-0.0122871902,
-0.0393075608,
0.0053143264,
0.0402738824,
0.0076900544,
0.085670732,
0.028654661,
-0.0131160133,
0.055477839,
0.0251476765,
-0.0246339217,
-0.0262281913,
0.0123204431,
0.0287928432,
0.0550430976,
0.0703293681,
-0.0052074813,
0.0568554662,
0.0225079041,
-0.0996012911,
0.0193836093,
-0.0274378695,
0.0769658238,
-0.0095192567,
0.0122345006,
0.0098416125,
-0.0173887964,
-0.0745458305,
0.0034576808,
-0.0164046157,
-0.0127501367,
0.0331394821,
-0.0069373562,
-0.0444798917,
-0.0276613794,
0.0195184685,
-0.0228367243,
-0.0109840594,
0.0101954564,
0.0143848229,
0.0096909208,
-0.0043715541,
0.0287200511,
-0.0696855113,
-0.010311516,
0.0162729397,
0.0009768305,
0.0466325246,
0.0169342887,
0.0364604294,
0.0058031948,
-0.032322675,
-0.022160247,
-0.0592237227,
0.0254902281,
-0.0020828384,
0.0401243716,
0.0366161726,
-0.0029529668,
-0.0162494779,
0.0092440359,
-0.0087190373,
-0.0251153857,
-0.0117004449,
0.0362258516,
-0.0189970918,
0.025184948,
-0.0182371903,
-0.004396813,
-0.0089817243,
0.0134149613,
0.0228952561,
-0.0528755113,
-0.0034355901,
0.0402213521,
2.923e-7,
0.0151413064,
0.0280570854,
0.013493821,
-0.0044606472,
-0.0113356719,
0.0317388438,
-0.0135435816,
-0.0352202579,
0.0233483687,
-0.0678749532,
-0.018915195,
0.0618312918,
0.04777411,
0.0450446978,
-0.1166574955,
-0.0018823529,
-0.0120633403,
-0.0198560413,
-0.0298864897,
0.0031918979,
0.071253784,
0.0327463783,
0.0159860402,
-0.0274383277,
0.0132669052,
0.0095874872,
0.017502455,
0.0131451841,
0.0291395318,
-0.000759409,
0.017647462,
-0.045771718,
0.0260647647,
0.0439595133,
-0.0226606578,
0.0084578022,
-0.0191314258,
0.0398840606,
-0.0019354738,
-0.0643480793,
-0.0208064765,
-0.0151438685,
0.0316259712,
-0.0812035501,
0.0697052553,
-0.0622726157,
-0.0165801365,
-0.0128225544,
-0.023074815,
-0.0260269716,
-0.0015764576,
0.0115028722,
-0.0077093262,
0.0230781045,
0.0010427207,
-0.0197148491,
0.0142052677,
-0.0937337279,
-0.0422149077,
0.0041166036,
-0.020414548,
-0.0441362783,
-0.0462486967,
-0.016036531,
0.0016840438,
-0.0588167422,
-0.0200143103,
2.875688513e-34,
0.0255766846,
-0.0341525301,
0.0381989628,
0.0366746858,
0.0220178068,
-0.006333054,
-0.0340326652,
0.0243604649,
0.0327833295,
-0.0558383763,
-0.0205080099
] |
http://arxiv.org/abs/2405.16057v1
|
http://arxiv.org/pdf/2405.16057v1
|
SPP: Sparsity-Preserved Parameter-Efficient Fine-Tuning for Large Language Models
| 2024-05-25T04:55:27 | 2024-05-25T04:55:27 |
Xudong Lu, Aojun Zhou, Yuhui Xu, Renrui Zhang, Peng Gao, Hongsheng Li
|
cs.CL
|
cs.CL, cs.LG
|
Large Language Models (LLMs) have become pivotal in advancing the field of
artificial intelligence, yet their immense sizes pose significant challenges
for both fine-tuning and deployment. Current post-training pruning methods,
while reducing the sizes of LLMs, often fail to maintain their original
performance. To address these challenges, this paper introduces SPP, a
Sparsity-Preserved Parameter-efficient fine-tuning method. Different from
existing post-training pruning approaches that struggle with performance
retention, SPP proposes to employ lightweight learnable column and row matrices
to optimize sparse LLM weights, keeping the structure and sparsity of pruned
pre-trained models intact. By element-wise multiplication and residual
addition, SPP ensures the consistency of model sparsity pattern and ratio
during both training and weight-merging processes. We demonstrate the
effectiveness of SPP by applying it to the LLaMA and LLaMA-2 model families
with recent post-training pruning methods. Our results show that SPP
significantly enhances the performance of models with different sparsity
patterns (i.e. unstructured and N:M sparsity), especially for those with high
sparsity ratios (e.g. 75%), making it a promising solution for the efficient
fine-tuning of sparse LLMs. Code will be made available at
https://github.com/Lucky-Lance/SPP.
|
SPP: Sparsity-Preserved Parameter-Efficient Fine-Tuning for Large Language Models: Large Language Models (LLMs) have become pivotal in advancing the field of
artificial intelligence, yet their immense sizes pose significant challenges
for both fine-tuning and deployment. Current post-training pruning methods,
while reducing the sizes of LLMs, often fail to maintain their original
performance. To address these challenges, this paper introduces SPP, a
Sparsity-Preserved Parameter-efficient fine-tuning method. Different from
existing post-training pruning approaches that struggle with performance
retention, SPP proposes to employ lightweight learnable column and row matrices
to optimize sparse LLM weights, keeping the structure and sparsity of pruned
pre-trained models intact. By element-wise multiplication and residual
addition, SPP ensures the consistency of model sparsity pattern and ratio
during both training and weight-merging processes. We demonstrate the
effectiveness of SPP by applying it to the LLaMA and LLaMA-2 model families
with recent post-training pruning methods. Our results show that SPP
significantly enhances the performance of models with different sparsity
patterns (i.e. unstructured and N:M sparsity), especially for those with high
sparsity ratios (e.g. 75%), making it a promising solution for the efficient
fine-tuning of sparse LLMs. Code will be made available at
https://github.com/Lucky-Lance/SPP.
|
[
0.0119839087,
0.0879917666,
-0.0037087679,
0.0418272726,
0.0114171831,
0.0018606255,
-0.0252065659,
0.0320178904,
-0.0439927168,
-0.0103142299,
-0.0492546111,
-0.0439422801,
0.0295043662,
0.0000846369,
0.0287307687,
-0.0070030843,
0.0298051164,
-0.0398318358,
-0.020217998,
-0.0082896324,
0.0244417526,
0.0171386357,
0.0135907363,
0.0076117897,
0.0183148384,
0.0177911837,
-0.0038989244,
-0.0080317752,
-0.0117578441,
-0.0027242117,
0.0311674904,
0.0750371292,
0.0145671312,
0.0405085012,
0.0000021325,
-0.0154870385,
0.018532766,
-0.0002383882,
-0.0347052924,
0.0857289433,
0.0627135411,
0.0102800643,
0.0054417439,
0.0147353206,
-0.0095115509,
-0.0231202412,
0.0492293388,
0.0467976667,
0.0538684577,
0.0545830876,
-0.0051509622,
-0.0343080908,
0.0313962884,
-0.0307480786,
-0.0215174053,
-0.0338070318,
0.0369043015,
-0.0442902036,
-0.0317128524,
-0.0292424653,
0.0061376467,
-0.0089032026,
-0.0174226984,
0.0245293751,
-0.02166491,
0.0165992025,
-0.0372485816,
-0.0355992168,
0.001259417,
0.0581825301,
-0.0323848911,
-0.0447343215,
0.021176124,
-0.0086405911,
-0.0142309489,
-0.0341019891,
-0.0477915518,
0.0028035697,
0.0346446969,
0.0351092927,
0.0083572436,
-0.0513750762,
0.0460324027,
0.0256820638,
-0.0492131971,
0.0418613032,
0.042550195,
-0.0046249009,
-0.0258651599,
-0.023110766,
-0.0213462096,
-0.0408035256,
-0.020820437,
0.0532327332,
-0.0210292842,
-0.0007154939,
-0.0520317778,
-0.0517163351,
0.0367016234,
-0.0113368798,
0.0984468758,
0.0448232628,
0.0058897394,
0.0325264186,
-0.0559290461,
0.0018499864,
0.015310484,
0.0211034399,
-0.0290577915,
0.0011323674,
-0.0006953969,
0.0225307867,
-0.0206350256,
0.0696972385,
-0.0158773717,
-0.0100687034,
-0.0303746089,
0.0105354832,
-0.0086436328,
0.0056442129,
0.0214134175,
-0.0067538084,
0.0039449516,
-0.0139262201,
-0.008613403,
0.042604208,
0.0035113078,
0.0420794263,
-0.0234027132,
-0.042127464,
-0.0222837273,
0.0199800041,
0.0152017623,
-0.0176348072,
0.0559934378,
0.0237339456,
-0.0139785707,
-0.062401861,
-0.0545354001,
0.0184442885,
0.0160911866,
-0.0793580562,
-0.0168988928,
-0.062029995,
0.0552724414,
-0.0195999779,
0.0109854853,
-0.0950697809,
-0.0016019661,
0.0172504745,
-0.0701138899,
0.0582845286,
-0.0232930873,
0.0316242613,
-0.0037940033,
-0.0075843544,
0.0364788026,
0.0755999759,
0.0348745957,
0.0468987897,
0.004740281,
-0.031864088,
-0.0533344895,
0.0343031473,
0.0315037444,
0.0492961966,
-0.0400818735,
-0.0203196984,
0.0177678447,
-0.0231167339,
0.0421434529,
0.0623630956,
-0.0230415557,
-0.009495547,
0.0521281287,
0.0468140319,
0.0359553359,
0.0574630722,
0.0304384213,
-0.0196125302,
0.0222944878,
0.0062747165,
-0.0376542732,
0.0051337187,
-0.055866491,
-0.0103803035,
0.027895337,
0.0770868212,
-0.0470065586,
-0.0273328032,
-0.0406701826,
-0.0270800777,
0.0115694124,
0.0334267765,
0.0373205394,
-0.0193767846,
-0.0660153776,
0.0979531407,
0.0681198314,
-0.0685282648,
-0.0348744057,
-0.0606988929,
0.0186289884,
0.0241869967,
-0.0024554946,
-0.0181142408,
-0.037798319,
0.0303377025,
0.0321910046,
-0.0358225927,
0.0326268561,
0.0011464917,
0.0307435654,
-0.0715951249,
-0.0042196224,
0.011883799,
0.0339714959,
0.0098849116,
0.0160155278,
0.0044906633,
0.0265993848,
0.0051103919,
-0.0486968942,
0.0551257022,
-0.0294988044,
0.0056257085,
0.0181607939,
-0.0220269188,
0.031418968,
-0.0055044964,
0.0025150401,
-0.0120062111,
0.0801194087,
-0.0285373498,
0.0246802326,
-0.0713800937,
-0.0244682133,
0.0018378432,
0.0600135401,
0.0678408667,
0.0311170295,
0.0570676588,
-0.0339060016,
-0.0371488817,
-0.003019487,
-0.0169228353,
0.0256066043,
-0.0666926801,
0.0453095622,
-0.0375481844,
0.0619253553,
-0.0154559985,
0.0281953365,
-0.0332049876,
0.0149449455,
0.0021031089,
-0.0168402456,
0.0280068424,
-0.0603085943,
-0.0581464469,
0.0053049223,
-0.0188780501,
0.0385299847,
-0.0083818147,
0.0263906419,
0.0139143327,
-0.037504673,
-0.0753625482,
0.0489173755,
-0.0394248143,
0.0001505394,
0.0091107124,
0.0172669999,
0.0100621348,
0.0018842033,
0.000524371,
0.0281306133,
0.009032744,
0.023385359,
0.0015916041,
0.0133986454,
-0.0446308143,
-0.0433062613,
-0.0383927934,
0.0303259715,
0.0251732077,
-0.0560374744,
0.0012425851,
-0.052833952,
0.0693707317,
-0.0456319451,
0.030239502,
-0.0076779407,
-0.0511182509,
-0.0304289777,
-0.036492113,
-0.0048318324,
0.0399907045,
-0.0209098645,
0.0429388173,
0.1009766608,
0.0416218489,
0.0048987367,
0.0062117991,
-0.0151634738,
-0.0108841434,
-0.0252401289,
-0.0160358902,
-0.0138906138,
0.0665719584,
-0.0437712036,
-0.0172373224,
-0.0432244055,
-0.0272646714,
0.0134426448,
-0.0129326899,
-0.0120827053,
0.0115810055,
-0.01616746,
0.0242587924,
-0.0149021447,
-0.054116156,
0.0262802839,
-0.0613646358,
-0.0014452966,
0.00437816,
-0.0500048883,
-0.0633240417,
-0.0076367147,
0.0280172639,
0.0112803346,
-0.0608465001,
0.0083184186,
-0.053035371,
-0.0246868636,
0.0055341059,
-0.0205273833,
0.0491653197,
-0.0198972914,
0.0294359028,
-0.0408958495,
-0.0313820317,
-0.0014488768,
-0.0354303047,
-0.0272422582,
0.0296664722,
0.0255960878,
-0.0366793871,
-0.0140389018,
0.0226240475,
-0.0064290767,
0.0269224495,
0.0117017347,
0.0333664976,
-0.023289213,
0.0103998911,
0.0550879091,
0.0458007157,
0.0315056555,
-0.0339584649,
-0.0173625108,
0.0160663165,
0.0117046339,
-0.0161902308,
-0.0426626466,
0.0079806298,
0.0610385947,
0.0157883745,
-0.0062261396,
0.0253746975,
0.0315083712,
-0.051474262,
0.0402477011,
0.0832447112,
0.0150033524,
-0.0397861674,
0.0174531899,
0.008858541,
0.0981984809,
0.0371712036,
0.0289662313,
-0.0367996804,
-0.0226776246,
0.012781458,
-0.0021739167,
0.0171521939,
-0.030825777,
-0.0125753153,
-0.0924422815,
0.010749151,
0.014272782,
-0.0177069604,
0.0038512626,
-0.0465825349,
0.0054125232,
-0.0261635799,
0.0222329777,
0.0213163774,
-0.0428690016,
-0.0132658323,
0.0162535887,
-0.0253510103,
-0.0665435493,
-0.0009603081,
0.00190019,
0.0371705815,
0.0017189137,
0.0141207455,
-0.0232770089,
-0.0589306131,
0.0593250692,
-0.0231372491,
-0.0380607024,
-0.0334643573,
-0.0004858179,
0.0366044901,
-0.0162302516,
0.0250687636,
0.013222592,
-0.0057752472,
-0.019734947,
-0.0135726761,
0.0865332782,
-0.0035136279,
-0.0277376547,
0.0184015539,
-0.0462257825,
-0.0487869605,
0.0623978674,
0.0279387888,
0.003233213,
-0.0385984331,
0.0072787344,
-0.0479286127,
-0.0158173461,
-0.0175405256,
0.018564146,
-0.0153401839,
-0.005066738,
-0.0103768352,
0.0013966177,
0.0475961603,
0.0235559959,
-0.0290040206,
0.0487285331,
-0.038458813,
0.001604129,
-0.0302708242,
0.0237659737,
-0.0461813472,
0.0112568671,
-0.0602772683,
-0.0145586664,
0.0208211951,
0.0106956279,
0.0027550983,
-0.0236859173,
-0.1307198107,
-0.0532154515,
-0.0202223659,
0.0499802381,
0.0150430426,
0.0151564833,
0.0150473053,
-0.0801340789,
0.0488929488,
0.0165646095,
-0.0384280011,
0.0751445219,
0.0131243663,
-0.0781555399,
0.0001700058,
0.0396169387,
-0.0407025963,
0.002111204,
0.0073963488,
0.0321341753,
-0.013180526,
-0.0253599454,
-0.0052660778,
0.0434012264,
-0.0063596903,
-0.0123027125,
-0.0340353437,
-0.0248641577,
0.0448278338,
-0.0084694829,
0.0259732325,
-0.022684142,
-0.0151671488,
-0.006788495,
0.0346909985,
-0.0063238079,
0.0192656387,
0.0169950947,
0.0614111461,
-0.0267843474,
0.0403325185,
0.0012102065,
-0.0278989002,
0.0573925152,
0.0194721818,
-0.024107486,
0.0155625818,
-0.0022549203,
0.0006292002,
-0.0317922235,
-0.0333587751,
0.0023002743,
0.0204202645,
-0.0999696478,
0.0676336735,
-0.0167454835,
-0.0739004835,
-0.0301584397,
0.016814502,
0.0659687966,
-0.0061352472,
0.0318821594,
0.0026950159,
-0.0334502161,
-0.0329307988,
0.0038484207,
-0.0086577823,
0.0159545802,
0.0324939899,
0.0273349304,
-0.0164898951,
0.0415652171,
0.080821,
0.0524853654,
0.0642652884,
0.0012802137,
0.0069568404,
-0.030360166,
-0.0259589739,
0.0105845397,
0.0548984855,
-0.0270566531,
-0.0210703053,
-0.016793441,
-0.0889299735,
0.0171759687,
-0.0260004047,
-0.0070365686,
-0.0340767801,
0.0353005677,
0.0366649032,
-0.0026227005,
0.0604625605,
-0.0094691347,
-0.0737214163,
0.0214800704,
-0.0301693603,
0.0195018426,
-6.791114453e-33,
-0.0053180843,
0.0157970488,
0.0305948593,
0.0111481128,
-0.032112468,
-0.0112211024,
-0.0375607312,
-0.012926531,
-0.0303084571,
0.000653763,
-0.031079445,
-0.0072913556,
-0.0045395992,
0.0100134695,
-0.0297152828,
0.0157649759,
0.0686949119,
0.0091065364,
0.0050762976,
0.0111313602,
-0.0129743982,
0.031835895,
0.0134017104,
-0.0174569059,
-0.0537955053,
0.014789138,
-0.0238791946,
0.0177249108,
-0.0028447772,
0.0428564288,
-0.0472899564,
0.0347014405,
0.0253920741,
0.0128694689,
-0.0222885609,
-0.0141811864,
-0.0557792895,
-0.0140238032,
0.0460248068,
-0.0188896526,
0.0061993138,
-0.047540687,
0.0438463651,
0.0202098843,
-0.0429667048,
-0.0150822615,
0.0042417366,
-0.0352823585,
-0.0562677532,
-0.0168424826,
-0.0426910371,
0.0073272805,
0.0106457584,
-0.0222939774,
0.031951081,
-0.0554831363,
-0.018375041,
0.0429978818,
-0.0364159159,
-0.013722959,
0.0440061353,
0.0838804618,
0.045608975,
0.0307627972,
-0.0218664464,
0.0196307097,
-0.0106589058,
0.0254133232,
0.0030867574,
0.0273985621,
0.0752124786,
0.0493290126,
0.0431550033,
0.0135878166,
0.0389866419,
-0.0428068303,
-0.1015372127,
0.0144424774,
-0.0036128082,
0.070272468,
0.0137841115,
0.0616637841,
0.035266418,
-0.0319819711,
-0.0518743992,
-0.000108109,
-0.049428001,
-0.0051136366,
-0.0037608745,
0.0386715159,
-0.025599055,
0.0051681427,
0.0141931875,
-0.046342805,
-0.0225616284,
0.0383838825,
-0.0320851617,
0.0076266415,
0.0144149754,
0.0127627226,
-0.0450534038,
-0.0583696291,
0.0012739325,
-0.0363103896,
0.018449001,
-0.0109229861,
0.0467756726,
0.0390348844,
-0.0635638535,
-0.0055970899,
-0.0177206919,
0.006803947,
0.0438692495,
0.0729426071,
0.0152534125,
0.0103816465,
0.0114431176,
0.0691734105,
0.0063098753,
-0.0034657896,
-0.0328499451,
-0.0296302177,
0.0360063724,
0.0165767409,
-0.024100285,
0.0095189754,
-0.0279905815,
0.0279089455,
0.0482257605,
-0.0359339267,
-0.0271230806,
-0.0063775168,
2.761e-7,
0.0081764022,
0.0106988065,
-0.0412737951,
-0.0351578332,
-0.0089474153,
0.0082895793,
-0.0231394842,
-0.0123520792,
-0.0314539969,
-0.0095661217,
0.0026449903,
0.0353997573,
0.0489257462,
-0.0135809211,
-0.0668607727,
0.025191199,
-0.0731718093,
-0.0085594589,
-0.0535748601,
-0.0263082888,
-0.0083800349,
0.126174897,
0.0101330392,
-0.0376911089,
-0.0137781594,
0.0065350723,
0.0388175398,
-0.0198006742,
-0.0038203883,
0.0191660114,
0.0480931774,
-0.008402558,
-0.0183208864,
-0.0282864533,
-0.0203143433,
0.041794274,
-0.0368379131,
0.0818485618,
-0.0337095521,
0.0113959424,
0.0327872075,
0.0004268459,
-0.0254723728,
-0.0347388014,
0.0847660303,
0.0048074778,
-0.0298443865,
-0.0317982547,
0.0137554873,
-0.0131368367,
-0.009793316,
-0.0195097122,
0.0174407903,
0.0389077105,
0.0126249949,
-0.0381054692,
-0.0332313552,
-0.0850296915,
0.0192165729,
0.0405035503,
-0.0419101007,
-0.0547500178,
-0.0441272445,
-0.0278595276,
0.0535699688,
-0.0463026688,
0.0370387211,
2.90633838e-34,
0.0356841572,
-0.0322694406,
0.025202632,
0.0560840778,
0.0161202624,
-0.0262269005,
0.0059997994,
0.0398561843,
-0.0173280109,
-0.1084362343,
0.0061150235
] |
http://arxiv.org/abs/2405.16043v1
|
http://arxiv.org/pdf/2405.16043v1
|
Theoretical Analysis of Weak-to-Strong Generalization
| 2024-05-25T03:48:12 | 2024-05-25T03:48:12 |
Hunter Lang, David Sontag, Aravindan Vijayaraghavan
|
cs.LG
|
cs.LG, cs.CL, stat.ML
|
Strong student models can learn from weaker teachers: when trained on the
predictions of a weaker model, a strong pretrained student can learn to correct
the weak model's errors and generalize to examples where the teacher is not
confident, even when these examples are excluded from training. This enables
learning from cheap, incomplete, and possibly incorrect label information, such
as coarse logical rules or the generations of a language model. We show that
existing weak supervision theory fails to account for both of these effects,
which we call pseudolabel correction and coverage expansion, respectively. We
give a new bound based on expansion properties of the data distribution and
student hypothesis class that directly accounts for pseudolabel correction and
coverage expansion. Our bounds capture the intuition that weak-to-strong
generalization occurs when the strong model is unable to fit the mistakes of
the weak teacher without incurring additional error. We show that these
expansion properties can be checked from finite data and give empirical
evidence that they hold in practice.
|
Theoretical Analysis of Weak-to-Strong Generalization: Strong student models can learn from weaker teachers: when trained on the
predictions of a weaker model, a strong pretrained student can learn to correct
the weak model's errors and generalize to examples where the teacher is not
confident, even when these examples are excluded from training. This enables
learning from cheap, incomplete, and possibly incorrect label information, such
as coarse logical rules or the generations of a language model. We show that
existing weak supervision theory fails to account for both of these effects,
which we call pseudolabel correction and coverage expansion, respectively. We
give a new bound based on expansion properties of the data distribution and
student hypothesis class that directly accounts for pseudolabel correction and
coverage expansion. Our bounds capture the intuition that weak-to-strong
generalization occurs when the strong model is unable to fit the mistakes of
the weak teacher without incurring additional error. We show that these
expansion properties can be checked from finite data and give empirical
evidence that they hold in practice.
|
[
0.01912513,
0.0428190082,
-0.0130304787,
-0.0220740717,
0.0085733226,
0.0408481173,
-0.0007408977,
-0.0014892907,
-0.0195049383,
0.0099083465,
0.0070873429,
-0.0210742913,
-0.0282282513,
-0.007989455,
0.0531313606,
-0.0496569611,
0.0491023473,
-0.0263779461,
0.0104626641,
-0.0224160459,
-0.0228198655,
-0.0731986314,
0.008003857,
-0.0096943118,
-0.0346297733,
0.0238914248,
-0.0154155465,
-0.0349982418,
-0.0249061771,
-0.004841215,
0.0216166321,
0.0289677046,
0.0276750121,
0.0580688231,
0.0000016456,
-0.0242334697,
-0.0236662999,
-0.0037401477,
0.0147123206,
0.000909018,
-0.0170403197,
-0.0554081425,
0.0039225179,
-0.0095536904,
0.0024918357,
-0.0739276409,
0.0423145667,
0.0733719543,
0.036898274,
0.0424692705,
0.0012511958,
-0.0597760752,
0.0182050839,
-0.0309215505,
0.0565268472,
-0.0872270167,
0.0409976654,
0.0142548624,
-0.0155184111,
-0.0336559489,
-0.0203095339,
0.0262355972,
0.0376980044,
0.0140290791,
-0.0129777566,
0.0317847654,
-0.047181949,
0.0063959965,
0.0184246842,
0.007016968,
0.0011655916,
-0.0110204509,
-0.0235646274,
0.009592236,
0.049632635,
0.0388020128,
-0.0337799117,
-0.0126910564,
-0.0272197966,
0.0159802064,
0.0049443706,
-0.0396895409,
0.0362670943,
-0.0029225545,
-0.0150720263,
0.0764555708,
-0.0064807627,
-0.0742239952,
0.0151594654,
0.0598942563,
-0.0766038969,
0.0190548059,
-0.0524180606,
0.012470318,
-0.0217731502,
0.0127941538,
-0.0256405249,
-0.0055864328,
0.0037352496,
-0.0209484585,
0.0274533089,
-0.0221750401,
-0.057088308,
0.0475910045,
-0.0372575261,
0.0260733645,
0.0089975204,
0.0882812589,
-0.003878071,
-0.0040125414,
-0.0743936151,
-0.0131491143,
-0.0257685594,
0.0575881861,
0.0589612052,
0.0517123528,
-0.0642875433,
-0.0155989965,
-0.0362740159,
0.0425243415,
-0.0351282954,
-0.0313643627,
-0.0680373386,
0.0191685781,
-0.0353561491,
-0.0038672173,
-0.0402412303,
0.0438303053,
0.017915139,
-0.0332089737,
-0.0039056998,
-0.0314727537,
-0.0103994152,
0.0127477059,
0.0133404499,
0.0498174764,
-0.0076736701,
-0.0190577693,
-0.1011724323,
-0.0241840482,
-0.0395270139,
-0.029773308,
-0.0024369305,
-0.0531492271,
0.0019076772,
-0.0087901317,
0.0061195511,
-0.0057506124,
-0.0052680373,
-0.0479005165,
-0.0464275815,
0.0214498434,
-0.0214789379,
0.0148005104,
0.0157246999,
0.0104259914,
-0.0512717031,
0.0378676206,
-0.0441819765,
-0.0207298305,
0.0228650197,
-0.0068241721,
0.031877961,
0.0177911408,
0.0373964682,
0.026464887,
-0.0016065524,
-0.0148525955,
0.0504980013,
0.0236025825,
-0.0169631727,
0.0343322977,
-0.039216809,
0.0020446535,
0.0263354219,
0.0000739419,
0.0122436397,
0.000523448,
0.0423567183,
-0.0440740846,
0.0386616625,
0.0724807754,
-0.0563955456,
-0.0638207719,
-0.0463629626,
-0.0119890869,
-0.0446893871,
-0.0432523936,
-0.0480177514,
-0.06588386,
0.0065311985,
-0.0012165965,
-0.0313574933,
-0.0212992541,
0.0989654213,
0.0505880527,
0.0445142016,
0.0129548004,
0.0574904382,
-0.0329346582,
0.0216265842,
-0.0593930744,
0.0323303491,
0.0162351523,
0.0113002025,
-0.0652396604,
-0.0071559744,
-0.0012863156,
0.05499845,
-0.0464226454,
0.0881271437,
-0.0174843092,
0.0241003782,
-0.0026094562,
-0.0227659345,
0.0071655558,
0.05983207,
0.0152740004,
-0.0019448908,
-0.0277261101,
0.014549979,
-0.053121943,
-0.0582371429,
0.0551493242,
0.0151931457,
-0.0458834879,
-0.0577130206,
0.0203865618,
0.0287832487,
-0.0184671208,
-0.040270444,
-0.0128675085,
-0.0052449652,
-0.008065599,
-0.0162909236,
-0.0226555541,
-0.0425150692,
-0.0221405569,
0.0146067496,
-0.0772311464,
-0.0269557443,
-0.0043749884,
-0.0469452627,
-0.0523366928,
-0.003247034,
-0.0095270537,
0.030118864,
-0.0256000664,
0.0292753223,
-0.0225764457,
-0.0344722122,
-0.0228636395,
0.0681421384,
0.0274664275,
0.0247032791,
0.0095659792,
0.0010252955,
0.0089849429,
-0.0140635287,
-0.0411404148,
0.0025119002,
-0.0390662663,
0.0211449396,
0.0312453471,
-0.0026846018,
-0.0458222479,
-0.0234028306,
-0.0003104051,
-0.0327430665,
-0.0202870257,
-0.0213789009,
-0.0158750843,
-0.0285740513,
-0.0000093485,
-0.0288221277,
-0.020985093,
-0.0173115786,
0.0092994813,
0.0190765373,
0.0151958875,
0.032385882,
-0.0148273669,
-0.0108157275,
-0.0169138964,
-0.0270244777,
0.0231263749,
0.0376575552,
-0.0166491233,
-0.0339627042,
0.0654929653,
-0.0262034293,
0.0336448029,
-0.0180773437,
-0.0215771496,
0.0244483277,
-0.034079574,
0.0055178087,
0.0624890514,
-0.0433546007,
0.0224646423,
0.0501679964,
0.012097151,
-0.0065981192,
-0.0008702945,
-0.043927554,
-0.1076469049,
-0.0421241783,
-0.0212517846,
-0.0247535165,
0.0376100875,
-0.0084433816,
-0.0398582779,
-0.0619462803,
-0.0111130541,
-0.0324317217,
-0.0342662185,
-0.0195325278,
0.0249465331,
0.0235990677,
0.0037631276,
-0.0479600206,
-0.0489358641,
0.0517060682,
0.0390703864,
0.0258362181,
-0.0078474702,
-0.0257058572,
-0.005208842,
0.0299233105,
0.0501875058,
0.011162811,
0.0235342123,
-0.0327399559,
-0.091819562,
-0.0153306667,
0.0519841909,
0.0050907396,
0.0121515039,
-0.057169836,
-0.0076315962,
0.0143821975,
-0.023515027,
-0.0084695034,
-0.0231814161,
0.021597486,
0.0011426504,
0.0200393293,
-0.0278259199,
-0.0118945967,
-0.0319650099,
0.0064347317,
0.0425830446,
0.0061052707,
-0.0240821447,
-0.0274750516,
-0.0447099321,
-0.0090551078,
-0.065159075,
-0.046296861,
-0.0913529396,
0.0097259097,
0.0197089296,
0.0422356203,
0.0136420587,
-0.0225358717,
0.0282334778,
0.0022659462,
0.0333556719,
-0.0470906012,
0.0327423513,
0.0537378713,
-0.0364849977,
0.0011637495,
0.0663742721,
-0.0064203739,
0.0157337617,
0.0002523848,
-0.0111468183,
0.0476621389,
0.0493547842,
0.009741609,
-0.0464601479,
0.0075207334,
0.0081945816,
0.0468828827,
-0.0154326698,
-0.0554022305,
0.0426358134,
-0.0824254602,
-0.0203173943,
0.0354277752,
-0.0591352396,
-0.0190086942,
-0.0205193087,
0.0116532389,
-0.0037474642,
-0.0052835424,
-0.0084165763,
0.0098424098,
0.0072013247,
0.0738853589,
0.0103207324,
-0.076380834,
-0.0162880961,
-0.0803387463,
0.020472208,
0.0456170514,
-0.0106630279,
-0.0483816713,
-0.0218432359,
0.0230118539,
0.0014383758,
-0.0015844331,
0.0304370448,
0.0803663582,
0.0089713307,
-0.0221518688,
0.0350399166,
0.0620395951,
0.0264775511,
0.0262266677,
0.0427788943,
0.0566373356,
0.0240597408,
-0.0584451817,
0.0181490388,
-0.0389958471,
-0.0032358514,
0.0454944819,
0.0074361223,
-0.0093936156,
-0.023776494,
-0.0162749868,
0.0166014899,
0.0315764472,
0.0122741647,
0.075431481,
-0.0213321075,
-0.0454270765,
-0.0094308928,
-0.0257388223,
0.0247129966,
0.0742802992,
0.0322184674,
-0.01108843,
-0.0161866788,
0.0090995803,
0.0194600485,
0.0206549894,
-0.0112938285,
-0.0251322296,
-0.0358426124,
-0.033511214,
-0.0032113786,
0.0070240847,
-0.0251388699,
-0.0273364186,
-0.020758871,
-0.0697927922,
0.0469714925,
0.0710198432,
0.0275595859,
0.0455788858,
0.0099805119,
-0.0563803129,
0.0507129468,
0.0580708943,
0.0220302325,
0.0174496416,
0.0783683658,
-0.0622112118,
0.0186735224,
0.0142426742,
-0.0524608344,
0.0261622686,
0.0424244776,
-0.0120337037,
-0.0179659389,
0.0343348868,
0.050348971,
0.0163570344,
0.0397943445,
0.0063767857,
-0.0018741075,
-0.0109286774,
0.0026896922,
0.0252210684,
-0.0131567828,
-0.0098640602,
0.0154829742,
0.0574456118,
0.0506993569,
0.0218941495,
-0.0346806012,
0.0304415729,
0.0837439299,
-0.0264671724,
-0.0220315959,
0.0307590999,
0.0376602001,
0.0571749285,
-0.0005995262,
-0.0127071794,
0.0012748218,
0.0177110974,
0.0164273977,
0.0222142059,
0.0194866844,
0.0580897704,
0.0399924144,
-0.0011554318,
-0.061310295,
-0.1216799617,
-0.0538098514,
-0.0083695566,
0.0516560338,
0.0001739854,
-0.0089701647,
-0.0296405572,
-0.0503607392,
-0.0135709876,
-0.0005520317,
0.0101307919,
-0.0146176275,
-0.0455269329,
0.0046089212,
0.003017775,
0.0072600679,
0.0271289907,
0.0654212311,
0.0690961555,
0.0930308104,
-0.0725027695,
-0.0213990379,
-0.0270277075,
0.006561399,
0.0762317628,
0.011006074,
0.0182741135,
0.0112165697,
-0.0125005152,
-0.0666585565,
0.0040777768,
0.010503117,
0.0401353091,
0.0277327355,
-0.00127814,
0.0824306905,
-0.0149709303,
0.0012855139,
0.0008388709,
-0.0208724849,
0.0805159584,
0.0218607746,
-0.010392366,
-5.313397566e-33,
-0.0293543581,
0.0317189991,
-0.0137809813,
-0.0200650916,
0.0255331025,
0.0004802091,
0.0221020915,
0.0169811901,
-0.0051944242,
0.0444237664,
-0.0015334537,
0.0160202403,
0.0085115097,
0.0405058153,
0.0241191313,
-0.0325096473,
0.058158502,
0.0099361641,
-0.012328092,
0.0119089214,
0.0241533034,
0.0348840915,
0.0793916583,
-0.0841531381,
0.0437264293,
0.0062919934,
-0.0169450976,
0.0233732909,
-0.0260648653,
0.019702509,
0.0418377742,
0.0343536623,
0.0033247212,
-0.0153532764,
-0.002396442,
0.0021471055,
-0.0771449804,
-0.0498117208,
-0.0186935104,
0.0059652664,
-0.0177800655,
-0.002534501,
0.0513662547,
-0.0123423589,
0.0106515726,
0.0109765958,
0.036742121,
-0.0559679978,
-0.0284378845,
-0.0365662836,
-0.0126041388,
0.0097298063,
-0.00276649,
0.0176312737,
-0.0332215317,
0.0489444286,
-0.0056462903,
-0.0252390578,
0.0125165796,
0.0175811891,
0.0743305534,
0.0126432786,
0.0136253536,
0.0364033841,
-0.0163632464,
0.0600509234,
0.0476488881,
0.0645304397,
0.012711959,
-0.0694001317,
0.0041658636,
0.0559671856,
0.0331356712,
0.0219444521,
0.0457182787,
0.0142652988,
-0.0621022694,
-0.0245608874,
0.0123910289,
0.0241058506,
-0.0109448107,
0.0122131435,
-0.0821385533,
-0.0451746136,
-0.0529967062,
-0.0270910449,
-0.0046988162,
-0.0028729562,
0.0271833502,
0.0215469413,
-0.0081801107,
0.0151517224,
-0.0154452119,
-0.0299753062,
-0.0415051617,
-0.0274727363,
-0.0292217415,
0.0025642607,
-0.0425833985,
0.0226952359,
-0.0647155643,
-0.0865596235,
0.005781617,
0.0155817689,
0.0317979641,
0.0134148616,
0.1211106479,
0.0269376915,
-0.0433521271,
-0.0252255034,
-0.0032077336,
-0.0377615504,
0.0524266846,
-0.0315643623,
0.0505010374,
-0.0155971469,
0.0004054781,
0.0087377196,
0.0266561471,
-0.0117313778,
-0.0217393171,
-0.0072257896,
-0.0017751366,
-0.0025759162,
0.016485136,
-0.0372203998,
-0.0087290993,
-0.0139560653,
0.0431900695,
-0.0587946549,
0.0009281386,
-0.0059312102,
2.454e-7,
-0.0046620523,
-0.0210709851,
-0.0170890633,
-0.0214733332,
-0.0176822059,
0.0280672684,
-0.0351799317,
0.015685197,
-0.0083207162,
-0.0380445383,
0.0315630287,
-0.0056924694,
0.0186766777,
0.0282086022,
-0.0251197293,
-0.0296785198,
-0.0461573601,
-0.0021952514,
-0.0207273681,
0.0428717099,
0.0738548487,
0.0809398368,
0.0499628484,
-0.0255118348,
0.0221714452,
-0.0855094567,
0.0409803391,
-0.0118272435,
0.014163441,
0.0239171945,
0.0758176968,
0.0338118486,
0.0259440169,
0.0697247162,
-0.007724727,
0.0142986448,
0.0128794573,
-0.0300177578,
0.0008617457,
0.0780471638,
-0.019363638,
0.0026817797,
-0.0041577923,
0.00396429,
0.054889448,
-0.0932245106,
-0.01776563,
0.0253044106,
-0.0107118757,
-0.0203025956,
-0.017058285,
0.0035839472,
0.0458582379,
-0.0116436407,
-0.0091832681,
0.0140204839,
-0.0190307256,
-0.0245851818,
0.0128986305,
0.0312250145,
-0.0167263038,
-0.0794081315,
-0.0043811933,
0.0465651192,
0.0055561699,
-0.052434437,
0.0326970592,
2.092093647e-34,
0.0625517145,
-0.0236245804,
0.036784403,
-0.0393022262,
0.0449556112,
-0.0426293947,
-0.0042713513,
0.0073200851,
0.0238235593,
0.0018123271,
0.0281880796
] |
http://arxiv.org/abs/2405.16042v1
|
http://arxiv.org/pdf/2405.16042v1
|
Incremental Comprehension of Garden-Path Sentences by Large Language Models: Semantic Interpretation, Syntactic Re-Analysis, and Attention
| 2024-05-25T03:36:13 | 2024-05-25T03:36:13 |
Andrew Li, Xianle Feng, Siddhant Narang, Austin Peng, Tianle Cai, Raj Sanjay Shah, Sashank Varma
|
cs.CL
|
cs.CL
|
When reading temporarily ambiguous garden-path sentences, misinterpretations
sometimes linger past the point of disambiguation. This phenomenon has
traditionally been studied in psycholinguistic experiments using online
measures such as reading times and offline measures such as comprehension
questions. Here, we investigate the processing of garden-path sentences and the
fate of lingering misinterpretations using four large language models (LLMs):
GPT-2, LLaMA-2, Flan-T5, and RoBERTa. The overall goal is to evaluate whether
humans and LLMs are aligned in their processing of garden-path sentences and in
the lingering misinterpretations past the point of disambiguation, especially
when extra-syntactic information (e.g., a comma delimiting a clause boundary)
is present to guide processing. We address this goal using 24 garden-path
sentences that have optional transitive and reflexive verbs leading to
temporary ambiguities. For each sentence, there are a pair of comprehension
questions corresponding to the misinterpretation and the correct
interpretation. In three experiments, we (1) measure the dynamic semantic
interpretations of LLMs using the question-answering task; (2) track whether
these models shift their implicit parse tree at the point of disambiguation (or
by the end of the sentence); and (3) visualize the model components that attend
to disambiguating information when processing the question probes. These
experiments show promising alignment between humans and LLMs in the processing
of garden-path sentences, especially when extra-syntactic information is
available to guide processing.
|
Incremental Comprehension of Garden-Path Sentences by Large Language Models: Semantic Interpretation, Syntactic Re-Analysis, and Attention: When reading temporarily ambiguous garden-path sentences, misinterpretations
sometimes linger past the point of disambiguation. This phenomenon has
traditionally been studied in psycholinguistic experiments using online
measures such as reading times and offline measures such as comprehension
questions. Here, we investigate the processing of garden-path sentences and the
fate of lingering misinterpretations using four large language models (LLMs):
GPT-2, LLaMA-2, Flan-T5, and RoBERTa. The overall goal is to evaluate whether
humans and LLMs are aligned in their processing of garden-path sentences and in
the lingering misinterpretations past the point of disambiguation, especially
when extra-syntactic information (e.g., a comma delimiting a clause boundary)
is present to guide processing. We address this goal using 24 garden-path
sentences that have optional transitive and reflexive verbs leading to
temporary ambiguities. For each sentence, there are a pair of comprehension
questions corresponding to the misinterpretation and the correct
interpretation. In three experiments, we (1) measure the dynamic semantic
interpretations of LLMs using the question-answering task; (2) track whether
these models shift their implicit parse tree at the point of disambiguation (or
by the end of the sentence); and (3) visualize the model components that attend
to disambiguating information when processing the question probes. These
experiments show promising alignment between humans and LLMs in the processing
of garden-path sentences, especially when extra-syntactic information is
available to guide processing.
|
[
0.0485829636,
0.0139557803,
-0.064125292,
0.042374,
-0.0612225905,
0.0091717411,
-0.0482949838,
-0.034178365,
-0.0224596206,
-0.0396245308,
-0.0012471344,
-0.0025481686,
0.0181160755,
-0.0244667139,
0.0047046407,
-0.0489192568,
0.0674645528,
-0.0375263207,
-0.0131866056,
0.0154678114,
0.0125488136,
0.0045479257,
0.009540244,
0.0284097008,
-0.0193850193,
-0.0736348331,
-0.0406792834,
-0.0030952308,
-0.0066926889,
-0.0299607851,
-0.0553963259,
-0.0039907447,
0.0104277134,
0.0499685593,
0.0000020104,
-0.0557632446,
-0.0187381525,
-0.0293902811,
-0.0728854015,
0.0402389541,
0.0039313221,
-0.0019673733,
0.020835245,
-0.0126565276,
-0.0181977544,
0.0275175758,
0.002588131,
0.0176162142,
0.0309489984,
0.0065893941,
0.0148907835,
-0.0353478082,
-0.0001179629,
0.0019780209,
0.0134762041,
-0.0336202867,
0.0199008919,
0.0024981599,
-0.0241219159,
0.0236671809,
-0.0655092597,
0.0229582693,
0.0181265175,
0.018482646,
-0.0177020971,
0.035709668,
0.0171387736,
-0.0329259448,
-0.0215350054,
0.0381837599,
0.0026903988,
-0.0415567644,
-0.0172650535,
-0.0090458095,
-0.0102923876,
0.0602516644,
-0.0645394176,
0.0441421457,
-0.0078704609,
0.0188628249,
0.037604019,
-0.0455209166,
-0.0145238228,
-0.0022839531,
0.0000164086,
0.1110478938,
-0.0043653892,
-0.0512286313,
-0.0160795059,
0.0232163984,
-0.050236024,
-0.0409761928,
0.0111988327,
0.0150787989,
-0.0027126635,
0.0576522723,
0.0022305064,
-0.0446152315,
0.022539204,
-0.0388764478,
0.028922908,
0.0379878432,
-0.0280098375,
0.0504399985,
0.0021639452,
0.0444930829,
-0.0417196341,
0.0466576964,
-0.038857054,
0.0234980881,
-0.0301795658,
-0.0436908714,
-0.0465958901,
0.051306542,
-0.0424704775,
-0.005246541,
-0.0185399111,
0.0228625201,
0.0298359692,
-0.0287016574,
-0.0280730408,
-0.028503513,
-0.038012486,
0.0144372722,
-0.0630689487,
0.0499619916,
-0.0629737899,
0.0568341762,
-0.0284341443,
-0.0664739758,
-0.0519543961,
0.0049636373,
-0.0181675777,
-0.0276332609,
-0.0477491692,
0.0316196568,
-0.0174598619,
-0.0299391467,
-0.0129171535,
-0.0516318567,
-0.0428986847,
-0.0370445475,
0.0055927318,
-0.0173160937,
0.0206584651,
0.0515562817,
-0.0027648686,
-0.0162935108,
0.001000397,
0.0361717418,
-0.0320414007,
0.0486222431,
0.0437846482,
0.0324348696,
-0.0068564424,
-0.0131665012,
-0.0095864944,
0.0326758809,
0.0171716195,
-0.0067493767,
-0.0039033738,
0.0236812476,
-0.005577906,
0.021643227,
-0.0219751522,
-0.007366586,
-0.0307912584,
0.0186436959,
-0.0021139015,
0.043599084,
0.060411416,
0.0462826751,
-0.0322658531,
-0.0098089166,
0.0551919602,
0.0594932958,
-0.0145819392,
-0.0123749916,
0.0279061906,
0.0090959435,
0.015843641,
0.065152131,
-0.0442315377,
-0.0004847088,
-0.0354746319,
0.0431640036,
-0.0838187784,
-0.0068856371,
0.0238372032,
-0.0336285532,
-0.0265062172,
0.0274333451,
0.0376929045,
0.0151813934,
0.0302744303,
0.0097985035,
-0.0263541117,
0.0551726855,
0.0131347748,
-0.0294414703,
-0.0053001978,
-0.0200695079,
-0.0041124313,
0.0578642525,
0.0218720473,
-0.020015059,
-0.0048986897,
0.0230572578,
-0.0052672438,
0.0303955711,
0.0113559831,
0.0076178126,
-0.0112173082,
0.0095069474,
0.0099351397,
-0.0309588481,
0.0199580006,
-0.0260522403,
-0.0530645102,
-0.0029548337,
-0.0073364908,
-0.0666700974,
0.0280606896,
0.0468636826,
-0.0015746148,
0.0226618256,
-0.0657485202,
-0.0076256925,
-0.0021380638,
-0.0058982396,
-0.006577705,
0.0146584557,
0.0311980098,
-0.0044102543,
0.0188408941,
0.0453955643,
-0.0338078961,
-0.039958261,
0.0191804077,
0.0057932953,
-0.0695755929,
0.0042531071,
0.0028585664,
-0.0351461358,
0.0012908939,
-0.0631249622,
0.0296410248,
-0.0168868136,
0.0565502495,
-0.0174842328,
-0.0098146144,
-0.0514209345,
0.060693711,
-0.0161183756,
0.0356645249,
-0.0095922779,
-0.0393449813,
0.0131657105,
-0.0035374535,
0.0325878449,
-0.0036407339,
0.0154211223,
-0.0557057299,
-0.0204670057,
-0.0138998916,
-0.1040593982,
-0.0283456072,
-0.0345100053,
-0.0037921565,
-0.0078741331,
0.0427905843,
0.0385574177,
-0.0189706124,
-0.0019285459,
0.0026169117,
0.0375045277,
0.0515930876,
-0.0337302275,
0.0257607792,
0.0452566929,
-0.0138088064,
-0.0176026896,
-0.0269949492,
0.0104799084,
-0.0645709634,
0.0594596751,
0.0014385406,
-0.0602801293,
-0.033259429,
0.013771018,
-0.0143627105,
-0.0266611613,
0.0188524984,
-0.0345567465,
0.0337475725,
0.0346604809,
0.0822985992,
0.0728290901,
-0.0284100622,
0.0525743552,
0.0539110079,
-0.0365474634,
-0.028639812,
0.0023201245,
0.0120738596,
-0.023922652,
-0.0419275388,
-0.0394492224,
-0.0220596865,
0.0871128738,
0.0386606082,
0.0009976308,
-0.0014600808,
0.025476085,
-0.0758681223,
-0.0139198294,
-0.0158613566,
-0.0518858172,
-0.039331995,
0.0469234437,
-0.0374704339,
-0.0258408654,
0.0696234107,
-0.0041294554,
-0.0192026515,
-0.0708667561,
0.0339414738,
-0.0120679224,
0.0886520818,
0.0190133918,
0.0185894221,
0.0147493929,
-0.0535972714,
-0.059546113,
-0.0145682003,
0.0297560412,
-0.0119662993,
0.0894498229,
-0.0635600686,
0.0216801204,
0.0240034387,
-0.0005655981,
-0.0207050499,
-0.0539638326,
-0.0449478775,
-0.0050644563,
-0.0133708557,
0.0562425256,
0.0322977789,
-0.0041529196,
0.0371318832,
0.0041976259,
-0.0192142371,
-0.0297241844,
-0.0333554111,
0.0250993706,
0.0442061424,
0.0386781208,
0.0316566303,
-0.0551436171,
0.0316306651,
0.040013127,
0.0168874413,
-0.0015678294,
0.0206794795,
0.0105366064,
0.0076999594,
-0.0593350306,
-0.0489748418,
0.034969382,
0.038350869,
-0.0710733309,
-0.0366005935,
0.0766771883,
0.0759682134,
0.0023385207,
0.0449798666,
-0.0757269487,
0.0277393684,
0.0220109243,
-0.002795069,
-0.0398141369,
-0.0190962553,
0.0336319692,
-0.027103059,
-0.0199321695,
0.0047415975,
-0.0181345493,
-0.0902330577,
0.0199133307,
-0.0010383418,
-0.0193275232,
0.0454352163,
-0.0014620259,
0.0253539197,
0.0523485877,
0.0313953534,
0.0403389148,
-0.0445841812,
0.0189270508,
0.0005836018,
0.0395716503,
0.0237093475,
-0.0178981405,
-0.0048209787,
0.0336981677,
0.0450301655,
-0.0028021063,
-0.0157872587,
-0.002890687,
0.0979590565,
-0.0426779985,
0.0385329723,
-0.0150207933,
-0.0065255747,
0.0565996878,
0.0155280195,
-0.0461864583,
-0.0173516925,
-0.0097117303,
-0.0433835573,
0.0227032993,
0.0683871508,
-0.0274509173,
-0.0149796121,
-0.0075564766,
0.0148631344,
-0.0557452329,
0.0603484809,
0.0488247499,
-0.0222778432,
-0.0449536219,
-0.0085194428,
0.0206612889,
0.0321302563,
-0.0346426852,
0.0566962101,
-0.0140737304,
-0.0001846511,
-0.040427506,
-0.0782608017,
0.0198145956,
0.0133967549,
0.0090469318,
0.0484615564,
0.0670260638,
-0.0348879024,
-0.0099632656,
0.0298277065,
-0.0012805924,
0.0312385671,
0.0441479385,
0.0139980977,
-0.0025991714,
-0.001516384,
-0.0325988717,
0.0051987558,
-0.0828333795,
-0.0296545494,
0.05299934,
-0.0189598221,
0.0285937022,
0.012248829,
0.0159449913,
-0.0147093534,
-0.049640134,
-0.0394807458,
-0.0125750806,
0.074860163,
-0.0297266301,
0.0622471161,
0.0059889327,
0.0338022821,
-0.0136634996,
-0.0139862197,
0.0143449055,
0.0462314747,
-0.0432626866,
-0.0194585864,
0.0588148721,
0.016037764,
0.0359547734,
0.0598299913,
0.0274228398,
-0.0043519796,
0.0060670637,
-0.0289359186,
-0.002831965,
0.0052829063,
-0.0418099351,
0.0307556689,
-0.0173508264,
-0.0360811502,
-0.02100618,
-0.019211499,
0.0092654265,
-0.0098618902,
-0.0386836343,
-0.0038275311,
-0.0371353477,
-0.0265259538,
0.0180767216,
-0.0027272305,
-0.0085122855,
-0.0348827802,
0.0277680401,
0.0335622318,
0.0592680648,
0.0059179571,
-0.0163514744,
-0.0676400065,
-0.0421912707,
-0.0261328369,
-0.0335411318,
0.0604092032,
0.0561082661,
0.0101918196,
-0.0149363736,
0.0862285793,
-0.0347203873,
0.0212744419,
-0.0174329989,
-0.0024912283,
0.0125838537,
-0.0409927256,
0.0594485924,
0.0081378892,
-0.0081156567,
-0.0331019647,
0.0585950613,
-0.0255396944,
0.0117344363,
-0.0016743117,
-0.0086249495,
-0.0150211193,
-0.0143482443,
-0.0589067824,
0.0058457735,
0.0044030794,
0.0331849679,
-0.0816169679,
-0.0417922251,
-0.0064575947,
-0.0164754018,
0.048834607,
-0.0670767203,
0.0246552173,
0.0462956391,
0.0223114938,
-0.1036678329,
-0.0122219417,
0.0522521958,
-0.0215792395,
-0.0100640422,
-0.0665985718,
-6.341200602e-33,
-0.0285446569,
-0.0382035263,
-0.0006099122,
-0.0128260013,
-0.098633185,
0.0142233633,
-0.0409935303,
0.0135113159,
0.0005975025,
0.0337871984,
0.0049640685,
0.0667036995,
0.012539275,
-0.0025943019,
0.0323729143,
-0.0393621959,
0.0348950177,
0.0043322095,
-0.0085548041,
-0.0240461901,
0.019391479,
-0.0016511115,
0.0090427306,
-0.0162559263,
0.0224949569,
-0.0113368239,
0.0012642068,
0.0299335197,
-0.0546583608,
0.0457664728,
0.0104993107,
0.0490836203,
0.0469564758,
-0.0243891664,
-0.029420713,
0.0281450581,
-0.0507928357,
-0.0289973039,
0.089696154,
0.0483684763,
-0.0631377101,
-0.0411053151,
0.0822539926,
-0.0135928839,
0.0025648614,
0.0173848532,
0.0267158635,
-0.0233714059,
-0.0153807942,
0.022701662,
-0.0521870255,
-0.0280542672,
-0.0315099023,
-0.0116960574,
-0.0034611635,
0.0063283397,
-0.0194034278,
0.0205400474,
-0.0442349687,
-0.0184473582,
-0.0010023329,
0.0200466458,
0.0238154083,
0.0129013574,
0.0474402383,
0.0252518449,
0.0191317443,
0.0130353905,
-0.0034303993,
-0.0209811069,
0.02044007,
-0.0063826526,
0.0080128396,
-0.0168039501,
0.0327481069,
-0.0007302622,
-0.0470757447,
-0.0519410297,
0.0397213958,
0.0670862868,
0.0144024035,
0.0090728588,
-0.0144621171,
-0.0216512978,
-0.0780818537,
0.0364732258,
-0.0500910096,
-0.0135060698,
0.0385188907,
-0.051289957,
0.0272000208,
-0.0543983132,
0.0013427443,
-0.0138214212,
-0.0254798364,
-0.0129989274,
-0.0280003008,
0.0542593859,
0.0062340451,
0.0613483079,
0.0028627093,
-0.0121102501,
-0.0227198545,
-0.0088670924,
0.0269607548,
0.0456561297,
-0.0088398904,
-0.028900411,
-0.0381556451,
-0.0047673057,
-0.0272022989,
-0.0470065437,
-0.0250292793,
-0.055428829,
0.0958047658,
0.0402455218,
-0.0198542923,
0.0228382647,
0.0496379621,
0.0030832968,
-0.0379916541,
0.023472745,
-0.0341197811,
0.035139177,
-0.0069278274,
0.0034679228,
0.0021113963,
0.0170594323,
-0.003008035,
-0.0674399808,
-0.0187038574,
0.0251886994,
2.854e-7,
0.0760412291,
0.0877155587,
-0.0021494308,
0.077011846,
0.0410127044,
-0.0049309954,
-0.0177915953,
0.0025176392,
0.0313637331,
-0.028307505,
-0.0256998204,
-0.0039684935,
0.0072908192,
-0.0078566866,
-0.0194205027,
0.0447643846,
-0.0364782922,
0.0620406009,
-0.0465949029,
0.0471968316,
0.0767528415,
0.0324006006,
-0.0203227494,
0.0373738222,
-0.0384085998,
-0.00442668,
0.0453426577,
-0.0047837431,
-0.0200497378,
-0.0055610789,
0.0255419649,
-0.0359664559,
0.0393410325,
0.0537827089,
0.0237574708,
-0.003524466,
0.0042393599,
0.0222629625,
-0.0077903355,
0.0056047756,
-0.0217153896,
-0.0327859297,
0.0205306802,
-0.067440547,
0.0494657122,
-0.0852287933,
-0.0384337977,
0.0372220241,
0.0099178571,
-0.016195124,
0.0057803593,
0.0549149029,
-0.0240791943,
0.015001894,
0.0338006467,
-0.0089558577,
-0.0151463756,
-0.0233640075,
-0.0427186452,
0.054623466,
-0.0284291301,
-0.0188245289,
-0.0365772322,
0.0195399653,
0.0369121507,
-0.040430665,
-0.0707922578,
2.269094857e-34,
0.0287702847,
-0.0234305989,
0.0710136369,
0.0330102593,
0.0102799339,
-0.0302426349,
0.0157744437,
0.0321947411,
0.0141364299,
-0.0883143246,
0.0000803661
] |
http://arxiv.org/abs/2405.15984v1
|
http://arxiv.org/pdf/2405.15984v1
|
Evaluating the Adversarial Robustness of Retrieval-Based In-Context Learning for Large Language Models
| 2024-05-24T23:56:36 | 2024-05-24T23:56:36 |
Simon Chi Lok Yu, Jie He, Pasquale Minervini, Jeff Z. Pan
|
cs.CL
|
cs.CL, cs.AI
|
With the emergence of large language models, such as LLaMA and OpenAI GPT-3,
In-Context Learning (ICL) gained significant attention due to its effectiveness
and efficiency. However, ICL is very sensitive to the choice, order, and
verbaliser used to encode the demonstrations in the prompt. Retrieval-Augmented
ICL methods try to address this problem by leveraging retrievers to extract
semantically related examples as demonstrations. While this approach yields
more accurate results, its robustness against various types of adversarial
attacks, including perturbations on test samples, demonstrations, and retrieved
data, remains under-explored. Our study reveals that retrieval-augmented models
can enhance robustness against test sample attacks, outperforming vanilla ICL
with a 4.87% reduction in Attack Success Rate (ASR); however, they exhibit
overconfidence in the demonstrations, leading to a 2% increase in ASR for
demonstration attacks. Adversarial training can help improve the robustness of
ICL methods to adversarial attacks; however, such a training scheme can be too
costly in the context of LLMs. As an alternative, we introduce an effective
training-free adversarial defence method, DARD, which enriches the example pool
with those attacked samples. We show that DARD yields improvements in
performance and robustness, achieving a 15% reduction in ASR over the
baselines. Code and data are released to encourage further research:
https://github.com/simonucl/adv-retreival-icl
|
Evaluating the Adversarial Robustness of Retrieval-Based In-Context Learning for Large Language Models: With the emergence of large language models, such as LLaMA and OpenAI GPT-3,
In-Context Learning (ICL) gained significant attention due to its effectiveness
and efficiency. However, ICL is very sensitive to the choice, order, and
verbaliser used to encode the demonstrations in the prompt. Retrieval-Augmented
ICL methods try to address this problem by leveraging retrievers to extract
semantically related examples as demonstrations. While this approach yields
more accurate results, its robustness against various types of adversarial
attacks, including perturbations on test samples, demonstrations, and retrieved
data, remains under-explored. Our study reveals that retrieval-augmented models
can enhance robustness against test sample attacks, outperforming vanilla ICL
with a 4.87% reduction in Attack Success Rate (ASR); however, they exhibit
overconfidence in the demonstrations, leading to a 2% increase in ASR for
demonstration attacks. Adversarial training can help improve the robustness of
ICL methods to adversarial attacks; however, such a training scheme can be too
costly in the context of LLMs. As an alternative, we introduce an effective
training-free adversarial defence method, DARD, which enriches the example pool
with those attacked samples. We show that DARD yields improvements in
performance and robustness, achieving a 15% reduction in ASR over the
baselines. Code and data are released to encourage further research:
https://github.com/simonucl/adv-retreival-icl
|
[
0.0159108602,
0.073361665,
-0.0133124841,
0.0790177435,
-0.019369619,
-0.0229311101,
-0.0039272443,
0.0299480632,
0.0248828977,
-0.0452110469,
-0.0432560071,
-0.0373750515,
-0.000386466,
-0.0233282521,
0.0680329725,
-0.0345744565,
0.0351522043,
-0.0273516811,
-0.0093509993,
-0.0007676428,
-0.0021723653,
-0.0282047316,
-0.0057991887,
0.0270405486,
0.011397888,
-0.0028785982,
-0.0150560336,
0.0096346112,
0.0247039385,
-0.0434355885,
0.0370788909,
0.0084832245,
0.0318673663,
0.0515738726,
0.0000020402,
-0.0318696611,
-0.0212871116,
0.013155912,
-0.0060233837,
0.0357789248,
0.0337132514,
-0.0235661622,
-0.0028317629,
0.0259544924,
-0.0425495319,
-0.0155955087,
-0.0092249373,
0.0570840724,
0.0558363907,
0.1226924658,
0.0085318126,
-0.0200523008,
0.0482582748,
-0.0476688147,
-0.0183849577,
-0.0517647751,
0.0180416871,
-0.0534571745,
0.0078787059,
0.0316874124,
-0.0033183845,
0.0320842303,
0.0326929428,
0.0304943416,
-0.0815383941,
0.0397973545,
-0.0431093946,
0.0249368344,
-0.0091212885,
0.0676062927,
-0.0213470068,
-0.0154553391,
-0.0150587615,
0.0002951354,
0.0268259812,
0.0020132479,
-0.051846277,
0.0104150875,
0.0139412312,
0.0270169899,
-0.0296655092,
0.0154403709,
0.0639502034,
0.0411252826,
-0.0081767812,
0.0318668038,
0.0112717673,
-0.0432960652,
-0.0035986095,
-0.0411939584,
-0.02878966,
-0.0130585739,
-0.0241802149,
0.0256843958,
-0.0214309432,
0.0229835287,
-0.0572494455,
-0.045187898,
0.0478751697,
-0.0176273994,
0.0607002899,
0.0177611541,
-0.0351824723,
0.0157932024,
-0.010769452,
0.0742935538,
-0.0240339283,
0.0757446885,
-0.0191489458,
0.001045338,
-0.0106381457,
0.0015332373,
-0.0092744417,
0.017240325,
-0.0303052124,
0.0091348952,
-0.0052488735,
0.0267732758,
-0.0351660587,
0.0164917149,
-0.0513530374,
0.0010958345,
-0.0447819307,
-0.0033742888,
0.0058418312,
-0.0360269025,
-0.0003943903,
0.014837659,
0.010620242,
-0.1039016396,
-0.0618508272,
0.0096769007,
-0.0351489894,
-0.0437623151,
-0.0096455133,
0.0952069461,
-0.0015764426,
-0.0942593738,
-0.0341397151,
-0.0233089998,
0.0046502976,
-0.0532502905,
0.0323679261,
-0.0534274466,
0.0060778544,
-0.0400127396,
-0.0111598969,
-0.0893337578,
0.0204705391,
0.0504679456,
-0.0225830711,
0.0617209077,
0.0246853977,
-0.0145956548,
0.0312405694,
0.0437776223,
-0.0259207971,
0.0473578013,
0.018859623,
0.0247049034,
0.0487115979,
0.0434776507,
0.0405318514,
0.0410012305,
0.0788540989,
0.048239734,
-0.0337827057,
0.0362149552,
0.0212880783,
0.0148472069,
-0.0023604138,
0.0574872233,
-0.0287794471,
-0.0559101738,
0.0788270682,
0.0277160592,
-0.0202501081,
-0.0195038859,
0.0048923702,
-0.0042613107,
0.0350726657,
-0.0107176797,
-0.0646180734,
-0.012471552,
-0.0493089706,
0.0027061484,
0.0316775478,
0.0086557521,
-0.0217497982,
-0.0219162256,
-0.0140866777,
-0.0189735014,
0.0031761769,
0.0258018896,
0.0341209956,
0.0457676724,
-0.0548445396,
0.0015838877,
0.0183881968,
-0.0225646868,
-0.0327012762,
-0.0784879029,
0.0042363163,
-0.0158320386,
0.0698281378,
-0.0572727062,
0.0172639228,
-0.0239322148,
0.0218547005,
0.0156077556,
0.0102594467,
-0.0034067011,
0.0082449783,
-0.0360425189,
0.0006475148,
-0.0591506623,
0.0138267037,
0.0272014644,
0.0675119534,
0.0172368865,
0.011888166,
-0.0239419658,
-0.0108456006,
0.050692942,
-0.0035806762,
-0.0826816708,
0.0020939137,
0.0125643434,
0.0527914949,
0.0022482672,
-0.0283619314,
0.0155363809,
0.0732129291,
-0.0161227547,
0.0213971138,
-0.0300049819,
-0.0213450789,
-0.0714750364,
0.0183769092,
0.0227699373,
-0.0196473878,
0.0547948442,
-0.0488050058,
0.003999583,
0.0066187866,
-0.0077688284,
0.0596007034,
-0.0385546573,
-0.0191632342,
0.0087182429,
0.0295552984,
-0.0540902093,
0.0717254281,
-0.0345262587,
-0.0299344677,
0.0327351801,
0.060459923,
0.006970156,
-0.0615044311,
-0.0345131755,
0.0498324372,
-0.0327835977,
0.0472189598,
-0.0122826183,
0.0012968738,
-0.0126913264,
-0.0059944754,
-0.0517738387,
0.0104368143,
-0.0548142791,
0.0118019991,
-0.0036278674,
-0.0049283835,
-0.0293878913,
-0.0222987309,
0.00855901,
-0.0273203049,
0.0022921432,
0.0171199366,
-0.0077198413,
0.0273788553,
-0.0077611948,
-0.0289270859,
0.0030532924,
0.012164915,
0.0797544941,
-0.0563328862,
0.0050270539,
0.0015512449,
0.030806955,
-0.005112642,
0.0600177199,
-0.0483081937,
-0.0137164472,
0.0369662195,
-0.011241057,
-0.006779131,
0.0781298503,
0.0144586535,
0.0669433549,
0.0398820527,
0.017912861,
-0.067949824,
0.0162454639,
0.0213804059,
-0.039797511,
-0.0251916554,
0.0276930872,
0.0283362288,
0.0225250833,
0.0327933207,
-0.0089034094,
-0.0100071523,
0.0133678969,
0.0078025986,
-0.0323481113,
-0.0383460596,
-0.0156733114,
-0.0603908822,
-0.0223034006,
-0.0326190367,
-0.0611418933,
0.0677644536,
-0.0383063518,
0.0247595571,
-0.0263405927,
-0.0496829003,
0.0088014239,
0.0513391085,
-0.0182820484,
0.0064782477,
-0.0161750372,
-0.0144069567,
-0.0587587692,
-0.0388084985,
-0.00722912,
-0.0393139273,
0.0517422296,
-0.0170926079,
-0.0004102304,
0.0025213722,
-0.0442639925,
-0.0856783986,
-0.0335928798,
0.0078090616,
0.0317846648,
-0.0460535586,
0.0071751163,
-0.0304788332,
-0.0497962423,
0.036659874,
0.0127729177,
-0.0233569816,
0.0032051906,
-0.0196556617,
0.0453318991,
0.0311849546,
-0.013790261,
-0.0180042125,
-0.0388136692,
0.0105232708,
0.0719495341,
0.0462917127,
-0.0122650452,
-0.0316640623,
-0.0207704362,
0.0179928485,
-0.0139098475,
0.0183772389,
0.0592274368,
0.0249544457,
-0.0521729402,
-0.0359016508,
0.0358735807,
0.0078283465,
-0.0212282371,
-0.0030008943,
0.0165948514,
0.022263838,
0.0102926036,
0.0186407939,
-0.0413892269,
-0.0117901508,
0.0252281148,
0.0591250882,
0.01067694,
0.0157736931,
-0.0094220424,
-0.1428333372,
0.0161480159,
0.0113122296,
-0.072129786,
-0.058584094,
-0.0350345522,
0.0027713894,
0.0322484933,
0.0803434402,
0.0430291407,
0.0131750023,
0.0053175655,
0.0142471967,
0.0413899794,
-0.0891391039,
0.0061210813,
-0.0353208147,
0.0206334069,
0.0826470256,
0.022078855,
-0.0147061422,
-0.0388702974,
0.0471945591,
-0.0326752514,
-0.0072003831,
0.0196228996,
0.0424157791,
-0.0151147889,
-0.0009195281,
0.0232459642,
0.0522561148,
0.001746446,
-0.0363323763,
0.0192504544,
0.027659066,
-0.0193993449,
0.0219943505,
-0.0374810211,
-0.0438516773,
0.0323777981,
0.0586121567,
0.0182760339,
-0.0239700843,
-0.0106023867,
-0.0552021451,
0.0045248941,
-0.016049616,
-0.0264744312,
0.018189827,
0.0293045472,
0.0223639701,
-0.0595275089,
-0.044783961,
-0.0355877616,
0.042917069,
-0.0144509664,
-0.0062652444,
-0.0691070259,
0.0043269824,
-0.0449955873,
0.0196617506,
-0.0329488739,
0.032434117,
-0.0423698537,
0.0213986859,
0.0331400484,
-0.0433232524,
-0.0211172421,
0.0078455964,
-0.0630090535,
-0.0803305358,
-0.0141639505,
0.0763124377,
0.0121752359,
0.0572281368,
-0.0119601265,
-0.0500217043,
0.0974089801,
0.0671080947,
0.0082975402,
0.0143254437,
0.0568780825,
-0.0124311084,
0.0166181959,
0.0348766074,
-0.0623788647,
-0.0297616255,
0.0435589813,
0.038849365,
-0.0199140944,
-0.001734908,
0.0248816069,
0.0013750893,
0.0787880346,
-0.0064439797,
-0.0183311366,
-0.008737932,
0.0370971002,
-0.0062383777,
-0.0141974445,
-0.0115789557,
0.0278019477,
0.0019901963,
0.0040515582,
0.0042679668,
-0.0196421221,
0.0198561717,
0.0408000126,
-0.0174892042,
0.0041265995,
0.0237666667,
0.0019437196,
0.0223509073,
0.0328458212,
-0.017095197,
0.030343229,
-0.0144396666,
0.0467472486,
-0.0100318622,
0.0515591167,
0.0240311697,
-0.0243268274,
-0.005013939,
0.0218109377,
-0.0551296622,
-0.0080183456,
-0.0240481943,
0.042198319,
0.0408240557,
-0.0025953224,
-0.0090809623,
-0.0309051536,
-0.0085628098,
-0.0197633337,
-0.0251165722,
-0.029349437,
-0.0529510416,
0.0392630026,
0.0197729133,
0.0129371947,
0.0421022214,
0.0785442293,
0.0026925118,
0.0546828061,
-0.0048567755,
0.0030079682,
-0.0353550762,
-0.0327843465,
0.0856620073,
0.0326660573,
-0.0212477706,
-0.0243159346,
-0.0215083603,
-0.0753159076,
0.0054281331,
0.0291290767,
-0.0187961757,
-0.0083119944,
0.021874724,
0.0511166379,
-0.0008513095,
0.0243880972,
-0.0398256853,
-0.0117878225,
0.0174171459,
-0.0766432583,
-0.011764111,
-5.945180963e-33,
-0.0095533105,
-0.0247156192,
0.008752848,
-0.0345194489,
0.0049812165,
0.0019945507,
-0.0390798077,
-0.0202002004,
-0.0320696607,
-0.0142152067,
-0.0212813783,
-0.0076370854,
-0.0132799,
0.0330364779,
-0.0451792069,
-0.0490961336,
-0.0006346925,
-0.0111875189,
-0.0306154545,
0.003043181,
0.0327632725,
0.0351508409,
0.0225663129,
-0.0803272426,
0.0469963066,
-0.0239155982,
-0.0289921518,
0.0006640073,
-0.0706579462,
0.0595237985,
-0.0043589137,
0.0458348095,
-0.0022814404,
0.0002441572,
0.0079498887,
0.0035432063,
-0.017216593,
-0.011960879,
-0.0248596333,
-0.002026187,
-0.0277344603,
-0.0301073771,
0.0429962315,
-0.0214760322,
-0.0247758254,
0.0103007089,
-0.0089203846,
-0.0012533345,
-0.0390922278,
-0.0190080293,
-0.001967571,
0.004921183,
0.0230535492,
-0.0328614525,
0.0038152908,
-0.0084009236,
0.0356197208,
0.0334254391,
-0.0761636347,
-0.00370753,
0.0041112285,
0.0177144613,
0.0407591313,
0.0033279026,
0.0064183697,
0.0676153824,
-0.0069743707,
0.0017322177,
0.0256141834,
-0.0587184764,
0.0404605418,
0.0262665153,
0.0446600989,
0.0854547173,
0.054796353,
-0.0133292964,
-0.0843057036,
0.0237770882,
0.013606145,
0.0292923767,
0.0191462263,
0.0386281647,
-0.068527557,
0.0139536392,
-0.0071664504,
-0.0305143818,
-0.0254642647,
-0.0311397742,
0.0601372235,
0.0300756805,
-0.024425257,
-0.025180053,
0.0181505587,
-0.0468876474,
-0.0149228005,
0.0446231849,
-0.043368835,
-0.0080216462,
-0.0153706251,
0.0138121592,
-0.0194042195,
-0.0284899492,
0.0066267783,
-0.0503293201,
0.0049051004,
0.0087342104,
0.0495802872,
0.0411130972,
-0.0618016385,
0.0028730978,
-0.0203974247,
0.0394914895,
0.0313885845,
-0.0098831272,
0.0376350433,
0.0084980195,
0.0029070675,
0.0199022423,
0.0309872031,
-0.0619314872,
-0.0426542163,
0.005136095,
-0.0314862989,
-0.0255633742,
-0.0263130814,
0.0096168658,
0.0030667882,
0.0636142343,
0.0171079654,
-0.0743994713,
0.0043787993,
-0.0064445944,
2.761e-7,
-0.0057692435,
0.0319816694,
0.0027599926,
0.0325401314,
0.0116237961,
0.0398752019,
-0.0420840606,
0.0767099336,
0.0141341556,
-0.0258249398,
-0.0032190375,
0.0239591505,
0.0516341999,
0.0286135431,
-0.103455022,
0.0066700368,
-0.0889971554,
-0.0123172915,
-0.0442610905,
-0.0367598236,
0.0297480989,
0.1211662963,
0.0197849311,
0.0064067617,
-0.0099494616,
-0.070866555,
0.0662878975,
-0.014006923,
0.0090264045,
-0.0112570832,
0.0217845775,
0.0410298817,
0.0161820967,
-0.015463301,
-0.0312336516,
0.0436040945,
0.0063047013,
0.0268528834,
0.0120500745,
-0.0632167011,
-0.0059934836,
-0.0116297072,
-0.0058932062,
-0.0400947221,
0.0692983419,
-0.0243743714,
0.0047372649,
-0.0026655679,
-0.0018488248,
-0.0380047671,
0.0181293581,
-0.0309001431,
-0.0038271667,
0.0033753272,
-0.01002088,
0.0007267653,
0.0088732475,
-0.0793137923,
-0.0420679413,
-0.0142499469,
-0.02968858,
-0.007032882,
-0.0264257714,
0.0424310118,
-0.0074597583,
-0.0262587834,
-0.004128179,
2.493284589e-34,
0.0171627086,
-0.030208908,
0.0068239239,
0.0264379289,
0.0301430505,
-0.008814103,
0.0300224107,
0.0146933068,
0.0140450122,
-0.0304409061,
-0.0321731977
] |
http://arxiv.org/abs/2405.15973v3
|
http://arxiv.org/pdf/2405.15973v3
|
Enhancing Visual-Language Modality Alignment in Large Vision Language Models via Self-Improvement
| 2024-05-24T23:09:27 | 2024-06-07T20:15:24 |
Xiyao Wang, Jiuhai Chen, Zhaoyang Wang, Yuhang Zhou, Yiyang Zhou, Huaxiu Yao, Tianyi Zhou, Tom Goldstein, Parminder Bhatia, Furong Huang, Cao Xiao
|
cs.CV
|
cs.CV, cs.AI, cs.CL, cs.LG
|
Large vision-language models (LVLMs) have achieved impressive results in
various visual question-answering and reasoning tasks through vision
instruction tuning on specific datasets. However, there is still significant
room for improvement in the alignment between visual and language modalities.
Previous methods to enhance this alignment typically require external models or
data, heavily depending on their capabilities and quality, which inevitably
sets an upper bound on performance. In this paper, we propose SIMA, a framework
that enhances visual and language modality alignment through self-improvement,
eliminating the needs for external models or data. SIMA leverages prompts from
existing vision instruction tuning datasets to self-generate responses and
employs an in-context self-critic mechanism to select response pairs for
preference tuning. The key innovation is the introduction of three vision
metrics during the in-context self-critic process, which can guide the LVLM in
selecting responses that enhance image comprehension. Through experiments
across 14 hallucination and comprehensive benchmarks, we demonstrate that SIMA
not only improves model performance across all benchmarks but also achieves
superior modality alignment, outperforming previous approaches.
|
Enhancing Visual-Language Modality Alignment in Large Vision Language Models via Self-Improvement: Large vision-language models (LVLMs) have achieved impressive results in
various visual question-answering and reasoning tasks through vision
instruction tuning on specific datasets. However, there is still significant
room for improvement in the alignment between visual and language modalities.
Previous methods to enhance this alignment typically require external models or
data, heavily depending on their capabilities and quality, which inevitably
sets an upper bound on performance. In this paper, we propose SIMA, a framework
that enhances visual and language modality alignment through self-improvement,
eliminating the needs for external models or data. SIMA leverages prompts from
existing vision instruction tuning datasets to self-generate responses and
employs an in-context self-critic mechanism to select response pairs for
preference tuning. The key innovation is the introduction of three vision
metrics during the in-context self-critic process, which can guide the LVLM in
selecting responses that enhance image comprehension. Through experiments
across 14 hallucination and comprehensive benchmarks, we demonstrate that SIMA
not only improves model performance across all benchmarks but also achieves
superior modality alignment, outperforming previous approaches.
|
[
0.0188803636,
0.0685454011,
-0.0088110901,
0.0544940494,
-0.0321439505,
0.0097578894,
0.0164442062,
-0.0346764922,
0.0135168675,
-0.0550690442,
-0.0706205368,
0.0108794477,
-0.0281513799,
0.0181633215,
0.0101991007,
-0.036637146,
0.0369075947,
-0.0655236095,
-0.0260631125,
-0.0092106601,
-0.0026061237,
-0.0047665699,
-0.0131124491,
0.0235423576,
0.0092184609,
-0.049804911,
0.0000271722,
0.0305863637,
0.0000284327,
-0.0169625785,
-0.0109904129,
0.0651368871,
0.0313903913,
0.0077359583,
0.0000017667,
-0.0383342654,
0.0318128504,
-0.0169723574,
0.0011734092,
0.0137675721,
0.0201552156,
-0.0198141728,
0.0038015705,
-0.0083148172,
-0.0397437997,
-0.019359136,
0.0653556064,
0.025283128,
0.0778618306,
0.0255896803,
0.005594519,
-0.0010376042,
-0.014725632,
-0.0098662097,
0.0131369112,
-0.0122163631,
0.0045622312,
0.0556043461,
-0.0079162754,
0.0069538462,
-0.0803086758,
0.0537250526,
0.0374793671,
-0.0230950024,
0.023535287,
0.0091535542,
-0.0308745112,
-0.0293082166,
-0.0409438014,
0.0670629293,
0.0203948934,
-0.0377014428,
-0.0395531915,
-0.0016767361,
0.0340138115,
-0.000283324,
-0.0522773825,
0.0063549709,
0.0022337488,
0.0055326484,
-0.0278404672,
-0.0125901522,
0.0085360846,
0.0204268917,
0.0127007728,
0.0133330254,
0.0187426265,
0.0176611692,
-0.0233289953,
0.0027652706,
-0.033762373,
-0.0421251878,
-0.0270879157,
0.0334275924,
-0.0465043336,
-0.0110577196,
-0.0415137708,
0.0140598537,
-0.0132868383,
-0.0059629921,
0.0530423708,
0.0389704034,
-0.0608979389,
0.069651261,
-0.0152709866,
0.0788456872,
-0.0248634517,
0.0767974108,
-0.0508609414,
-0.0104473168,
-0.0177937895,
-0.0151196849,
-0.041725941,
-0.0164829958,
-0.0498943888,
-0.0402942449,
-0.0249066651,
0.0449003056,
-0.033806704,
0.0000602967,
-0.0059140571,
0.0364280455,
-0.0461233668,
-0.0210129954,
-0.0709610954,
0.0530457497,
-0.0265204292,
-0.0093946811,
0.0014456528,
-0.0336862318,
-0.0116356285,
0.0320093147,
-0.0043178978,
-0.018640114,
-0.0071458924,
0.0814895555,
-0.0454542972,
-0.0036489968,
-0.0033868283,
0.0016189098,
-0.0178560112,
-0.0455836207,
0.0185068902,
-0.0516293049,
0.0091693439,
-0.0068091955,
-0.0273404624,
-0.0342412405,
-0.0088859815,
-0.0117094796,
0.0081109004,
0.1003504321,
-0.0063051749,
0.0072329314,
0.0150808627,
0.0115811164,
0.0096458578,
0.0843362212,
0.0010702561,
0.0226757973,
0.0425616652,
-0.0333169959,
0.0222832263,
0.0239511821,
0.0007871515,
-0.0059600282,
-0.0827724561,
0.0498623438,
-0.0009053695,
0.0336640328,
0.0493516475,
0.0290474407,
-0.0295472853,
-0.0711088479,
0.0584209859,
0.0183232184,
0.0640004426,
-0.0037464013,
0.0248755943,
-0.0174558349,
0.011321024,
0.0090676844,
-0.0799979493,
0.0957230851,
0.0111676231,
0.0049507669,
0.0078562638,
0.0289524719,
-0.0459313467,
-0.0612751432,
0.0060345535,
0.0250219256,
-0.0140491519,
0.0613607764,
0.023897564,
0.0375807807,
-0.0120206214,
-0.0194523241,
0.0112400921,
-0.0190797746,
0.0039704763,
-0.0577990673,
0.068265751,
0.0744378939,
0.0432338677,
0.0010933419,
-0.0277837496,
0.0151851857,
0.0417812467,
0.0544226579,
0.021065183,
-0.0044145491,
0.0232092198,
-0.0377770551,
-0.0590461828,
-0.0387851372,
-0.0151985046,
-0.0082599884,
0.0169738047,
-0.0321280882,
-0.0042459425,
-0.037796732,
0.0090985168,
0.0305123031,
0.0362965912,
-0.0152667686,
-0.0033007937,
0.0094417837,
0.0001613055,
-0.0252096821,
0.0070360932,
0.0033803736,
0.006943515,
-0.0179707929,
0.02194627,
0.0249206293,
-0.069358103,
-0.0154790208,
0.0546972826,
-0.0398620404,
-0.0216402486,
0.0358344167,
-0.0212121494,
-0.0201189145,
0.0385193825,
-0.0086750705,
-0.0227974784,
-0.0552769713,
0.0041852435,
0.0125736855,
0.0172901414,
-0.030519234,
0.0356526859,
0.0003549908,
-0.0195519626,
0.0183390789,
0.0537691675,
0.007868873,
-0.0210321061,
0.022480756,
-0.0080454685,
-0.0104102474,
0.0153261991,
0.0093588466,
0.0026189634,
-0.0727431029,
-0.0176371951,
-0.0844872743,
-0.012142594,
-0.0251406189,
-0.0398874842,
0.0369531177,
0.0056814668,
-0.0032480776,
-0.0656341687,
-0.0107495673,
0.0477363579,
0.0119189844,
0.012840718,
0.0314982384,
-0.0053222058,
0.0020441301,
-0.0284323134,
-0.0122457212,
0.0234118626,
0.0227650926,
0.0277855396,
0.0378262252,
-0.0528456867,
0.0001438548,
-0.0445302911,
0.0649734661,
-0.0080768876,
-0.0413943231,
0.017139215,
-0.0157786701,
0.0471977107,
0.102577284,
0.0112590995,
0.0124896253,
0.0499291718,
-0.012132979,
-0.0540620163,
0.0136007527,
0.0132003175,
-0.0379647613,
-0.0131800836,
-0.0068688518,
-0.0452525355,
0.0387085974,
-0.0264815502,
0.0032917499,
-0.0693982989,
0.0485628732,
0.016966017,
-0.0537026897,
-0.0447672904,
0.0113080218,
0.0258356296,
0.0428518131,
-0.054918658,
-0.0971186757,
0.0771258399,
-0.016637411,
-0.0239229444,
0.0043894309,
0.0335671715,
-0.0198025908,
0.0626804233,
0.0293376558,
0.0184218232,
-0.0481898002,
0.0042462652,
-0.0858957544,
-0.0349778198,
0.0446564108,
-0.0034756535,
0.0761611238,
-0.0637935698,
0.0166551117,
-0.0305445567,
-0.0451386832,
-0.0433085747,
-0.0193761718,
-0.0188085847,
-0.0262374599,
-0.0704766884,
-0.0098707788,
-0.009790116,
0.019987395,
0.015467112,
0.0132497456,
-0.0326622538,
-0.0194001868,
-0.0072197705,
0.0027241628,
0.0856737718,
0.0740572736,
0.0041772919,
-0.0087098517,
0.0410633385,
0.0590510666,
0.0486269109,
0.0005449216,
-0.0571104959,
0.049844414,
0.0251899138,
0.0179531854,
0.0380399935,
0.0294206999,
0.0745409578,
-0.047017619,
-0.019990975,
0.0725487992,
0.0098336507,
-0.0256802607,
-0.0260440465,
-0.0420443043,
0.0476181842,
0.0638073832,
0.0409152284,
-0.0679876134,
-0.0410099141,
-0.0323529802,
0.0029211813,
0.025722092,
0.0137432069,
-0.0136430701,
-0.0853074566,
0.080377914,
0.0232553892,
-0.0236435011,
-0.0325878374,
0.0212225523,
0.0027780035,
-0.0233221222,
0.012455429,
0.0099940235,
-0.0196895115,
0.0072898031,
-0.0051924079,
0.0374502391,
-0.0002159943,
0.0293590669,
-0.0368398614,
0.0276536979,
0.0532074496,
-0.0324351378,
0.0235582888,
-0.0509449132,
0.0402394533,
-0.0113909543,
0.0066330894,
-0.0131429071,
0.0456125475,
-0.0056056944,
-0.0550870113,
-0.0089083463,
0.0391355678,
-0.0006007262,
-0.0371128917,
-0.0119340019,
0.0335485972,
-0.0060787406,
0.008045732,
-0.0488690361,
0.0167262498,
-0.0185002107,
0.0300092716,
0.0237847324,
-0.0307000931,
-0.0401397198,
-0.009939787,
-0.0121177658,
-0.0716841593,
-0.0443585254,
0.0229242034,
0.069269985,
-0.0047353175,
-0.0507690571,
-0.0296099782,
0.0290913731,
0.0501346402,
-0.0233815406,
-0.0077384878,
-0.0470596887,
0.0057365163,
-0.0624779724,
0.0832494274,
-0.0426674113,
-0.0039554336,
-0.0526486747,
-0.0113875596,
0.0190905854,
-0.0425642431,
-0.0012202667,
0.0129856244,
-0.0689335987,
-0.0309877917,
-0.0216179863,
0.0938724875,
0.0125797288,
0.0678409711,
0.0129035385,
-0.048206985,
0.0049802647,
0.026809087,
-0.0093117533,
0.0020830203,
-0.0128384531,
-0.0203199722,
-0.0066113998,
0.0306534786,
-0.0189548396,
0.0111164963,
0.0532700606,
0.0421252102,
0.0240928382,
-0.0071845399,
0.0487917662,
-0.0002056074,
0.0167079736,
-0.0026336331,
0.0313729607,
-0.0177128017,
0.0147520788,
-0.0050346614,
-0.0163549203,
-0.052195169,
0.0104291979,
0.0288116056,
-0.0347519629,
0.0257662237,
0.0115509331,
-0.0146789011,
-0.0014164544,
-0.0500667021,
0.0134735582,
0.0293719899,
0.0062826825,
-0.0112961968,
0.02103227,
0.000088382,
0.0063567976,
0.0212878119,
0.0549951307,
0.0120247416,
0.0846283659,
0.0260998718,
-0.0075090285,
-0.043333441,
0.0414847732,
-0.042028822,
-0.0566236787,
-0.0208528396,
0.0215144493,
0.0232462715,
0.0021608998,
-0.015281708,
0.0158365238,
-0.0315255262,
-0.0107332021,
0.0291772317,
0.0124661382,
-0.0154618006,
0.0769248307,
0.0387216285,
0.0087432237,
-0.0272258855,
0.0388358086,
-0.002067972,
0.0267025661,
0.0144430697,
0.0250771008,
-0.0497119203,
-0.0499534644,
-0.0080873929,
-0.0334055573,
-0.0197986308,
0.0435812064,
0.0009363948,
-0.0810991898,
-0.0377063826,
0.0607557744,
-0.0270282831,
0.0152744949,
0.0302264616,
0.062588796,
-0.0011604524,
-0.0555643328,
-0.0071534663,
-0.003743971,
-0.0105096838,
0.0087072467,
-0.0009275713,
-5.746842373e-33,
-0.0339087397,
-0.0303434711,
0.070617646,
-0.0405594781,
0.0123151792,
-0.0289884247,
-0.0173556656,
0.014770316,
-0.0259910282,
0.0001128356,
-0.0001514073,
0.0044336868,
0.0067025218,
0.0370894931,
0.0081200087,
-0.0190762691,
0.0332578085,
-0.0079734903,
-0.0596514679,
-0.0126226675,
-0.0217580479,
-0.0194032434,
0.0369731635,
-0.0632325038,
-0.0400854312,
0.0401075482,
-0.0247877687,
0.0458824262,
-0.0080532711,
0.107451342,
-0.0078588752,
0.0204114206,
0.0124633303,
-0.0047313948,
-0.0313282348,
-0.0093524633,
-0.0813708007,
-0.0252446756,
0.0356509797,
0.0283593535,
-0.1133574545,
0.0034449806,
0.0356211849,
-0.0090658423,
-0.0296237227,
-0.0403538942,
0.0239927471,
-0.0276629757,
-0.0084307324,
0.0756596625,
-0.0454545282,
-0.0177486148,
-0.0286113471,
-0.0513808355,
0.043022614,
0.0423632823,
0.0206036456,
0.0268390458,
-0.0270215217,
0.0265854094,
-0.0016594438,
0.0265487507,
0.1013295352,
0.0739912838,
0.0181772411,
0.0163937956,
0.0512837768,
-0.0506091453,
-0.0062927171,
-0.0644606873,
0.0099639231,
0.0551097728,
0.0405847356,
0.0729200989,
0.0045964341,
0.0243558381,
-0.0332846083,
0.014253174,
-0.0559818596,
0.0516554415,
-0.0040235235,
0.0174020194,
0.0019428562,
0.0056266305,
-0.0461790524,
-0.0049640043,
-0.0669077411,
-0.011141533,
0.0737687945,
0.0291899256,
-0.0258075874,
0.0215649735,
0.0467892922,
-0.0479355007,
-0.0104045477,
0.0191556606,
-0.034499485,
0.0440481156,
0.014989539,
0.0386199988,
-0.027430268,
0.0192092881,
0.0481723249,
0.0096492693,
-0.0231807772,
0.012554246,
0.0019141953,
0.017464133,
-0.0636243373,
0.0054157111,
-0.057019718,
0.0073421882,
-0.0204342082,
-0.01271486,
0.0117400521,
0.0156735033,
-0.0641531125,
0.0206546281,
-0.0103703011,
-0.0287721064,
-0.0329185575,
-0.001328314,
-0.0086750854,
0.0212637112,
0.0045695943,
0.001719252,
-0.0363727398,
-0.0234071948,
-0.0090581756,
-0.046352502,
-0.009233607,
0.0420870855,
2.318e-7,
0.0225521307,
0.0124392183,
0.0058721327,
0.0154808946,
0.0169060398,
0.0419033393,
-0.0231764596,
-0.0089417668,
0.0460574925,
-0.0521526597,
-0.0030215271,
0.0776036978,
0.0127458554,
-0.0149166482,
0.0187911987,
0.0624722578,
-0.0096416632,
0.0488354862,
-0.0320838466,
-0.0276277438,
0.0207262095,
0.0753064752,
-0.0375664979,
-0.0097292336,
0.0452455282,
-0.0515993051,
0.0820216462,
-0.0463342667,
0.0002540148,
0.0167399514,
-0.0178820565,
0.0259398632,
0.0024356106,
0.029390391,
0.0114356996,
-0.0130968252,
0.0031904227,
0.0177100562,
-0.0064593423,
-0.0593161359,
-0.0572540499,
0.0488913171,
0.0096916435,
-0.0586845949,
0.0783832818,
-0.0575704388,
-0.0288409069,
-0.0319407284,
-0.0075102216,
0.0085377609,
0.0349243842,
-0.0197073128,
-0.0504842773,
0.0105774729,
0.0532624088,
0.0018101722,
-0.0277182832,
-0.0619829334,
-0.013657093,
-0.0352644883,
-0.0269826204,
0.0131799821,
-0.0734660104,
0.0179996975,
0.0240124725,
-0.0251801684,
-0.041626744,
1.58586976e-34,
0.0119010713,
-0.0097832382,
0.0566707626,
-0.0102930414,
0.0103711635,
-0.0155581972,
-0.0095047699,
-0.0007757518,
0.0188515782,
-0.0557077564,
-0.0163691603
] |
http://arxiv.org/abs/2405.15964v1
|
http://arxiv.org/pdf/2405.15964v1
|
A hierarchical Bayesian model for syntactic priming
| 2024-05-24T22:26:53 | 2024-05-24T22:26:53 |
Weijie Xu, Richard Futrell
|
cs.CL
|
cs.CL
|
The effect of syntactic priming exhibits three well-documented empirical
properties: the lexical boost, the inverse frequency effect, and the
asymmetrical decay. We aim to show how these three empirical phenomena can be
reconciled in a general learning framework, the hierarchical Bayesian model
(HBM). The model represents syntactic knowledge in a hierarchical structure of
syntactic statistics, where a lower level represents the verb-specific biases
of syntactic decisions, and a higher level represents the abstract bias as an
aggregation of verb-specific biases. This knowledge is updated in response to
experience by Bayesian inference. In simulations, we show that the HBM captures
the above-mentioned properties of syntactic priming. The results indicate that
some properties of priming which are usually explained by a residual activation
account can also be explained by an implicit learning account. We also discuss
the model's implications for the lexical basis of syntactic priming.
|
A hierarchical Bayesian model for syntactic priming: The effect of syntactic priming exhibits three well-documented empirical
properties: the lexical boost, the inverse frequency effect, and the
asymmetrical decay. We aim to show how these three empirical phenomena can be
reconciled in a general learning framework, the hierarchical Bayesian model
(HBM). The model represents syntactic knowledge in a hierarchical structure of
syntactic statistics, where a lower level represents the verb-specific biases
of syntactic decisions, and a higher level represents the abstract bias as an
aggregation of verb-specific biases. This knowledge is updated in response to
experience by Bayesian inference. In simulations, we show that the HBM captures
the above-mentioned properties of syntactic priming. The results indicate that
some properties of priming which are usually explained by a residual activation
account can also be explained by an implicit learning account. We also discuss
the model's implications for the lexical basis of syntactic priming.
|
[
0.0192120578,
0.0101399021,
-0.0318164937,
-0.0058230613,
-0.0637211502,
0.0121668531,
-0.0433279984,
0.0357482545,
-0.0023730311,
-0.0666432008,
-0.0484796017,
-0.0586133227,
0.0317968838,
0.0090443995,
0.0360864215,
-0.0660077482,
0.0613720715,
0.0015240423,
0.012114984,
-0.0231506974,
-0.0069512525,
-0.0369840823,
-0.0161636882,
0.012973085,
-0.0180305596,
-0.0101719135,
0.0237527583,
-0.0146596376,
-0.0248392783,
-0.0238417219,
0.0434741303,
0.0511240065,
0.005293462,
-0.0251494478,
0.0000017227,
-0.0767264068,
0.0135444123,
-0.0441839807,
-0.006993142,
-0.016646456,
0.0527554229,
-0.0250250883,
0.0074283686,
-0.0127308089,
-0.0524369664,
-0.0256841592,
0.0295301154,
0.0333580598,
-0.0297092255,
0.0382034518,
-0.003734875,
-0.0189108606,
-0.0005735055,
0.0076771034,
-0.0037500181,
-0.0232564621,
0.0385457762,
-0.0326233096,
-0.0290123001,
0.0025101395,
-0.0507764928,
-0.0040890486,
0.0019310086,
0.0564974062,
0.0494886898,
0.0405363478,
0.0578622892,
0.0061102798,
-0.0614393689,
-0.0464437827,
-0.0260212924,
-0.0649055988,
0.0310404673,
-0.008079173,
0.0019167642,
0.0359881595,
-0.009626979,
0.0363159366,
0.0065788655,
0.0397145785,
0.06212271,
-0.0484299436,
-0.0079353182,
0.0150014209,
0.0552361645,
0.1051073074,
0.0154220108,
-0.062441919,
0.0375294238,
0.0345201455,
0.0007788169,
-0.0746860132,
-0.0306693986,
-0.0240964368,
-0.0037380988,
0.0496275015,
-0.0318956561,
-0.0428530201,
0.0038868792,
-0.0436052233,
-0.0215659514,
-0.0028612195,
0.0000188534,
-0.0265292302,
-0.0143234851,
-0.0030045055,
-0.0001865347,
0.0071656173,
-0.0330089182,
0.0713647231,
-0.0162327662,
-0.0298673157,
-0.0359035507,
0.0276327096,
0.0314515904,
-0.0096660638,
-0.0630876273,
-0.0291741323,
0.0183882006,
-0.0533988699,
-0.0534006804,
0.0213277247,
-0.072368212,
0.0008963223,
-0.0643263608,
-0.0257381778,
-0.0507127009,
0.0764946491,
-0.0002887759,
-0.0339779928,
-0.0243591201,
0.0078142369,
0.0338623039,
0.0180773456,
-0.002841858,
0.0122110629,
-0.0448403917,
-0.0342940502,
-0.0385585055,
-0.0316712819,
-0.0274622906,
-0.0013684916,
0.0218316037,
-0.0120277293,
-0.0581978858,
0.0066476176,
0.0075383643,
-0.023768669,
-0.0197431706,
0.0321548097,
-0.0275491197,
0.0552231371,
0.0579192787,
-0.0243728049,
0.017898256,
-0.0038421394,
-0.0219551083,
0.0650942475,
-0.0096398359,
0.0044247168,
-0.0006195459,
-0.0141984457,
0.0074939816,
0.0238179844,
0.012501006,
0.0126162292,
-0.0579897091,
-0.0353374407,
0.0024110707,
-0.0162522588,
0.0349884033,
0.0176864639,
-0.0227174945,
0.0034120791,
0.0445725322,
0.0530735962,
-0.0441153944,
-0.0102485688,
0.0359504335,
0.0101318387,
0.0046114228,
0.0775939003,
-0.0331087708,
0.0031136763,
-0.0249444284,
-0.0236492716,
-0.0280972887,
-0.0232543312,
0.0342810452,
-0.0179884061,
0.0056328895,
-0.0054200618,
-0.0183052588,
-0.0294248145,
0.0182707384,
0.0784083903,
-0.0134130185,
0.0233179051,
0.0550786592,
-0.0082965083,
0.0068912902,
-0.0038715361,
-0.0118518258,
-0.0194528475,
0.0186574776,
-0.003400201,
0.0415081978,
-0.0156045575,
-0.075658381,
0.0388444625,
0.0058805826,
0.0135801286,
-0.01031191,
-0.0284673348,
0.0054547484,
0.0168478191,
0.0360160656,
-0.0005682216,
-0.0357724316,
-0.0170703772,
-0.0178149939,
0.0041746232,
0.0215586815,
0.062544547,
0.0122070257,
-0.0368699022,
-0.0183381531,
-0.0090755532,
0.0486798957,
-0.0480158329,
0.0009965496,
0.0383045599,
0.0196760204,
0.0111317309,
0.0089104408,
0.0069340696,
-0.0607874952,
-0.0435457565,
-0.0085414806,
-0.0146907326,
0.0068776696,
0.0295265988,
0.0257391445,
-0.0480902307,
-0.0087347953,
-0.0373816192,
0.0977633446,
-0.0120981522,
0.0502069406,
-0.017403435,
0.0281351116,
-0.0017462795,
0.0537746772,
0.0075944811,
0.0027317312,
-0.0231998172,
-0.0130402157,
0.0042359731,
-0.0234836247,
0.0177145228,
-0.0469482765,
-0.0301991813,
-0.0352013223,
0.0616918504,
-0.0577251017,
-0.0820400119,
0.0423054025,
-0.0594853573,
-0.0327243283,
0.0066348822,
0.0064049568,
0.0099303834,
-0.0346954204,
-0.0281437561,
-0.0021727341,
-0.026315419,
0.0401353762,
0.0026474188,
0.0411265343,
0.0269476343,
-0.0237034392,
-0.0390547551,
-0.0485786758,
0.0108477734,
-0.0301997606,
0.0209019277,
0.0133823268,
-0.0070049679,
-0.0098363999,
0.0691751689,
0.0350059271,
0.0309960898,
0.0191085599,
-0.0381871238,
0.0110913729,
0.0262416862,
0.0744249225,
0.0839046687,
0.0078072776,
0.0052675223,
0.0557978489,
0.0161085036,
-0.0246672295,
-0.0246402211,
-0.0585269146,
-0.0123151094,
-0.0445572399,
-0.0281437673,
0.0398457535,
0.1019643992,
0.0628738329,
-0.0127777848,
-0.0351322889,
0.0576500967,
-0.0545914285,
-0.049897939,
-0.005458544,
-0.0692699477,
-0.063416101,
0.0351950191,
-0.0271545593,
-0.098319158,
0.0706040934,
0.026394818,
0.0029250279,
-0.0614002012,
0.0293975472,
0.0334082693,
0.0166429933,
0.0534089208,
0.0248723179,
0.0712610334,
-0.0386613309,
-0.0835258514,
-0.0332837366,
0.0329400599,
0.0262531023,
0.0147241689,
-0.0511790849,
-0.0014473472,
0.0618073493,
-0.0391193107,
-0.0400583372,
-0.0901634619,
-0.0000811069,
-0.0020889088,
0.0505383536,
-0.0039094486,
0.0679520518,
-0.0481275059,
-0.0149687417,
0.0209737923,
0.0132148238,
-0.040715903,
-0.0226119701,
0.0155722378,
0.0143711809,
-0.0479915328,
0.0091406545,
-0.000920912,
-0.0337717347,
0.0152731193,
0.0715132728,
0.012790259,
-0.0119715156,
0.0449309424,
0.0460958816,
-0.0480454452,
-0.0343514755,
0.0480295792,
-0.016924629,
-0.0071496652,
-0.0101784738,
0.0518525057,
0.0245512985,
-0.06583976,
0.0464870073,
0.0131103816,
0.024387259,
-0.0156587698,
-0.027835628,
-0.0782394782,
-0.0072511095,
0.018540306,
-0.0198134705,
-0.0475952066,
-0.0382210054,
0.0121252974,
-0.0753377974,
0.0064850724,
-0.0136653269,
-0.0006247217,
0.0196599364,
-0.0059247427,
0.022729665,
0.0329766423,
0.0002864696,
0.0040144385,
-0.0601279698,
0.0251304489,
0.0380473509,
-0.0067214398,
-0.0751827061,
-0.0036688608,
-0.0344811082,
0.0676893741,
0.0356070995,
0.0132329408,
-0.0266332868,
-0.020790549,
0.0146270255,
-0.0161193833,
0.0127206026,
-0.0172897168,
0.0026796546,
-0.001197713,
0.0157064125,
-0.0506630279,
0.0247906391,
0.0087124212,
-0.0565572307,
-0.0136752762,
0.0254002735,
-0.0295429584,
-0.0249846019,
0.0062531834,
0.0217930507,
-0.0019750916,
0.0758721232,
0.0162650105,
-0.0212144535,
0.0326984897,
-0.0157998614,
0.0285187867,
0.0285868421,
-0.0166245736,
0.0030952289,
-0.0518107526,
-0.0455943234,
-0.0951623917,
-0.0181006212,
-0.0210762955,
0.0125577347,
0.0310090445,
0.0251996368,
-0.0104363663,
-0.0201333407,
0.0224808715,
0.0100352494,
-0.0217108876,
-0.0146686109,
-0.0257965215,
-0.0386974029,
0.0285287686,
-0.0373425595,
-0.0277888197,
-0.0102441311,
0.0009634651,
-0.0795701593,
0.0509823151,
0.0424463935,
0.0197384264,
0.0037528209,
0.0268229824,
-0.0534662083,
0.0318443067,
-0.004163404,
-0.0277441498,
0.0469106548,
0.0206709653,
-0.0270901304,
0.0220120382,
0.0300168917,
0.0172667447,
0.0252651013,
0.0359863825,
-0.0013923836,
-0.021209022,
-0.024455525,
0.0788194388,
0.0202750824,
0.075489983,
-0.069746688,
0.030940406,
-0.0432377756,
-0.0096088341,
-0.0213204063,
-0.0151339881,
0.0268382002,
-0.0099234469,
0.0620420091,
0.0763597712,
-0.0268133692,
0.0051566903,
-0.0280425716,
0.0104396436,
0.0482582822,
-0.0305635985,
0.0493202284,
-0.0103644636,
0.0420354344,
0.0439732373,
0.0149012636,
0.0083506377,
-0.0163172614,
0.0344882831,
0.0180324186,
0.0410999656,
0.0385340825,
-0.0444952697,
-0.0547763892,
-0.0547449291,
-0.0184419621,
0.0046041505,
0.0143076973,
-0.0023308082,
0.0472510345,
-0.076618433,
0.0281282943,
-0.1046634465,
0.0324490927,
0.020985065,
-0.0547490902,
0.0001952419,
-0.0163685009,
0.0456276946,
-0.0101880608,
-0.0438437127,
0.0080707483,
0.0816574842,
-0.0078251883,
0.0383820161,
-0.0294138379,
-0.0205230564,
-0.0160404015,
0.0158571396,
0.0726975501,
0.0290486757,
0.0132573266,
0.0125585599,
-0.0525909141,
-0.041095674,
-0.0284543931,
-0.0130094867,
0.0239878464,
-0.0519767068,
0.066364944,
0.062762484,
-0.0328977779,
-0.0563969351,
-0.0328638703,
0.0256619863,
0.0038408802,
0.0317196138,
-0.038102027,
-5.843613476e-33,
-0.0055944752,
-0.0021424596,
-0.0291125085,
0.0161690079,
0.0234305169,
0.0069131679,
0.0062489319,
0.0219267681,
0.0299745072,
0.0706975386,
-0.0275415406,
0.0000302077,
0.0318367109,
-0.0089769131,
0.0267443098,
-0.0027592888,
0.0321028642,
-0.0499508157,
-0.0429831296,
-0.0549721047,
0.0168471187,
0.0149962045,
0.0313537121,
0.0365607515,
0.0551379137,
-0.0276000928,
-0.0410275944,
0.0283084102,
-0.0240660273,
0.0172868315,
-0.0215804819,
0.0486298054,
0.0429907292,
-0.0040106224,
0.0011496915,
0.0871990323,
-0.0088033974,
-0.0350862518,
0.0985720754,
0.0347664244,
-0.0164795313,
-0.0096381987,
0.0253031161,
-0.0167617183,
-0.0833024532,
0.0634396374,
0.013338591,
-0.0030674653,
0.0192240048,
-0.0267252997,
-0.0601558387,
-0.0311627239,
0.0068010897,
-0.0297586434,
0.0297092516,
-0.0087143863,
-0.0331874453,
-0.0439931452,
0.0065353792,
0.0450666212,
0.0162591338,
0.0648157597,
0.049015943,
0.0666016936,
0.0267721675,
0.0442262702,
0.0612803958,
0.0324288383,
0.0333139859,
0.0063755997,
0.0375249423,
-0.023279829,
-0.0029257971,
-0.0029861149,
0.0680109486,
0.0116428891,
-0.0094310073,
0.0011768157,
0.0775700584,
0.0211239085,
-0.0225330051,
0.0233662575,
-0.0191415939,
-0.0518137552,
-0.0620098747,
-0.0004126798,
0.0044220276,
-0.0225127768,
0.0188829098,
-0.0154302558,
0.0400625542,
0.0169036146,
0.0006333705,
0.0140625108,
-0.0254195929,
-0.0251290668,
0.008426901,
0.0185323432,
-0.0063495571,
0.0154127711,
-0.0754674673,
0.0083698193,
0.0027049081,
-0.0297144633,
-0.0152712949,
0.054232806,
-0.0217416715,
0.0073460722,
-0.050330583,
0.0086655002,
-0.016540667,
-0.033207424,
0.0402329601,
-0.0149429478,
0.0440481305,
0.0336832032,
0.0048566367,
0.046804782,
0.0474418737,
-0.0114742657,
-0.0263575576,
0.0012171433,
-0.0473311357,
0.0601169281,
0.0171136633,
-0.0066992976,
-0.0112104192,
0.0064032525,
0.0362317488,
-0.0335629731,
-0.0060211285,
0.0178935155,
2.54e-7,
0.0288840123,
0.0319357365,
0.0069446857,
-0.0115600741,
-0.005162878,
0.0442148335,
-0.0095003406,
0.0340064317,
0.0227099191,
-0.0372601449,
-0.0126048969,
-0.0145312315,
-0.0149157252,
-0.0109216236,
-0.0709172562,
-0.0253746472,
-0.053295739,
0.0030912687,
-0.019410966,
-0.0014259542,
0.02103379,
0.0382074192,
-0.0096347611,
-0.0000036625,
-0.0310586039,
-0.0247453805,
0.0305297486,
-0.0095433081,
-0.0400635451,
0.0421905369,
-0.0125916163,
-0.0249582808,
0.0031259032,
0.0883487239,
-0.034901686,
0.0174313784,
-0.0082469108,
-0.0100734662,
-0.009630152,
0.0141291879,
-0.0062932908,
-0.0008178987,
-0.0325873457,
-0.0150856627,
0.0098886183,
-0.0668563247,
-0.02280063,
-0.0019345665,
-0.0326888114,
0.0103445137,
0.0218278654,
0.0044955751,
-0.052072078,
-0.0164576098,
0.0092489384,
0.0651422888,
-0.0159707349,
-0.051687818,
-0.0306350328,
0.0946027413,
-0.0054265121,
-0.0436342619,
-0.0092043998,
-0.0162199307,
0.0383352339,
-0.0581983291,
-0.0274681617,
2.034828246e-34,
0.0115910256,
-0.0362012237,
0.0543803684,
0.0268284306,
0.0227885842,
-0.0275603198,
0.0330886208,
0.0784410611,
0.0654493123,
-0.0608195812,
0.0400507338
] |
http://arxiv.org/abs/2405.15943v1
|
http://arxiv.org/pdf/2405.15943v1
|
Transformers represent belief state geometry in their residual stream
| 2024-05-24T21:14:10 | 2024-05-24T21:14:10 |
Adam S. Shai, Sarah E. Marzen, Lucas Teixeira, Alexander Gietelink Oldenziel, Paul M. Riechers
|
cs.LG
|
cs.LG, cs.CL
|
What computational structure are we building into large language models when
we train them on next-token prediction? Here, we present evidence that this
structure is given by the meta-dynamics of belief updating over hidden states
of the data-generating process. Leveraging the theory of optimal prediction, we
anticipate and then find that belief states are linearly represented in the
residual stream of transformers, even in cases where the predicted belief state
geometry has highly nontrivial fractal structure. We investigate cases where
the belief state geometry is represented in the final residual stream or
distributed across the residual streams of multiple layers, providing a
framework to explain these observations. Furthermore we demonstrate that the
inferred belief states contain information about the entire future, beyond the
local next-token prediction that the transformers are explicitly trained on.
Our work provides a framework connecting the structure of training data to the
computational structure and representations that transformers use to carry out
their behavior.
|
Transformers represent belief state geometry in their residual stream: What computational structure are we building into large language models when
we train them on next-token prediction? Here, we present evidence that this
structure is given by the meta-dynamics of belief updating over hidden states
of the data-generating process. Leveraging the theory of optimal prediction, we
anticipate and then find that belief states are linearly represented in the
residual stream of transformers, even in cases where the predicted belief state
geometry has highly nontrivial fractal structure. We investigate cases where
the belief state geometry is represented in the final residual stream or
distributed across the residual streams of multiple layers, providing a
framework to explain these observations. Furthermore we demonstrate that the
inferred belief states contain information about the entire future, beyond the
local next-token prediction that the transformers are explicitly trained on.
Our work provides a framework connecting the structure of training data to the
computational structure and representations that transformers use to carry out
their behavior.
|
[
0.0144351665,
0.0730150715,
-0.0207884274,
0.0209838264,
-0.0518802106,
0.0006528133,
-0.0293157026,
-0.0215510465,
-0.0467244461,
-0.0311173983,
-0.0079723997,
-0.11429324,
0.0161820743,
0.0536083467,
0.0163517911,
-0.0737859309,
0.0368447267,
-0.0442375802,
-0.0009555181,
-0.0216826592,
-0.0263847485,
-0.0241340101,
-0.0027537132,
0.0517398156,
-0.0349853225,
-0.0140903508,
-0.0324015059,
-0.0206939504,
-0.0099084517,
-0.0212761983,
-0.0380312353,
0.0069319718,
0.0051960703,
0.0822618604,
0.0000016871,
-0.0342004895,
0.0021751623,
-0.0280874241,
-0.0171723496,
-0.0102443676,
-0.0163440965,
-0.0195867587,
-0.0404043049,
-0.0049115065,
-0.008889352,
-0.0266284607,
0.0567272455,
0.0688624233,
0.0303702001,
0.0805428028,
-0.0169743057,
-0.0652353987,
0.0183889493,
-0.0236601159,
-0.0160070844,
0.0404885709,
0.0496057943,
-0.0437365361,
-0.0697534159,
-0.0201728903,
-0.0484473519,
-0.0136290789,
0.0268555842,
0.0154519752,
-0.067157276,
0.0552478917,
0.000648573,
-0.046830751,
-0.0206426326,
0.0179334115,
0.0190912038,
-0.0370687395,
0.0267528463,
0.0148393735,
0.0040554209,
0.0600943863,
-0.0172935836,
0.0218676589,
-0.006869792,
0.0363793224,
0.0069694845,
0.0004976683,
0.0369881094,
-0.0356371477,
-0.0687718913,
0.1065722108,
0.0014441679,
-0.0243891105,
-0.047574386,
-0.0167595875,
-0.0111118909,
-0.0827521533,
-0.0198546145,
0.0036374403,
0.0097191157,
0.0153185613,
-0.0252302121,
0.0128533384,
-0.0513192937,
-0.0039642248,
0.0202736743,
0.0307866707,
0.0030468565,
0.0063795336,
-0.002086635,
0.0296318717,
-0.0509922802,
0.0373107456,
-0.0171435084,
0.0290621407,
0.0041651782,
-0.0156024471,
-0.0216343589,
0.020958228,
0.0278271064,
-0.0049798912,
-0.0540338494,
0.0203521233,
-0.0093339207,
-0.005755187,
-0.0263814554,
0.0069160834,
-0.0107664224,
-0.0253750347,
-0.0543718636,
0.0645446256,
0.01019309,
0.0066568633,
0.0003008198,
-0.0419991985,
-0.0139375785,
0.0186648462,
-0.0103809759,
0.0264453087,
-0.0237502977,
0.040708404,
0.0166410469,
-0.0128884055,
-0.0284529198,
-0.0129525214,
-0.0269898754,
-0.0682986453,
0.051809106,
-0.0320793614,
-0.0167014394,
0.0136737488,
0.0094563225,
-0.0246108733,
-0.0085443482,
-0.0093029365,
-0.0002656092,
0.0750547796,
0.0073382594,
0.0189737119,
-0.0071774577,
-0.0102661224,
-0.0143311955,
0.0535516925,
0.0379703455,
0.0445691608,
-0.0094141578,
-0.0068275859,
0.0194296613,
0.0213498902,
0.0132681048,
-0.0081850421,
-0.0608920604,
-0.0100371409,
-0.0140144639,
0.0194438715,
0.0257273782,
0.0465061739,
-0.0286672805,
-0.0239082836,
0.032591477,
0.0678086355,
-0.0091742678,
0.0204783678,
0.0251361355,
-0.0071136747,
0.0277355053,
0.0779721513,
-0.0538674667,
0.0217600185,
0.0363834985,
-0.0005297224,
0.0016481273,
0.0496070832,
-0.0325323492,
-0.0339728482,
-0.0198692568,
-0.0256946124,
-0.0371503942,
-0.0480228327,
0.0438452587,
-0.0025832809,
-0.0255593117,
0.0554622337,
0.0020044092,
-0.0006510719,
-0.011791151,
-0.0500406139,
0.0320565812,
0.0090559674,
0.0100777522,
0.0118618216,
-0.0234859493,
0.0198956616,
-0.0037824279,
0.0404547416,
0.0041384827,
0.0063188141,
-0.0274705347,
-0.0211645477,
-0.0130588571,
0.0344891846,
0.0340842232,
0.025586402,
-0.0330379568,
-0.0174755398,
-0.0050684046,
-0.0310789701,
0.0272459164,
0.0236044377,
0.0126282778,
0.0193809103,
-0.0221581627,
0.0279325321,
0.0527791865,
-0.0035785667,
-0.0101316711,
0.0257282387,
0.0576987639,
0.0223790854,
-0.0169134196,
0.0786595419,
-0.0733246803,
-0.0174269062,
0.0270200465,
-0.023947604,
0.0461813174,
-0.003484491,
0.0104553066,
-0.0523492172,
0.0511182472,
-0.0535217486,
0.1001142263,
-0.0088747414,
0.0230930336,
-0.0014250857,
0.0260861088,
0.014288594,
0.0301007852,
-0.0315801241,
-0.0116395159,
0.0221304316,
-0.0097017856,
0.0245464351,
0.0079276413,
-0.0009357524,
0.0261538681,
-0.0283755343,
-0.0076717152,
0.0266531035,
-0.0721859634,
-0.0134988986,
-0.0510759726,
-0.0000691853,
-0.0694283321,
-0.0183422342,
0.0167923067,
0.0073995101,
-0.0141557455,
-0.0034494265,
-0.0238358211,
-0.0222408753,
0.0262921732,
0.0432022624,
0.0272927824,
0.0101114754,
-0.0165806543,
-0.0406705923,
-0.0210025385,
-0.0221186597,
-0.0268748719,
0.0539260544,
0.0002706604,
-0.0099729085,
-0.0790056661,
-0.0357598849,
-0.0377031676,
0.0104398383,
-0.038763389,
-0.0459993221,
0.0177924931,
-0.0289165825,
0.0569630191,
-0.0031335012,
0.0100303506,
0.0319265611,
0.0434821323,
0.0201785993,
-0.0136409374,
-0.0097520389,
-0.0799671486,
0.006854408,
-0.0820267946,
-0.0251850747,
0.0214318056,
0.106621705,
0.011977803,
-0.0113749066,
-0.0698645562,
0.0016903728,
0.0062651355,
-0.0293036662,
0.0396068618,
-0.024405187,
-0.0434771366,
0.0636683106,
-0.0133800125,
-0.0817007571,
0.0611741878,
0.0059022973,
0.0007346688,
0.0093727978,
0.0398692898,
-0.0023176158,
0.0505574681,
0.0615140609,
-0.0102291359,
0.0356840864,
-0.0055719647,
-0.048565045,
0.0127037084,
-0.0299071185,
-0.0011997118,
0.017566653,
-0.0051121069,
0.0033589171,
0.0221500639,
-0.0277456511,
-0.0546943061,
-0.0086484598,
0.0058965115,
-0.0274022352,
0.0161032397,
-0.0471269712,
-0.0408448465,
0.0010032455,
0.0027603453,
0.0163568072,
-0.0135570737,
-0.0227847304,
-0.0125925252,
0.0255192313,
0.0423840731,
0.0283917543,
0.0785331652,
-0.0634060726,
0.0183463022,
0.0352327749,
0.0789127797,
0.0470421873,
-0.0027925451,
-0.0248490311,
0.0465747565,
0.0182113573,
-0.032686118,
0.0579105467,
0.0491642989,
0.0009679497,
0.0018281714,
0.1080478951,
0.0065959259,
-0.0439855382,
0.0183258411,
0.0100841904,
0.0482930802,
0.0489241183,
0.0049899439,
-0.0743348673,
0.015691001,
-0.0040816362,
-0.0045022243,
-0.0018863865,
-0.0342459008,
-0.0022575448,
-0.0490063541,
0.013192866,
0.0036256094,
0.0115172546,
-0.0197026972,
-0.0350878127,
0.0160566419,
0.0019251389,
-0.0248708222,
0.0115030864,
-0.0310522299,
0.0124779083,
0.0422823727,
-0.0224909186,
0.0126804495,
-0.0227799919,
-0.0060713626,
0.0209067687,
0.0318062156,
0.0348342508,
-0.0093415929,
-0.0687133744,
-0.0023186107,
0.0261234399,
0.0144197894,
0.022191124,
0.0368505009,
0.0253787376,
0.0041392515,
-0.0054317163,
-0.0057739187,
-0.0318623707,
-0.0264151711,
-0.0253688376,
0.075587213,
-0.045537211,
-0.0807387605,
0.0059341015,
0.0006036568,
-0.0187568422,
0.0658953115,
0.0134020308,
-0.0353213139,
-0.03816716,
-0.0512219034,
0.05799786,
0.0107223978,
-0.0216267183,
0.0003334749,
0.0617949143,
-0.0461319312,
-0.063465178,
-0.0005283809,
-0.0299371891,
0.0356431603,
0.0144814942,
-0.0574591607,
0.0450012572,
-0.0135746505,
-0.0472369045,
-0.0024245409,
-0.0118512772,
0.00427646,
-0.0779171586,
-0.0043252697,
0.0113905538,
-0.0404337421,
-0.0333300531,
-0.0155465789,
-0.0822002664,
-0.0562377237,
0.0689034835,
0.0166101865,
0.0180396661,
0.0135514131,
0.0336324126,
-0.0138862561,
0.0711645633,
0.0116132544,
-0.0433126613,
0.0427450687,
0.0423811041,
0.0011002086,
-0.0095861824,
0.0527771376,
0.0241352413,
0.0149941752,
-0.0134593574,
-0.007633531,
-0.0228769686,
-0.0224812347,
0.0546728149,
-0.0210622754,
0.092238076,
0.0012897582,
0.0280342884,
0.000217363,
-0.0012137167,
0.0053529427,
0.0171235446,
-0.0365805738,
-0.0032762345,
0.0356382765,
0.0340434276,
-0.008173896,
-0.0141248945,
0.0171230529,
0.0095117455,
0.0283955969,
-0.0298546758,
0.0354406573,
-0.0034036359,
0.0548489653,
-0.011677755,
0.0414269045,
0.0370716862,
-0.0172375906,
0.0138294725,
0.0197482705,
0.0403789543,
0.0404212549,
0.0038817334,
-0.0373165347,
-0.0096120341,
-0.1079916283,
-0.0616651848,
-0.021370355,
-0.0020453783,
0.0228007156,
0.0131748505,
0.0252212286,
-0.0830817148,
0.021336427,
-0.0099507496,
0.0249676444,
0.0012136243,
-0.0590256639,
0.0278017614,
-0.0099553652,
-0.0460599661,
-0.0058123679,
0.073155582,
-0.0060468363,
0.083531335,
-0.0328659229,
-0.0012804444,
-0.0373713002,
-0.0471414216,
-0.0356028713,
0.0470255092,
0.0007800366,
0.0133008119,
-0.0698343068,
-0.0689318031,
0.0292541869,
0.0030798595,
-0.002510055,
-0.0741146877,
0.0401978083,
0.0683140978,
-0.0233150572,
-0.0112669822,
-0.0296489336,
-0.003826384,
0.0607175976,
0.0231008623,
-0.0300098192,
-5.556217964e-33,
0.0027747317,
-0.0125780329,
0.0010224268,
0.0298922975,
0.0077278945,
-0.0028445316,
-0.0187225323,
0.0083867889,
0.0130474018,
0.0289432574,
-0.0053426633,
0.0231468789,
0.0190352835,
0.0406822674,
0.0382641852,
-0.0295173246,
0.0678131059,
-0.0566993728,
-0.0143411392,
0.0163345449,
0.0427537262,
-0.0090383356,
0.0951199234,
-0.0209319238,
0.0143405711,
0.0214135163,
-0.0159004331,
0.0455281548,
-0.0386108644,
0.0204013605,
-0.043060191,
0.009134775,
0.0042495853,
-0.0391929895,
-0.0349545442,
0.0047338274,
0.0005932587,
-0.0357218273,
0.0172574613,
0.0365465842,
-0.0446771048,
-0.0090471515,
0.0283264369,
-0.0230005272,
-0.0584293306,
0.0117213884,
0.0266898982,
0.0228167288,
-0.0043269279,
-0.0111685311,
-0.0208974462,
0.0034177634,
0.0036357529,
0.0028685206,
0.0277948603,
0.0406235158,
0.0274183061,
-0.0451936498,
-0.0640067384,
0.0765959173,
-0.0279201046,
0.0481286868,
0.0236308426,
0.0549484305,
0.0277327094,
0.0753149986,
0.1277735233,
0.0016339505,
-0.0354359895,
-0.0259884875,
0.0458898358,
-0.0041007977,
-0.0292686298,
0.0072102295,
0.0455320813,
-0.0013532842,
-0.0033218935,
-0.0287017077,
0.0060367314,
0.0281703416,
0.0222714841,
-0.006948716,
-0.0178135224,
-0.0364124365,
-0.0533457696,
-0.0122541664,
-0.0042835688,
0.0157511253,
-0.0278000496,
0.0351635665,
0.0555053167,
0.0048319325,
0.0329430364,
-0.0356835574,
-0.0336396247,
0.0033714662,
-0.0203426387,
0.02148249,
0.0208687205,
0.0172951482,
-0.0468282886,
0.0108915642,
-0.0261806566,
-0.0034242442,
-0.0190268885,
0.0358437039,
-0.0098873954,
0.0443236418,
-0.0672215521,
-0.0016484773,
-0.0184276234,
-0.0382800028,
0.0023118993,
-0.0358520709,
0.0966956541,
0.0235180203,
-0.0185023807,
-0.0444444753,
0.0121046109,
0.0288523249,
-0.0241778605,
0.0492864102,
-0.0035083941,
0.0687314868,
-0.0496676117,
-0.0449149609,
-0.1028748453,
-0.0055680913,
0.0387321413,
-0.0782666877,
-0.0385970511,
0.0263874959,
2.55e-7,
0.0139979608,
0.0489221998,
0.0326149985,
0.0135113103,
0.0756179988,
0.0154168289,
0.0046002879,
-0.0118240779,
0.0519926585,
-0.0299729854,
0.002668869,
-0.0159224197,
0.0295018367,
0.0212321002,
-0.1018415019,
0.0048220586,
-0.0641256943,
0.0111984992,
-0.0059623923,
0.0366246812,
0.0021194543,
0.0816887543,
0.0346253477,
-0.0099090422,
-0.0040616393,
0.0359269679,
0.0672234371,
-0.0022480616,
0.0361863673,
0.0369270593,
-0.0143675404,
-0.0440355577,
-0.0038729955,
-0.0141846137,
0.0175919775,
0.021407472,
-0.048301518,
0.0176492911,
0.0090727797,
-0.0065102158,
-0.0061196405,
0.0111453123,
-0.0388829298,
-0.0062265527,
0.0203839839,
-0.0664049685,
-0.0304347202,
-0.0377664454,
0.0125649842,
0.0107943909,
-0.0007082755,
0.0414873324,
-0.0277099479,
0.0295670871,
-0.0043179137,
-0.0014958925,
0.0281081833,
-0.0438550003,
-0.041992344,
-0.0124890255,
-0.0788502619,
-0.0379060544,
-0.0639646873,
-0.0327411927,
0.0620747581,
-0.0676539168,
-0.0058668512,
2.488669166e-34,
0.0355089381,
-0.0219019931,
0.0047901929,
-0.0063165855,
0.0011635268,
-0.0629288554,
0.0708111599,
0.0158661846,
0.0399755016,
-0.053055875,
-0.030747233
] |
http://arxiv.org/abs/2405.15936v1
|
http://arxiv.org/pdf/2405.15936v1
|
Zero-Shot Spam Email Classification Using Pre-trained Large Language Models
| 2024-05-24T20:55:49 | 2024-05-24T20:55:49 |
Sergio Rojas-Galeano
|
cs.CL
|
cs.CL, cs.AI, I.2.7
|
This paper investigates the application of pre-trained large language models
(LLMs) for spam email classification using zero-shot prompting. We evaluate the
performance of both open-source (Flan-T5) and proprietary LLMs (ChatGPT, GPT-4)
on the well-known SpamAssassin dataset. Two classification approaches are
explored: (1) truncated raw content from email subject and body, and (2)
classification based on summaries generated by ChatGPT. Our empirical analysis,
leveraging the entire dataset for evaluation without further training, reveals
promising results. Flan-T5 achieves a 90% F1-score on the truncated content
approach, while GPT-4 reaches a 95% F1-score using summaries. While these
initial findings on a single dataset suggest the potential for classification
pipelines of LLM-based subtasks (e.g., summarisation and classification),
further validation on diverse datasets is necessary. The high operational costs
of proprietary models, coupled with the general inference costs of LLMs, could
significantly hinder real-world deployment for spam filtering.
|
Zero-Shot Spam Email Classification Using Pre-trained Large Language Models: This paper investigates the application of pre-trained large language models
(LLMs) for spam email classification using zero-shot prompting. We evaluate the
performance of both open-source (Flan-T5) and proprietary LLMs (ChatGPT, GPT-4)
on the well-known SpamAssassin dataset. Two classification approaches are
explored: (1) truncated raw content from email subject and body, and (2)
classification based on summaries generated by ChatGPT. Our empirical analysis,
leveraging the entire dataset for evaluation without further training, reveals
promising results. Flan-T5 achieves a 90% F1-score on the truncated content
approach, while GPT-4 reaches a 95% F1-score using summaries. While these
initial findings on a single dataset suggest the potential for classification
pipelines of LLM-based subtasks (e.g., summarisation and classification),
further validation on diverse datasets is necessary. The high operational costs
of proprietary models, coupled with the general inference costs of LLMs, could
significantly hinder real-world deployment for spam filtering.
|
[
0.0516937897,
0.0328351259,
0.0046606804,
0.0329618044,
-0.0431324728,
0.027656246,
0.0685346648,
-0.0026786251,
0.0597373433,
-0.0978553891,
-0.0107006542,
-0.0433178321,
-0.0256283656,
0.0442821831,
-0.0095752766,
-0.0319321416,
-0.0099526923,
-0.0176519845,
-0.013993551,
0.0053724512,
0.0015647419,
0.0154101355,
0.0107148802,
0.0036667164,
-0.0180163383,
0.0070425374,
0.0025985593,
0.0076665287,
0.0028949631,
-0.0203645881,
0.0183454771,
0.0242856275,
0.0155846542,
0.0479059629,
0.0000019144,
-0.0835378766,
-0.0012276544,
0.0112693096,
-0.0600018576,
0.0323298723,
0.0722993016,
-0.0344492011,
0.0062109632,
0.0572703518,
-0.0280727521,
-0.015154724,
0.0213849433,
0.0616595782,
0.0624192022,
0.0934640765,
-0.0191410389,
-0.0507769547,
0.0493904948,
-0.051192075,
0.0074529196,
-0.0073117889,
0.0585048981,
-0.0688471645,
-0.0090759704,
0.0721802339,
0.0100846253,
0.0095971907,
-0.0075175702,
-0.0052768704,
-0.0124074742,
0.0259949639,
-0.0362274908,
0.0082866568,
-0.0084372666,
0.0084049879,
-0.0366370939,
-0.0460997932,
0.0120450268,
0.0013648951,
-0.014674155,
-0.0386601947,
-0.0434429124,
0.0750650242,
0.0295280162,
0.0248609614,
-0.0050127306,
0.0074914903,
0.0053488715,
0.0609074347,
-0.0416924879,
0.0079509327,
-0.0074662529,
-0.037654113,
-0.0120816063,
-0.0025666966,
0.0436209179,
-0.0795774981,
-0.0024961913,
0.0279587097,
0.0432048813,
-0.0175246634,
-0.0522773713,
-0.0258494206,
0.0402783081,
-0.0747310668,
0.0097508188,
-0.0002845633,
0.0139650134,
0.0573823862,
-0.0185261983,
-0.0057833688,
0.0123619223,
0.0259622913,
-0.0373797156,
0.0047539854,
-0.0744052082,
-0.0290046185,
-0.0399217531,
0.0648166016,
0.0050393422,
-0.003236454,
-0.101564914,
0.0111607797,
-0.0110331997,
0.0141883064,
-0.0044527734,
-0.0364204422,
0.0224460028,
0.0068506654,
-0.0066516199,
0.0013499922,
0.0007297994,
0.0336142592,
-0.0101971207,
-0.0401436388,
-0.0177141819,
0.0309234373,
-0.0182026289,
-0.0409837179,
0.0436720178,
0.0038752665,
0.01359843,
-0.0813524574,
-0.0106028747,
-0.009248117,
-0.0488942154,
-0.04541976,
-0.0392937809,
-0.0330859385,
-0.0319203734,
-0.005007131,
0.008743111,
0.0145843774,
-0.0321936235,
0.0190137383,
-0.0680473074,
0.047376588,
0.0157719366,
0.0084979963,
0.0185124576,
-0.0249597766,
0.0525271595,
0.0326553434,
-0.0445648991,
0.08322566,
0.0256533902,
-0.0119256238,
0.0399950519,
0.0415167212,
0.0348835848,
0.0134074222,
-0.0159076061,
-0.0121368514,
0.0163835138,
-0.0565572418,
-0.0026328349,
0.0276400931,
-0.0061360784,
-0.0069439169,
0.0835706517,
0.0399690792,
-0.0073954077,
0.0034995372,
0.0116896592,
0.0413540564,
0.0065030581,
-0.0337354764,
-0.0622543767,
-0.0290894937,
-0.0822589248,
-0.0068878741,
0.0269726701,
0.0253432281,
-0.0613687448,
-0.0333236009,
-0.0015412789,
-0.0161599889,
0.0012865023,
0.0356224068,
0.0907141194,
0.0090107331,
-0.0368942879,
0.0254725218,
0.0023091568,
0.0026089884,
-0.0280539393,
-0.0459802449,
0.0338614881,
0.0037418844,
0.0476335436,
-0.0209449567,
-0.0134649426,
-0.0415097252,
0.0195050053,
-0.0050460193,
-0.0016757684,
0.0443276763,
0.0074143526,
-0.0170444138,
-0.0131431371,
-0.0139601212,
0.0119604133,
-0.0338094272,
-0.0064739375,
0.018174734,
0.0550911203,
0.0489224754,
-0.0387082212,
0.0238254517,
0.0318329185,
-0.0311029572,
0.0615763441,
0.0003142635,
0.0526063964,
0.0308338776,
-0.0359725542,
-0.0163510349,
0.0464282446,
-0.042648375,
-0.0013658316,
-0.0762233883,
0.0137600917,
0.0131737767,
0.0574482083,
0.0205786247,
0.009437914,
0.0381675139,
-0.0461855531,
0.015362801,
0.0151949441,
0.0274264421,
0.0020403762,
-0.0727630034,
0.0009830921,
-0.0158250779,
0.0383133739,
-0.0147904176,
0.0341476984,
0.0260249581,
-0.0116453031,
-0.0193061903,
-0.0342128798,
0.0048264856,
0.0010675596,
-0.0538781285,
0.0086809499,
-0.0339904241,
0.0117617464,
0.0078988224,
0.0417796113,
-0.0098614385,
-0.0331666917,
-0.1179494709,
0.0245097913,
-0.024214359,
0.0279934667,
-0.0039624367,
0.0119487811,
-0.0114003103,
0.004528007,
-0.0508400835,
-0.0033709118,
0.0332886204,
0.0025235373,
0.0250240806,
-0.0173869245,
-0.010095248,
-0.0031726537,
-0.0100178802,
-0.0033092247,
0.0309361331,
0.0308006741,
0.0020620173,
0.0142243905,
0.0297296476,
-0.000174064,
0.0128234169,
0.0157694835,
-0.0669888854,
-0.0282634478,
-0.0409119949,
-0.0302555002,
0.006439561,
0.0518197268,
0.0301407073,
0.0224262308,
0.0511116199,
-0.0330124944,
0.0095583815,
-0.0389725305,
-0.0250791889,
-0.024938222,
0.0142393894,
0.0439010039,
0.0569871292,
-0.0019951349,
-0.0365735888,
-0.0019251729,
0.0667278543,
0.015194701,
0.0151562393,
-0.0554074422,
-0.0197419897,
0.0028095962,
0.0063373023,
-0.0068993615,
-0.0918052495,
0.0427478552,
-0.0596779287,
-0.0147611881,
-0.0384334251,
-0.0536556877,
-0.0388629474,
0.0398643203,
0.0200660266,
0.0061362619,
0.0253269412,
0.0139473584,
0.0180693101,
-0.0054015936,
0.0181036796,
0.0258099306,
0.0194651671,
0.0113166831,
0.0638031065,
-0.0085946769,
-0.0039722701,
-0.057214465,
-0.0982377157,
-0.0383674391,
-0.0069070728,
-0.0187841598,
0.0145786665,
-0.0669017434,
0.0317546576,
0.0205524918,
0.0479170606,
-0.005625145,
-0.0136270337,
0.0067509632,
-0.0051244949,
0.0039359732,
-0.0752711147,
-0.0108010704,
-0.0310086794,
0.0043568024,
0.0360646471,
0.0100392215,
0.0026117773,
-0.0202177335,
-0.0120501556,
0.0129666161,
-0.0625488684,
-0.0065421211,
-0.003009021,
-0.0072840843,
-0.0281979237,
0.0018651703,
0.1181851476,
0.0389028527,
-0.0450789966,
-0.0231353249,
0.0401957184,
0.0143972924,
-0.0069343657,
0.0195685327,
-0.0667321682,
-0.022230003,
0.0074646166,
-0.0085036131,
-0.0342833586,
-0.0180303846,
0.0006717988,
-0.128420487,
0.0496334918,
0.0071794176,
-0.0536803789,
-0.0260233786,
-0.0511851721,
-0.0456470661,
0.0088337986,
-0.025375763,
-0.0381647162,
-0.0408056043,
-0.0038250524,
-0.0112528782,
-0.043721959,
-0.0176095795,
-0.0169463884,
-0.0300882328,
0.0305873509,
0.0399888866,
0.006891143,
-0.0755705535,
0.0205583982,
0.0452464186,
-0.0483815297,
-0.0274538659,
-0.0254704375,
0.0926759988,
-0.0156878829,
0.0039904411,
-0.0181129202,
-0.0024186906,
-0.0345766023,
0.016838897,
0.018792944,
0.063140817,
-0.0149298292,
-0.0203753579,
-0.0168604944,
-0.0088481344,
-0.0139327757,
0.089617081,
0.0234283339,
-0.0547431521,
-0.0334122851,
-0.0174719002,
-0.0248707309,
0.0365772508,
0.0017125221,
0.0354633629,
-0.048828654,
-0.0621122979,
-0.0416436009,
-0.0273594614,
0.0390981734,
0.0216728076,
-0.0045508542,
0.0052411961,
-0.0484513976,
0.0016236393,
-0.0177199189,
0.0179673657,
-0.0280497354,
0.0559031218,
-0.0767605603,
0.0100186728,
0.028703738,
-0.0457879417,
0.0047253133,
0.0011935841,
-0.0793374702,
-0.0562995113,
-0.0352640189,
0.0568452589,
0.0277076084,
0.0093582785,
-0.0248910058,
-0.0750766024,
0.0853563473,
0.0343008265,
-0.0138715012,
0.002010324,
0.0487590879,
-0.0282145683,
0.0169996545,
0.0185050759,
-0.0139469132,
0.0603853911,
0.0746137127,
-0.0552490316,
-0.0349184312,
-0.0139605822,
0.007924323,
0.0376484431,
0.0487213805,
0.0125435553,
0.0009427143,
-0.0074602691,
0.0008096263,
-0.0217613429,
-0.0363373533,
-0.0269716494,
-0.0218934193,
0.0332522504,
0.0022071614,
-0.002566657,
0.0179102719,
-0.0407425165,
0.031667836,
0.0057785455,
-0.0104410984,
-0.016363943,
-0.0015275999,
0.0428326204,
0.0153581006,
-0.0154393315,
0.001079174,
0.009262519,
-0.0231379475,
-0.0082004033,
0.0168062709,
-0.0044146185,
0.0160220005,
-0.0481616072,
0.0230319034,
-0.0247446392,
-0.0183881149,
-0.0080656549,
0.0293779448,
0.0560508817,
-0.0448038131,
0.0354578085,
-0.0235405453,
0.0045772092,
-0.0486659221,
0.0215150602,
-0.0004338417,
0.0002912916,
-0.0195407383,
-0.0064871237,
-0.0513868406,
0.0124180475,
0.08175309,
0.0017336025,
-0.0017411431,
0.0559935682,
-0.0517628044,
-0.0114820004,
0.000326819,
0.0409014039,
0.0298799537,
-0.0237078164,
0.0189305358,
-0.0466000214,
-0.0486012846,
0.0082285879,
-0.0315670557,
0.0051904339,
0.014572111,
0.0306294449,
0.0619032793,
0.0027645638,
0.0433970802,
-0.0513788313,
-0.0176323168,
0.0441279747,
-0.0532077029,
0.0019938634,
-5.98205365e-33,
0.0579329021,
-0.0070114168,
-0.0065406775,
0.0166976899,
-0.0052809012,
0.0193283875,
0.0122742616,
-0.0111368578,
0.0085940994,
0.0171329062,
-0.0285922326,
-0.0254976377,
-0.003807896,
0.0835511908,
0.0036622151,
0.0169688612,
0.0585221611,
-0.0031847609,
-0.0062996261,
0.0137694525,
0.0419619568,
0.0125738094,
0.0017633463,
-0.0089347744,
0.0232580025,
-0.0036205114,
-0.0396126434,
0.0078435084,
0.0316601209,
0.0542597845,
-0.0417113453,
0.0670030266,
0.0599835366,
-0.0161491372,
0.0047733188,
0.0156204989,
-0.0028567205,
-0.0231893174,
-0.0035975864,
-0.0131468894,
0.0430578813,
-0.0314429067,
0.0451204367,
0.0253256224,
-0.008261756,
-0.0020368027,
0.0249930173,
-0.0339918472,
-0.0029099379,
-0.0605588816,
0.0042866454,
0.0037006401,
-0.0197122544,
-0.0114795379,
0.0522740893,
0.044644706,
0.0023878657,
0.0509305075,
-0.0494354889,
0.0332799405,
0.0452976264,
-0.0282154437,
0.0155098541,
-0.0013917357,
-0.0108423242,
0.0375386141,
0.0114437444,
0.0217469092,
-0.0473547801,
0.0103241103,
0.0228220038,
0.0044377712,
0.0292308126,
0.0515873544,
0.0568308644,
-0.023085665,
-0.0768996701,
0.024657758,
-0.0012374026,
0.0129772257,
0.0324469022,
-0.0095267026,
-0.0143511333,
-0.0164900571,
-0.0085053761,
0.0234324969,
-0.0310712922,
-0.0248455107,
0.0145574678,
0.034187831,
0.0132950312,
-0.0176228713,
0.0310487524,
-0.0450362749,
-0.0305071864,
0.0105193462,
0.0129246591,
0.0193148497,
0.0060835192,
-0.0064989962,
-0.0352601372,
-0.0212946143,
-0.0004195625,
-0.0570783019,
-0.0201306492,
-0.0129554272,
0.055650264,
0.0073949853,
-0.0774228647,
0.0009299094,
0.0513902754,
0.0510509796,
0.0237443112,
0.0657335296,
-0.0248380285,
0.031206226,
0.0454399139,
0.0735766515,
-0.0015436976,
-0.0327124335,
-0.00440779,
-0.0000313576,
0.0034166754,
0.0042852252,
-0.0339564718,
-0.0150845004,
-0.0277268179,
-0.0091162585,
0.0549165457,
-0.0725953653,
0.008036308,
-0.0186317377,
2.644e-7,
0.0342013724,
0.0443832576,
-0.028303843,
-0.0120318653,
0.0074630035,
0.0689306408,
-0.0009581614,
0.0442037433,
-0.0286127068,
0.0233337376,
0.0058565773,
0.0183459893,
0.0337084942,
0.006424265,
-0.0809489265,
0.0018487069,
-0.0447963625,
-0.0046714591,
-0.0879845992,
0.0453024507,
0.0703153163,
0.1324021369,
0.0641365796,
-0.0229565986,
-0.0172062889,
-0.0369469561,
0.0344348587,
-0.0297156572,
-0.0513113514,
0.0204998665,
0.0718565881,
-0.0197237227,
0.0039317114,
0.0427502654,
-0.0339856967,
0.0165014919,
-0.0160839725,
0.0628587753,
0.0061063715,
0.0373098776,
0.0114865368,
-0.0066630137,
0.0018228963,
-0.0815808475,
0.0953316912,
-0.0717076138,
-0.014366529,
-0.0155710774,
-0.0097229769,
-0.0600095205,
0.0507475249,
-0.0301596709,
-0.0143291513,
0.0380956978,
0.0299889967,
-0.0473623089,
0.0247037876,
-0.0787698701,
0.0296026021,
0.0051876171,
-0.014973809,
0.0061580846,
-0.012874906,
-0.0284475237,
0.0689387172,
-0.016662797,
0.0390880592,
2.689737934e-34,
0.0434974544,
-0.0349564143,
-0.0095892092,
0.0159089118,
-0.0052277013,
0.0059754523,
0.0114498865,
0.0234818645,
-0.0029138024,
-0.1079752967,
0.0281004794
] |
http://arxiv.org/abs/2405.15924v3
|
http://arxiv.org/pdf/2405.15924v3
|
SLIDE: A Framework Integrating Small and Large Language Models for Open-Domain Dialogues Evaluation
| 2024-05-24T20:32:49 | 2024-05-30T02:13:56 |
Kun Zhao, Bohao Yang, Chen Tang, Chenghua Lin, Liang Zhan
|
cs.CL
|
cs.CL
|
The long-standing one-to-many problem of gold standard responses in
open-domain dialogue systems presents challenges for automatic evaluation
metrics. Though prior works have demonstrated some success by applying powerful
Large Language Models (LLMs), existing approaches still struggle with the
one-to-many problem, and exhibit subpar performance in domain-specific
scenarios. We assume the commonsense reasoning biases within LLMs may hinder
their performance in domainspecific evaluations. To address both issues, we
propose a novel framework SLIDE (Small and Large Integrated for Dialogue
Evaluation), that leverages both a small, specialised model (SLM), and LLMs for
the evaluation of open domain dialogues. Our approach introduces several
techniques: (1) Contrastive learning to differentiate between robust and
non-robust response embeddings; (2) A novel metric for semantic sensitivity
that combines embedding cosine distances with similarity learned through neural
networks, and (3) a strategy for incorporating the evaluation results from both
the SLM and LLMs. Our empirical results demonstrate that our approach achieves
state-of-the-art performance in both the classification and evaluation tasks,
and additionally the SLIDE evaluator exhibits better correlation with human
judgements. Our code is available at https://
github.com/hegehongcha/SLIDE-ACL2024.
|
SLIDE: A Framework Integrating Small and Large Language Models for Open-Domain Dialogues Evaluation: The long-standing one-to-many problem of gold standard responses in
open-domain dialogue systems presents challenges for automatic evaluation
metrics. Though prior works have demonstrated some success by applying powerful
Large Language Models (LLMs), existing approaches still struggle with the
one-to-many problem, and exhibit subpar performance in domain-specific
scenarios. We assume the commonsense reasoning biases within LLMs may hinder
their performance in domainspecific evaluations. To address both issues, we
propose a novel framework SLIDE (Small and Large Integrated for Dialogue
Evaluation), that leverages both a small, specialised model (SLM), and LLMs for
the evaluation of open domain dialogues. Our approach introduces several
techniques: (1) Contrastive learning to differentiate between robust and
non-robust response embeddings; (2) A novel metric for semantic sensitivity
that combines embedding cosine distances with similarity learned through neural
networks, and (3) a strategy for incorporating the evaluation results from both
the SLM and LLMs. Our empirical results demonstrate that our approach achieves
state-of-the-art performance in both the classification and evaluation tasks,
and additionally the SLIDE evaluator exhibits better correlation with human
judgements. Our code is available at https://
github.com/hegehongcha/SLIDE-ACL2024.
|
[
0.0499519184,
0.050833147,
0.0103687141,
0.0624702945,
-0.0265056938,
0.0003731829,
0.0059696171,
-0.0102519942,
0.0408517905,
-0.0655859262,
-0.05454183,
-0.0420318544,
0.0011108537,
0.0398259163,
-0.0046691764,
-0.0596570857,
0.0363966301,
-0.0176741406,
0.015664719,
-0.0082414038,
-0.0008511105,
0.0056310012,
-0.0163427368,
0.0681895688,
0.0394594446,
-0.04250101,
0.0026849329,
-0.0051167221,
0.0195276123,
0.0049331053,
0.0104244649,
0.0556870326,
0.019168919,
0.0477933586,
0.0000015675,
-0.0513190329,
0.0394001119,
-0.006317371,
0.0060758963,
0.009698634,
0.0486491099,
0.0307039842,
0.0124780815,
0.0350221135,
-0.057477925,
-0.0299478266,
0.0295092333,
0.0446605869,
0.0782374516,
0.0654711872,
0.0052334201,
-0.048817981,
0.0069890628,
-0.0203873571,
-0.0318040065,
-0.085005179,
0.0068026362,
0.0284769367,
0.009100195,
0.021222597,
-0.0028614192,
0.0166918859,
0.0086021759,
-0.0604244955,
-0.0095930602,
0.0001373266,
-0.0395120457,
-0.0036130345,
-0.0264057945,
0.1230966672,
0.0383593775,
0.0274426434,
-0.023919668,
-0.0096054571,
0.0263838582,
-0.0166811403,
-0.0517059043,
-0.0069348947,
0.0304749608,
-0.0029409018,
-0.0219589043,
0.0423835106,
-0.0077384398,
0.0499853194,
0.0014986962,
0.0284290891,
0.0047177016,
-0.0157186929,
-0.0197266098,
-0.0051794089,
-0.0445969589,
-0.0335432775,
0.0025935457,
0.0502587445,
-0.0515945889,
0.0064919032,
-0.0316001326,
-0.0083035547,
0.0265333299,
-0.0700807944,
0.0312660374,
0.0271481201,
-0.0205130167,
0.075600639,
-0.0051139905,
0.0772764087,
-0.0224050917,
0.0014127638,
-0.0265006199,
-0.0142781325,
-0.0482244305,
0.0078254407,
-0.017450029,
0.03611736,
-0.0857412666,
-0.0359951407,
-0.0195499342,
0.0229018535,
-0.020421667,
0.0217033457,
0.0207441505,
0.0420313887,
0.0024094237,
-0.0251317918,
-0.0479144156,
0.0795526654,
-0.039377965,
0.0173911769,
0.0130441263,
-0.0457318313,
-0.0149977449,
0.0491364338,
-0.0086656408,
-0.0088271564,
-0.0191204213,
0.0705212951,
-0.0033781563,
-0.0200017411,
-0.0089696385,
-0.0053156819,
-0.0572187193,
-0.055940453,
-0.0296311714,
-0.0469818413,
0.0097516794,
0.0107661653,
-0.0265118852,
-0.0428285114,
0.0174112096,
0.0274366885,
-0.0199929811,
0.0329344571,
0.0403144881,
0.0007521664,
0.0343193933,
0.0250548981,
-0.0213224813,
0.0464588851,
0.0297197215,
-0.0104604568,
0.0484545827,
-0.0174393896,
0.0082921935,
0.0104475729,
-0.0039236336,
-0.0040972284,
0.0086220093,
0.0648114309,
0.005325289,
0.0344273411,
0.031827908,
0.0592677854,
-0.034329813,
-0.0388994999,
0.0647912621,
0.0516849793,
-0.0066965879,
0.0262681749,
0.0441190377,
0.0521500446,
0.0054774205,
0.0059087882,
-0.0108577283,
0.0150238518,
-0.0118505303,
-0.0105430167,
0.0731616318,
0.0067374134,
-0.0391991138,
-0.045389358,
-0.0039676377,
-0.0031453497,
0.0132532613,
-0.0042552752,
0.0161832888,
0.0173233524,
-0.0554177649,
0.0378284864,
0.0206684377,
-0.034508191,
-0.0019027854,
-0.0445143916,
0.0114705451,
0.0229205154,
0.0379500017,
-0.0042789183,
-0.0180311091,
-0.0496537685,
-0.0022613395,
0.0717065707,
0.0798123479,
-0.0302192364,
0.0173976794,
-0.0637003258,
0.0023429377,
-0.0333598182,
0.0005369643,
0.027637599,
0.0124709988,
-0.0043385997,
0.016410552,
-0.023810247,
0.0104439817,
-0.0202190541,
-0.0049306033,
-0.0279973093,
0.0304655917,
-0.0147622013,
0.0400002412,
0.0401095897,
0.0007407147,
0.0107912673,
-0.0062414645,
-0.0143922996,
-0.0209061708,
-0.0178596359,
-0.0711463094,
-0.0556336045,
0.0352718979,
0.0075756386,
-0.0112240994,
0.0337192863,
-0.0432772264,
-0.0387170054,
0.0404137634,
-0.0258896612,
0.0084970016,
-0.0672236457,
-0.0081142206,
-0.0059040911,
0.0083323587,
-0.0475710258,
0.0704297423,
0.0080803698,
0.0288961139,
0.0220454764,
0.0253536403,
0.0468538217,
-0.0619916767,
-0.0458194353,
-0.0015755245,
-0.0367093198,
0.0514179803,
0.0117152352,
0.023161063,
-0.0396243706,
-0.0433010086,
-0.0896377861,
0.0082273819,
-0.0255771466,
0.0176678523,
0.0055917227,
-0.006350223,
0.0188696627,
-0.0079252077,
-0.0412205979,
-0.0113414172,
0.0448613651,
0.0161826741,
0.0302070174,
0.0249988493,
-0.0115951849,
-0.0250732563,
0.0104573527,
0.0133650741,
0.0752303675,
-0.0183327924,
0.0473546647,
-0.0464434884,
0.0207229219,
-0.0393997841,
0.0136829177,
0.0034780174,
-0.0442802198,
0.0124445939,
0.0252152029,
0.012465612,
0.0580824651,
0.0334975161,
0.017202979,
0.0020892357,
-0.0024066947,
-0.0531766191,
0.0213098619,
-0.0317237936,
-0.0210658852,
-0.0433326326,
0.0364771821,
0.0024122267,
0.0389778055,
0.0182830822,
0.0126305791,
-0.0655649751,
-0.0004559423,
0.0067604138,
-0.0341621712,
-0.0221696198,
-0.0244101621,
-0.0358186811,
-0.0012392094,
-0.0104830889,
-0.0124360155,
0.0753783509,
-0.0423678197,
-0.0141704772,
-0.0647936463,
-0.0106526697,
-0.0522217117,
0.0273658764,
-0.039779298,
0.0235665925,
-0.006100859,
-0.0345369726,
-0.0352898054,
-0.0361615978,
0.0098756636,
-0.055111941,
0.0479831249,
-0.0280633643,
-0.010731387,
-0.0122461347,
-0.0663577393,
-0.0549808927,
-0.0419486761,
-0.0217339508,
0.0304017998,
-0.0135429911,
-0.0109049864,
-0.0489517972,
0.0179786105,
0.0042201434,
0.010666674,
-0.0031804487,
0.0178190246,
-0.003882128,
0.0730284154,
0.0400456637,
0.0696404725,
0.0165234078,
0.017014239,
0.0642092898,
0.0370703526,
0.0249789786,
-0.0083596194,
-0.0311448779,
-0.0097743319,
0.0174782127,
-0.0033746059,
0.0036793654,
0.0168666467,
0.0668716952,
-0.025762625,
-0.0198632777,
0.1196701601,
0.0124126738,
-0.0141488388,
-0.0107678501,
-0.0477596372,
0.0210254062,
0.0076322462,
0.0071487855,
-0.0526421443,
-0.0527409799,
-0.0189703442,
0.0249549225,
0.005254955,
-0.0381522663,
0.005510991,
-0.1308920532,
0.0115510216,
-0.0092928763,
-0.025392523,
-0.0509666987,
-0.0241393391,
0.0238341428,
-0.0179433059,
0.0328412093,
-0.0088265827,
0.0153407054,
-0.0157504752,
0.0012498099,
0.0138704358,
-0.0500194058,
0.0478719436,
-0.0184872635,
0.062642552,
0.0601572692,
0.0468159616,
0.0292557199,
-0.0336135887,
0.0181907229,
-0.0589909069,
-0.0101374816,
-0.0008413855,
0.0316202268,
-0.0019650585,
0.0014530215,
-0.0294185355,
0.0176561251,
-0.0175527968,
-0.0243517235,
-0.0087971548,
0.0543767698,
-0.0172746051,
0.0135789746,
-0.0281386804,
0.0509874597,
0.0281121656,
0.1008260772,
0.0402384326,
-0.030194724,
-0.025681939,
-0.0229007453,
-0.0338018015,
-0.0178804547,
-0.0305208191,
-0.0162729472,
0.060345754,
0.0296441521,
-0.0671640038,
-0.0368448123,
0.0026819475,
0.00719631,
-0.038300734,
-0.0126769198,
-0.0234960113,
0.0309656169,
0.0012209589,
0.0428153053,
-0.0309708137,
0.0064023412,
-0.072130397,
0.0379111357,
0.0699667186,
-0.0170371495,
-0.0559777915,
0.0369211547,
-0.0701268911,
-0.0547199398,
-0.0348030217,
0.0330596752,
0.0187398251,
-0.0140877496,
-0.0335793793,
-0.0429827198,
0.0583670512,
0.0276410729,
-0.0572894402,
-0.0091089066,
0.0633125976,
0.0395140462,
-0.0252040662,
0.0542861335,
-0.0236137509,
-0.0090780789,
0.0260746609,
0.0696069524,
-0.0120505821,
-0.0937832966,
-0.0349176377,
-0.0204230808,
0.0406937562,
0.0351478495,
-0.0274293199,
-0.0061701494,
0.0201001484,
0.0022469424,
-0.0236224327,
-0.0665383041,
-0.0015156931,
0.0160479117,
-0.0261830259,
-0.0170147121,
0.0150699271,
-0.0315163024,
0.0670084283,
-0.0191639811,
0.0245722234,
0.0653050765,
-0.0209703278,
-0.0022761303,
-0.0050206571,
-0.0612382293,
-0.0016781804,
-0.0050888816,
0.0086950529,
0.0193382595,
0.0972568095,
0.0019507973,
-0.0123867057,
0.0399174727,
0.0232527088,
-0.0616142973,
-0.0140116811,
-0.0237849765,
0.038781397,
0.025417747,
0.0209946949,
0.0359938964,
0.0125737432,
-0.0125207528,
-0.0372288972,
0.0488890186,
-0.0188431852,
-0.0449468382,
0.0546400584,
0.0293860529,
-0.0173897408,
0.0311111137,
0.0646702722,
-0.0369062237,
0.0051379227,
0.0270744935,
0.0107285231,
-0.0412001871,
-0.0673680976,
0.0227062665,
0.0243895222,
0.0016066249,
0.024955228,
-0.0643186271,
-0.0421183258,
0.0006351227,
-0.0039552348,
-0.0304240361,
0.0168507472,
0.0352184772,
0.0472447649,
-0.0026016475,
-0.0045013535,
-0.0375466347,
-0.0061457041,
-0.0316588916,
-0.0162635949,
-0.0622717701,
-5.380215236e-33,
-0.0277810842,
-0.0398123302,
0.0002282593,
-0.0210700333,
-0.0133001227,
-0.0297338758,
-0.0359268039,
0.0275010876,
-0.0559935458,
0.012115337,
0.0124802915,
0.0213654321,
0.0009575593,
0.0381663963,
-0.0398358069,
-0.0598631091,
0.0367853418,
0.0033200746,
-0.0285773668,
0.0414606817,
0.0184357706,
-0.0177523587,
0.0121438578,
-0.0690215006,
0.0282075331,
-0.0001074214,
0.0019399539,
0.0326700136,
0.0205951929,
0.0868676752,
-0.0244916119,
0.0809044167,
0.0151931047,
-0.0377448089,
0.0058107222,
0.0520501025,
-0.0434124023,
-0.0240254253,
0.0272744503,
-0.0285042003,
0.0054978505,
-0.0163121354,
-0.0011216888,
0.0057478631,
-0.0113451993,
-0.041040089,
0.0157844499,
-0.0274157245,
0.005341419,
0.0085377339,
-0.0283299629,
0.0393048935,
-0.0126235951,
-0.0413033403,
-0.0229268149,
0.0317531452,
0.0278257281,
0.0151863415,
-0.026527375,
0.0134475995,
-0.0208461937,
-0.003577143,
0.042333249,
-0.0166452564,
-0.0209809262,
0.0576870218,
-0.0526068769,
0.0036641534,
0.0008778863,
-0.0531504229,
-0.0213469937,
0.0317657366,
0.0567568876,
0.0942830071,
-0.028333202,
0.0352419913,
-0.0652817264,
0.0268870853,
-0.0242002215,
0.0537860282,
0.0404891744,
-0.0003292423,
-0.0296959113,
0.0129186111,
-0.0517723709,
-0.0023737319,
-0.0476581454,
-0.0306067374,
0.0328551456,
-0.0136256916,
-0.0256876443,
-0.0113770682,
0.059725225,
-0.0554169565,
-0.0007399348,
0.0672925562,
-0.0033409218,
0.0003665427,
0.0095035033,
0.0304938089,
-0.0288557373,
-0.0104428558,
0.0475235097,
0.0049460935,
-0.0144127924,
0.0083691906,
0.013325572,
-0.0168949235,
-0.0540784486,
0.0109535726,
0.0011287279,
0.0140691381,
-0.0708753094,
0.0550668016,
0.0052504735,
0.0120684654,
-0.0231711622,
0.0858050361,
-0.0231869351,
0.0065233437,
-0.0212763678,
-0.0454210863,
0.0147313047,
-0.0008429564,
-0.0582172796,
0.0158689544,
0.0002837765,
-0.0059347269,
-0.0063163145,
-0.0027410535,
-0.0070064492,
-0.0070636342,
2.231e-7,
0.0185032114,
0.0113567729,
-0.0046043657,
0.0346778445,
0.0020190815,
0.063321963,
0.0154507924,
0.0520340577,
0.0325759687,
-0.005964342,
-0.0101607833,
0.0251715034,
0.0565386526,
0.0046853041,
-0.0878702924,
0.0237162858,
-0.0775287002,
-0.0106343022,
-0.01232675,
-0.0040525724,
0.0733308271,
0.0829642341,
0.0106523177,
0.0151315024,
-0.013463879,
-0.0651349649,
0.049549263,
-0.0202172678,
-0.0271321125,
0.0070982375,
0.021956237,
-0.0155581804,
0.020666955,
0.0120074125,
0.0107798344,
0.0589261763,
0.0283287968,
0.0798289999,
0.0279376619,
-0.0756186396,
-0.0293699969,
-0.038228713,
0.0283453893,
-0.0576311946,
0.0446178205,
-0.075957343,
-0.0337454043,
-0.0135966521,
-0.0262297504,
-0.0259851795,
0.0429778062,
-0.0269744843,
-0.0054534404,
-0.0097707333,
0.0206187796,
-0.0104018552,
0.0171418265,
-0.0658748671,
-0.0192815438,
-0.0011947559,
-0.0506414771,
-0.0731608346,
-0.0073709846,
-0.0164998583,
0.0222721696,
-0.0496219061,
0.0288677886,
1.873201102e-34,
0.0480670966,
-0.0033824337,
-0.0076248953,
0.0129085835,
0.043854624,
0.0147576286,
-0.0181895345,
-0.0317576267,
0.0716023743,
-0.1052576229,
0.0186447669
] |
http://arxiv.org/abs/2406.00031v1
|
http://arxiv.org/pdf/2406.00031v1
|
AMGPT: a Large Language Model for Contextual Querying in Additive Manufacturing
| 2024-05-24T20:03:32 | 2024-05-24T20:03:32 |
Achuth Chandrasekhar, Jonathan Chan, Francis Ogoke, Olabode Ajenifujah, Amir Barati Farimani
|
cs.CL
|
cs.CL, cs.LG
|
Generalized large language models (LLMs) such as GPT-4 may not provide
specific answers to queries formulated by materials science researchers. These
models may produce a high-level outline but lack the capacity to return
detailed instructions on manufacturing and material properties of novel alloys.
Enhancing a smaller model with specialized domain knowledge may provide an
advantage over large language models which cannot be retrained quickly enough
to keep up with the rapid pace of research in metal additive manufacturing
(AM). We introduce "AMGPT," a specialized LLM text generator designed for metal
AM queries. The goal of AMGPT is to assist researchers and users in navigating
the extensive corpus of literature in AM. Instead of training from scratch, we
employ a pre-trained Llama2-7B model from Hugging Face in a Retrieval-Augmented
Generation (RAG) setup, utilizing it to dynamically incorporate information
from $\sim$50 AM papers and textbooks in PDF format. Mathpix is used to convert
these PDF documents into TeX format, facilitating their integration into the
RAG pipeline managed by LlamaIndex. Expert evaluations of this project
highlight that specific embeddings from the RAG setup accelerate response times
and maintain coherence in the generated text.
|
AMGPT: a Large Language Model for Contextual Querying in Additive Manufacturing: Generalized large language models (LLMs) such as GPT-4 may not provide
specific answers to queries formulated by materials science researchers. These
models may produce a high-level outline but lack the capacity to return
detailed instructions on manufacturing and material properties of novel alloys.
Enhancing a smaller model with specialized domain knowledge may provide an
advantage over large language models which cannot be retrained quickly enough
to keep up with the rapid pace of research in metal additive manufacturing
(AM). We introduce "AMGPT," a specialized LLM text generator designed for metal
AM queries. The goal of AMGPT is to assist researchers and users in navigating
the extensive corpus of literature in AM. Instead of training from scratch, we
employ a pre-trained Llama2-7B model from Hugging Face in a Retrieval-Augmented
Generation (RAG) setup, utilizing it to dynamically incorporate information
from $\sim$50 AM papers and textbooks in PDF format. Mathpix is used to convert
these PDF documents into TeX format, facilitating their integration into the
RAG pipeline managed by LlamaIndex. Expert evaluations of this project
highlight that specific embeddings from the RAG setup accelerate response times
and maintain coherence in the generated text.
|
[
-0.001232422,
0.0467419028,
-0.0231725331,
0.041018568,
0.0096413484,
0.017383663,
0.0151363062,
0.0108703906,
-0.0424702726,
-0.0237022452,
-0.0541317351,
0.0167209841,
-0.0181956422,
0.0373883396,
0.0539054237,
-0.0012183278,
0.0079799714,
-0.0124835828,
0.0209377538,
-0.0071524302,
-0.0013733227,
0.0082530146,
0.0070112981,
0.009249283,
-0.0166158564,
-0.006233112,
-0.0114801675,
-0.0294007156,
0.0298844334,
-0.0189024974,
0.007807537,
0.0282677561,
0.0419036299,
0.0376108214,
0.0000024122,
-0.0322635993,
0.0107802087,
0.053670235,
0.022010427,
0.0113440696,
0.0664705113,
-0.0406767167,
-0.0235239789,
0.0153519912,
-0.0326910205,
0.0140845459,
0.0182563644,
0.0490520336,
0.0183197632,
0.0209010243,
-0.0100801,
-0.0604891777,
0.060711626,
-0.0362562761,
0.0809044167,
-0.0344320014,
-0.0075917714,
0.0246608537,
0.0755591244,
0.0677767545,
0.0274884589,
0.083828181,
0.0181125049,
-0.0535671674,
0.0731844306,
0.0406747013,
-0.0297315307,
0.0168991219,
0.0121432981,
0.0739813522,
0.0549415872,
-0.0161667522,
-0.0131881097,
-0.0191069692,
-0.0339098759,
0.0551664494,
-0.086620748,
0.0253068693,
-0.0196924079,
-0.0105840191,
-0.0163813382,
-0.0008673383,
0.0235742629,
0.0028470808,
-0.0256348625,
0.0525431372,
-0.0132114049,
-0.0325726084,
-0.004947816,
0.0130005768,
-0.0196494963,
0.0002300453,
0.0178108271,
0.0272393636,
0.0078137871,
0.0184517074,
-0.0217306428,
-0.0574020073,
0.0579746068,
-0.0414552987,
0.0096785296,
0.0444615148,
-0.0299404003,
0.040171206,
-0.0736015141,
0.007267612,
-0.0098426621,
-0.011978168,
-0.0718871579,
-0.0130056432,
-0.0351481065,
-0.073906675,
-0.0552217849,
0.0207589529,
-0.0444449075,
-0.0389732495,
-0.0262289792,
0.0228890721,
0.0057317857,
0.0022063197,
0.0219211765,
-0.0054587414,
-0.0536775552,
-0.0306065921,
-0.0496950261,
-0.0253805816,
-0.0210292824,
0.0167455375,
-0.0067869569,
-0.022720838,
-0.0177137014,
0.0281011779,
-0.0006947431,
-0.0268076416,
-0.0074488381,
0.0820596442,
-0.0390590951,
-0.0242473185,
-0.0397647321,
-0.0386638679,
0.0187575296,
-0.0247392394,
0.0065369448,
-0.0497639216,
0.0512332954,
0.0043166075,
-0.0162748117,
-0.0060880035,
0.0347701013,
0.0267333351,
-0.000461716,
0.0573103055,
0.0003528082,
-0.0340389498,
0.054566402,
-0.0109468997,
0.0480247997,
-0.0012283402,
0.0282372143,
0.0473175831,
0.0128188059,
-0.0047392682,
-0.0235352218,
0.0045654089,
0.0255307909,
-0.0115387635,
-0.0237886943,
0.002404771,
-0.0032737881,
0.0383970514,
0.0248845425,
0.0170675088,
-0.0412745588,
0.0072156005,
0.0152619304,
0.0389443822,
0.0263845734,
0.0200981814,
0.0491318777,
0.0321252048,
0.0108904354,
-0.0342739262,
0.0046472186,
0.0383743085,
-0.0421053246,
0.0037857816,
0.0224862788,
-0.0019998737,
-0.0004375909,
-0.0482785702,
-0.0033679556,
0.0046738675,
0.0130797131,
-0.0073262663,
0.0514766797,
0.0203235727,
-0.0371266939,
-0.0290423017,
-0.0062192073,
-0.0344954766,
-0.0241052806,
-0.0148598803,
0.0307170898,
0.0611696392,
0.0583146662,
-0.0599315651,
-0.0419507511,
0.0290612895,
0.0016141876,
0.0728608444,
0.0010658338,
-0.0344084464,
-0.0074223885,
-0.0511160642,
-0.025267221,
-0.0382523946,
-0.0073361374,
0.0153609626,
-0.0055856705,
-0.0122373672,
0.0265386142,
-0.046601437,
-0.0379227065,
0.049742721,
-0.0675807744,
-0.079056561,
0.0174275003,
0.0075209881,
0.0034138428,
0.0251890179,
0.0201209541,
0.0304996725,
0.0918140709,
-0.0512136258,
-0.045479238,
0.0169086959,
-0.0159454197,
0.0083836569,
0.0322473198,
-0.0497300886,
0.041563727,
-0.080121398,
-0.0378450528,
-0.043718081,
0.0239211638,
-0.0195671991,
0.0333513394,
-0.0283675678,
-0.0094754137,
0.025583066,
0.0119845467,
-0.0787644237,
0.0787106082,
0.0235108957,
-0.012937787,
0.0246798806,
0.0446265973,
0.0272045881,
-0.0476367548,
0.0093550747,
0.0301154945,
0.0090446938,
0.0082199425,
0.03649785,
0.0057408558,
-0.0272765364,
-0.0414867215,
-0.0392099544,
0.0152343698,
0.0165528096,
0.0236071553,
0.0110653862,
0.0043855268,
-0.0192467719,
-0.0149608767,
0.0014818354,
-0.0161957946,
0.0061036157,
-0.0450439677,
0.0568604693,
0.0660866871,
0.0152233969,
-0.0177970957,
-0.0277386885,
-0.0109794494,
0.0067085056,
0.0584100038,
0.0066793626,
-0.0591073371,
0.0279921982,
-0.0161851887,
0.0055988203,
-0.0347083583,
-0.0078642815,
0.0294289179,
-0.0396640785,
0.0400215685,
0.0470224023,
0.0106694093,
0.0560915843,
0.0128185209,
-0.0336300693,
-0.0640382245,
0.0044724667,
-0.02186854,
0.0068292203,
0.0255931038,
-0.0195213538,
0.0398002714,
0.0197759625,
0.0000148787,
0.0105079208,
-0.0427685305,
0.0446560681,
0.008446021,
-0.0409222133,
-0.0438285097,
-0.0011497054,
0.0310381465,
0.0392136574,
0.0346195884,
-0.0306884218,
0.0634922832,
-0.0488658175,
0.0094325133,
-0.0376410969,
-0.020267874,
-0.0605645776,
0.0234779399,
0.0141572608,
0.0240326263,
-0.0058664414,
-0.0181584302,
-0.0293338001,
0.0583134033,
0.0313996747,
-0.0146669019,
-0.0087119993,
-0.0435529277,
0.0465095416,
-0.0183164682,
-0.04257847,
-0.0193287525,
-0.0038156779,
0.0065617729,
0.0173530579,
0.0105241807,
-0.0500962995,
-0.0581039488,
0.0312032308,
0.0793841854,
-0.0028709851,
-0.0187216997,
-0.0297754332,
-0.0414866246,
0.03775765,
0.0416012481,
-0.0122824032,
0.004709763,
-0.0197209083,
-0.0247742832,
0.0315556936,
0.0102136508,
-0.0305604059,
-0.0494422354,
-0.0025818478,
0.0203675237,
0.0012057999,
-0.0180802327,
-0.0297501944,
-0.0668805167,
-0.0645968691,
0.0132008391,
0.0169850625,
0.0447924063,
-0.0101439673,
-0.0126085794,
-0.0278236493,
0.0525491685,
-0.0098509295,
0.0503837131,
-0.1132258102,
0.0121551873,
-0.0040594395,
0.0702050179,
0.0002755444,
0.0041721133,
0.0052343467,
-0.1137004271,
0.0868239924,
-0.0146556608,
-0.0591153726,
-0.0630056337,
-0.0071628178,
-0.000536528,
0.0343294069,
-0.0218106732,
0.0438179635,
-0.0537019931,
0.0085724015,
0.0341630876,
0.0513780192,
-0.100814864,
-0.0086435564,
0.0478532091,
-0.0376720503,
0.042069532,
0.0427575149,
-0.0083831772,
-0.0537349507,
0.0570859239,
-0.0055368096,
-0.0012843289,
-0.0577652305,
0.0068403026,
0.0072771269,
0.0044497829,
-0.0162163749,
0.0516125076,
0.0071763447,
-0.0370664299,
0.0034104944,
0.0600853264,
-0.0317490548,
0.0578399785,
-0.0555478483,
-0.0421070606,
-0.0256130639,
0.0312075466,
-0.00253442,
-0.0066624018,
-0.0618042275,
-0.0405469723,
0.0413424708,
-0.0343537815,
-0.0241440535,
0.0092538623,
-0.036875207,
0.0090207262,
-0.0564171001,
0.0100142313,
0.0168374199,
0.0371257737,
-0.0019854445,
-0.0002858712,
-0.001904511,
-0.0309200231,
-0.0644730255,
0.0150534213,
-0.0566881336,
0.0294224434,
-0.0489810854,
-0.0020536939,
0.0263501834,
-0.0809890106,
-0.0449741036,
0.0251061935,
-0.0180742107,
-0.0584752336,
-0.044423189,
0.0278006364,
0.0248272214,
0.0238657407,
-0.0375840776,
-0.0649345517,
0.0731115043,
0.0349982418,
-0.0324612372,
-0.0035053785,
-0.0144301625,
-0.0168444403,
-0.0193367917,
-0.0267292485,
-0.0352796912,
0.0322102644,
-0.0408065952,
-0.0007935517,
0.0103687095,
-0.0437340103,
0.0207307991,
0.0505325384,
-0.0320715606,
-0.0080259237,
0.0414434262,
-0.0569758639,
0.0526847802,
0.0061496939,
0.0023783962,
-0.031632185,
0.0237943958,
0.0055323406,
-0.0801825076,
0.0119070485,
-0.0337221175,
0.0062430203,
0.1131256595,
-0.0479483306,
0.0595552213,
-0.0343719237,
0.04285134,
0.0627891198,
0.0005956991,
-0.0034231145,
-0.0004626116,
-0.0207607485,
0.1047806069,
-0.0258756895,
0.0196402986,
0.0339637063,
0.0132040186,
-0.015575571,
0.0342947282,
-0.060350325,
-0.0592516325,
-0.0603901558,
-0.0031773909,
-0.0115507673,
0.0041284785,
0.0388814546,
-0.0366813876,
-0.0250808168,
-0.0084592337,
0.051097393,
0.0149912043,
-0.0129711144,
0.0147810811,
0.0278011058,
0.0122311935,
-0.0505342185,
0.0664249286,
0.0010265353,
-0.0241333004,
-0.0249611884,
-0.0197993722,
-0.0669028759,
-0.0279135983,
0.0579988994,
0.0367493257,
0.0092760976,
-0.027501164,
-0.0460154861,
-0.0127559546,
-0.0195494648,
0.0292285904,
-0.0430450961,
-0.0306040253,
0.0305430666,
0.1140458211,
0.0090427361,
-0.0033379737,
-0.0198698863,
0.039085716,
-0.052391544,
-0.0320598707,
0.0037536817,
-6.717743034e-33,
-0.0644465834,
-0.0223529842,
0.0381634906,
-0.0010350088,
-0.037956886,
-0.0155191245,
-0.0420174934,
-0.0439900346,
0.0224384721,
0.0046610716,
-0.0135419462,
-0.0104695503,
0.0065310034,
-0.0042140526,
0.009037016,
0.0286146514,
0.0320205837,
0.0215513594,
-0.0142087266,
0.0615272708,
0.0065828064,
0.0264185499,
0.0249931756,
-0.0676618889,
0.0002108813,
0.0088740243,
-0.0058783055,
-0.0191745572,
-0.0202121306,
0.0703864023,
-0.0573963672,
0.0321634822,
0.0040233484,
-0.0084241824,
-0.0161905717,
0.0263378657,
-0.0774417371,
-0.0312979594,
-0.0165329352,
-0.0242772344,
0.0027301519,
-0.0214706995,
0.0338414684,
-0.0304586813,
-0.0245502293,
0.015540787,
0.0578132756,
-0.039841257,
-0.0179898627,
-0.0089661013,
-0.1077484787,
0.0198564604,
0.0173256099,
0.0521265306,
0.0433618017,
0.044215247,
0.01791583,
0.0913995132,
-0.0650022253,
0.0297072008,
0.0350595638,
0.0274712816,
0.0364831276,
0.017146403,
0.0370740592,
-0.0029511496,
-0.0091982009,
-0.0013207715,
0.020558279,
0.0136671793,
0.0217548646,
0.0634200573,
0.04075858,
0.0388376005,
0.025227664,
-0.0011353253,
-0.0705772191,
0.04166409,
-0.0252345372,
0.0601225048,
0.0360836722,
0.0311432406,
0.0191338062,
-0.0189638194,
-0.0204715431,
0.0336670727,
-0.0166643038,
-0.0104443701,
0.0156019032,
0.0412597209,
-0.04868095,
-0.0205599312,
0.0111332769,
0.0039105243,
-0.0130383531,
0.0248531718,
-0.0390672497,
0.0262553003,
-0.0174083281,
-0.0046232659,
-0.0827872902,
-0.0179901998,
-0.0177438576,
-0.001991241,
0.0207817983,
-0.006473721,
-0.004127684,
0.0244417451,
-0.0432649516,
0.0331476629,
-0.0403899364,
0.0140391458,
-0.0092626913,
0.0311633758,
-0.0060321563,
0.023347849,
0.0087933643,
-0.0173774101,
0.0047271824,
0.0111576496,
-0.0118379155,
-0.027260894,
-0.022550419,
0.0121298032,
-0.0194825772,
0.0451488979,
-0.0048665819,
0.0541368909,
0.0854683667,
-0.0724334866,
0.0044448101,
0.0551629402,
3.168e-7,
0.0599496029,
0.0440611616,
0.0288389921,
0.0364332944,
-0.0006494183,
0.0318423696,
-0.008834675,
0.0233231727,
0.0343964174,
0.0267942082,
0.018747054,
0.0209328178,
0.0383242927,
0.0253690444,
-0.0770912319,
0.0065562101,
-0.0318112224,
-0.022725245,
-0.0527110212,
0.0205441546,
0.0408052914,
0.0476682745,
-0.0390119106,
-0.0420551784,
0.0063318773,
-0.0459638126,
0.0743405744,
-0.0023660953,
0.003305323,
0.0193082038,
0.0155139342,
0.0152211376,
-0.0206223223,
-0.0032989511,
-0.0287652034,
0.0146749616,
0.0050154687,
0.019470742,
0.0102910167,
0.0250732321,
0.0284087807,
-0.0096847042,
0.0279104449,
-0.0357506461,
0.021249406,
-0.0300116893,
0.0102692582,
0.0078675468,
-0.0200994704,
-0.0170683917,
-0.0013379257,
-0.0455477275,
-0.0025269475,
-0.0031839225,
0.0156992227,
0.0028088833,
-0.000815493,
-0.073147811,
-0.0200173836,
0.0107483091,
-0.0631241202,
-0.0324345231,
-0.0474792384,
0.0043880558,
-0.0036065497,
-0.0125635974,
-0.0117130885,
3.248474047e-34,
0.0472602658,
0.0100507345,
0.0031311566,
0.0599125214,
0.0287115406,
0.0066342549,
-0.0463877954,
0.0682574287,
0.0251186825,
-0.1097483262,
-0.0141577646
] |
http://arxiv.org/abs/2405.15902v1
|
http://arxiv.org/pdf/2405.15902v1
|
Hacc-Man: An Arcade Game for Jailbreaking LLMs
| 2024-05-24T19:55:20 | 2024-05-24T19:55:20 |
Matheus Valentim, Jeanette Falk, Nanna Inie
|
cs.CR
|
cs.CR, cs.AI, cs.CL, cs.HC
|
The recent leaps in complexity and fluency of Large Language Models (LLMs)
mean that, for the first time in human history, people can interact with
computers using natural language alone. This creates monumental possibilities
of automation and accessibility of computing, but also raises severe security
and safety threats: When everyone can interact with LLMs, everyone can
potentially break into the systems running LLMs. All it takes is creative use
of language. This paper presents Hacc-Man, a game which challenges its players
to "jailbreak" an LLM: subvert the LLM to output something that it is not
intended to. Jailbreaking is at the intersection between creative problem
solving and LLM security. The purpose of the game is threefold: 1. To heighten
awareness of the risks of deploying fragile LLMs in everyday systems, 2. To
heighten people's self-efficacy in interacting with LLMs, and 3. To discover
the creative problem solving strategies, people deploy in this novel context.
|
Hacc-Man: An Arcade Game for Jailbreaking LLMs: The recent leaps in complexity and fluency of Large Language Models (LLMs)
mean that, for the first time in human history, people can interact with
computers using natural language alone. This creates monumental possibilities
of automation and accessibility of computing, but also raises severe security
and safety threats: When everyone can interact with LLMs, everyone can
potentially break into the systems running LLMs. All it takes is creative use
of language. This paper presents Hacc-Man, a game which challenges its players
to "jailbreak" an LLM: subvert the LLM to output something that it is not
intended to. Jailbreaking is at the intersection between creative problem
solving and LLM security. The purpose of the game is threefold: 1. To heighten
awareness of the risks of deploying fragile LLMs in everyday systems, 2. To
heighten people's self-efficacy in interacting with LLMs, and 3. To discover
the creative problem solving strategies, people deploy in this novel context.
|
[
0.0628729835,
0.0284229126,
-0.0202601179,
0.0550147966,
-0.0572285876,
-0.0345532708,
-0.0086165909,
-0.018822154,
0.036927294,
0.0101318313,
0.011022415,
0.0253533386,
0.0127749043,
0.0393070057,
0.0311930813,
0.0201142207,
0.0326000713,
-0.0641875267,
-0.0246199444,
0.0288792998,
-0.0100816507,
0.007589546,
-0.020465441,
0.0492215082,
0.0271298066,
-0.0310407337,
-0.0245314501,
0.0384662747,
0.016876841,
-0.0416692682,
-0.0102549121,
0.0058322465,
0.0096298968,
0.0310537778,
0.0000022683,
-0.0603381135,
-0.0274582412,
0.0198567752,
-0.0362574011,
0.0282426998,
0.0016424245,
0.0278857723,
-0.0482194982,
0.0065297582,
-0.0474218614,
0.0409360602,
0.0285429098,
0.020291714,
0.0895810053,
0.0380187482,
0.0201522876,
-0.0110309888,
0.0151398964,
0.0004638546,
0.0159576535,
-0.0639389902,
0.0106580723,
0.001174918,
0.0834946036,
0.0130372876,
0.0084716361,
-0.0131501723,
0.0149037074,
-0.0125775877,
-0.0396958217,
0.0060575637,
-0.0030966827,
-0.0076286737,
-0.0051529114,
0.0060550217,
-0.025241768,
0.0000110432,
0.0182682984,
0.0354344212,
0.0252888203,
0.0305995606,
-0.0243444927,
0.0588820055,
-0.0034863381,
-0.0141715137,
0.0004099017,
-0.0186510701,
0.0229125842,
-0.0162853077,
0.060120713,
0.0589857437,
-0.0182278436,
-0.0262649059,
-0.0238870624,
-0.0194059722,
0.0501461439,
-0.043012701,
-0.012035151,
0.0418951735,
-0.0011831395,
0.0181333926,
-0.037321046,
0.0064312825,
0.0260549709,
-0.0789770782,
-0.0064175664,
0.0135835279,
0.0299175251,
0.0416515358,
0.0169648379,
-0.0067703761,
-0.0304685216,
0.0048352689,
-0.0009701857,
0.0379046537,
-0.0295012984,
-0.0597691499,
-0.0293584354,
0.0324170776,
0.000767081,
0.0045596096,
-0.014408472,
0.0559366345,
-0.000096023,
0.0489829928,
-0.0488925949,
0.0055127842,
0.0164274368,
-0.0158339217,
0.0132297538,
-0.0626279339,
-0.0144419679,
0.0084861573,
-0.0117763653,
-0.0607235506,
-0.0058497237,
-0.0187648274,
-0.0579422936,
-0.0003748365,
0.0132980356,
0.132684201,
0.0117725581,
-0.001510203,
-0.0867902264,
-0.0691084862,
-0.0002715574,
-0.0282823127,
0.0076085157,
-0.0944432169,
-0.0185657144,
-0.0188005995,
0.0252374355,
-0.0315280072,
0.0235180277,
0.0025902553,
-0.0166873187,
0.0758847594,
-0.0306344461,
-0.0104644205,
0.0424097367,
0.0612554885,
0.0396652371,
0.0103524597,
0.0241253134,
0.0538830943,
0.0332602262,
0.0058106766,
0.1059499905,
0.0452704765,
-0.0638346523,
0.0109892823,
-0.0024226808,
-0.0039794222,
0.0381721854,
0.015260498,
-0.0556228422,
0.0156634133,
0.0040387153,
0.0398370996,
0.0688475966,
0.0687408,
-0.0273663756,
0.0185503997,
-0.0268790722,
0.0086137354,
-0.0003344935,
-0.0492994748,
0.0120828729,
0.0415284969,
-0.0759481043,
-0.0063332133,
-0.0577175468,
0.0328038298,
0.0083412407,
0.0013487098,
-0.018744994,
-0.0287182368,
-0.011158308,
-0.0047048857,
0.0036994657,
0.0223080218,
0.0164063759,
0.0353390463,
0.0122112371,
-0.0558966175,
-0.0320090353,
-0.0934588984,
0.0205991603,
0.0952266231,
0.0368854962,
0.0171364918,
0.0036359983,
-0.0218854863,
-0.0001669873,
0.0302695744,
-0.0297813937,
0.0302153919,
-0.0153334821,
0.0150970099,
0.0072005182,
-0.0188186616,
-0.0082435859,
0.0385793261,
-0.0430380851,
0.0001828538,
-0.0155749526,
0.0211555026,
-0.0113271037,
0.0130148437,
-0.0040548919,
-0.0481036715,
-0.0227868687,
0.0208823923,
-0.012491595,
0.0234786961,
-0.0453905165,
0.026082404,
0.0675314814,
-0.0152941206,
-0.0024156033,
-0.0063968124,
-0.0061330851,
0.0238346085,
0.0454594642,
-0.1219809055,
0.0171616748,
0.0310161132,
-0.0137137407,
0.0178832915,
0.0922789574,
-0.0661224052,
0.0113750789,
0.0048748134,
0.0417217053,
-0.0150170969,
-0.0192879979,
-0.0278769135,
-0.0026263609,
-0.0296766479,
-0.0216996111,
-0.0181635022,
-0.0310187191,
0.0258136634,
-0.0300826821,
-0.0261537861,
0.0347911976,
-0.0230361838,
0.0244728979,
-0.0046923053,
0.0556020588,
-0.002925948,
-0.0354840197,
-0.0317879096,
-0.0113549521,
-0.0010759671,
0.0241186917,
0.0382698961,
0.0052509168,
-0.0390307456,
-0.0275250431,
0.0104791271,
-0.02617516,
-0.0025918554,
-0.0206634738,
0.0160344969,
-0.0037600773,
-0.0672532916,
0.0165004935,
-0.0174324624,
0.0006997787,
0.0457715355,
-0.0351302586,
-0.0154430876,
-0.0491837114,
-0.0085267946,
-0.0075247502,
-0.0226214752,
-0.0030868794,
-0.0396916009,
0.0463063791,
-0.0159269106,
-0.0202125963,
0.0027269488,
0.0048075905,
0.0481018722,
0.0646843612,
-0.0285594109,
-0.0074374378,
-0.0107282344,
0.0727329031,
-0.0524651818,
-0.0148090133,
-0.0771678835,
0.0464282036,
0.0469220392,
0.0147994356,
0.0393170901,
-0.0029285846,
0.0539686941,
-0.007917692,
-0.0109318737,
-0.022594668,
-0.0606345944,
-0.0633681193,
-0.0093559008,
0.0013939304,
-0.0438225791,
0.0734280348,
-0.0579318479,
-0.0060328916,
0.0195892137,
-0.0555166341,
-0.0321941003,
0.088713631,
-0.0442001745,
0.0141958771,
0.0356316343,
-0.0411012843,
-0.0171354115,
-0.0473656133,
0.021714611,
-0.0593288615,
0.0013420722,
-0.0738681555,
-0.0182330478,
0.0055757728,
-0.0233097859,
0.0026175766,
-0.0176600292,
-0.0623925328,
-0.0283620171,
0.0418420322,
-0.0140813487,
-0.0230444185,
-0.0472381823,
0.0843583718,
-0.0189845189,
0.0396535322,
-0.0305001158,
-0.006734421,
0.0048509794,
0.0038862156,
-0.0192232709,
-0.0225907266,
-0.0010274122,
0.0049658334,
0.0715558976,
0.0194407739,
-0.0108279306,
-0.000033658,
0.0274010934,
0.0322035216,
0.017313974,
0.0033693423,
0.0372982398,
0.020454051,
-0.0195375811,
-0.0125156092,
0.0181708112,
0.0157931931,
0.0446435064,
0.0365837477,
-0.031599585,
-0.0237384569,
0.0287010204,
0.0476814508,
-0.0977352411,
-0.0187946837,
0.0173872281,
0.0610801019,
-0.0364516675,
-0.0237302966,
-0.0244022924,
-0.0679291859,
0.0114545822,
0.0414134711,
-0.0152124558,
0.0003866535,
-0.0201329272,
-0.0412814803,
0.0122587811,
0.0085656904,
0.0088936659,
0.0014575218,
0.002203383,
0.0203466248,
-0.0493655019,
0.0142412689,
0.0246307086,
0.0134942792,
0.0894173533,
0.0919487625,
-0.0146272574,
-0.0161657967,
-0.028325092,
0.0334285647,
0.0372063927,
0.0393759906,
-0.0156739876,
-0.0323408805,
0.0005554497,
-0.0064069554,
-0.028107645,
0.0060862792,
-0.0244587231,
-0.0616237558,
0.0392025486,
0.0593186058,
-0.0055566095,
-0.013817329,
-0.0556056015,
-0.0420934409,
0.0356998295,
0.0211862251,
0.0102081485,
-0.0315783955,
-0.0177504513,
-0.0734276101,
-0.0616568029,
-0.003129259,
-0.0801484957,
-0.0254721791,
0.0508984961,
0.0126865059,
-0.0118949683,
-0.0148889283,
-0.0100845592,
0.0285574589,
0.01672641,
-0.0098712984,
-0.0158514176,
0.0556660295,
-0.0264212433,
-0.040564537,
-0.042326685,
0.0484512784,
-0.0550478771,
0.036750216,
-0.0314876325,
-0.0476470329,
0.0006277664,
0.0493338853,
0.0041689659,
-0.0684997886,
0.0381696746,
0.0659229904,
0.0360166579,
-0.0824328661,
0.0028504904,
-0.0698239133,
0.0320554413,
0.0722476393,
-0.0582169257,
0.0476538278,
0.0206243545,
0.0208586138,
0.0282328185,
-0.0150653478,
-0.0268038586,
-0.0514683574,
0.0741590783,
0.035897173,
-0.0698005706,
-0.0316400863,
0.0176729504,
0.0548084378,
-0.0293834154,
-0.0382352993,
0.014347733,
0.0430927128,
0.0529974326,
-0.0207713693,
-0.0396499075,
-0.0094289742,
0.0046134987,
0.0084543284,
-0.041318357,
0.0603166334,
0.0100135226,
-0.0200539324,
0.0116823865,
0.0181727018,
0.0511695519,
0.0377385095,
0.0428985283,
0.0480281934,
0.0272122845,
-0.0047071292,
0.0835596472,
-0.0274246372,
0.0386200324,
0.0248652436,
0.0325398929,
-0.0170890838,
-0.0209849663,
-0.020333454,
0.0513715148,
-0.0453323238,
-0.0668384656,
0.0416115224,
0.1023047194,
0.0303923283,
0.049835708,
-0.025024347,
-0.0155997761,
0.0291389171,
-0.036007721,
-0.0224274304,
-0.0394872501,
-0.0554098561,
0.0052651051,
0.0231446512,
-0.015642209,
-0.0246731397,
0.0453854948,
-0.0049611656,
0.0228311345,
-0.0042232391,
0.0115078026,
-0.0610103905,
0.0060799671,
0.0545289554,
0.0165520366,
-0.0444014892,
-0.0237860028,
-0.084279649,
-0.0292399358,
-0.0280547701,
0.0337846056,
0.0413626134,
-0.0137511007,
0.0123278731,
0.0297216065,
0.0326467082,
-0.0268160123,
-0.0514119826,
0.0298906304,
-0.0037561518,
-0.0383357219,
0.0084233508,
-6.263354958e-33,
0.0140946181,
-0.0831210464,
-0.0169421155,
-0.0139206275,
0.0125379907,
-0.0524850674,
-0.0318061709,
0.0766501948,
0.0152198002,
-0.0328187607,
-0.0002426809,
-0.0168778803,
0.0195010472,
-0.0148599409,
-0.0109502943,
0.0072855288,
0.0077573555,
0.0362983271,
-0.0069819544,
-0.0075289691,
0.0363272466,
0.0296398513,
0.0294836722,
-0.0554640256,
0.0672535375,
-0.0380836017,
-0.0318281651,
0.0314044915,
0.0010759614,
0.0114650112,
0.0198672973,
0.00474294,
0.0064563365,
-0.0169373509,
-0.0033232754,
0.0406846479,
0.0380235836,
-0.0188355837,
0.018909432,
0.0121631492,
0.0266615637,
-0.07850077,
0.0123047251,
0.0036106799,
-0.0334554315,
0.0213517211,
0.0124897566,
-0.0093097128,
-0.0116095413,
-0.0063233301,
-0.0151084671,
-0.0127438204,
-0.0454394631,
-0.0019208976,
-0.0255375057,
-0.0409122072,
-0.0300440863,
-0.0668776855,
-0.0484338738,
0.0807146505,
0.0479549393,
-0.0172828026,
0.0649960488,
-0.0065931622,
-0.0093414644,
0.0380700231,
-0.0375730507,
-0.0260055959,
-0.0543228984,
0.0024075867,
-0.0145548936,
0.0632736906,
0.0480336174,
0.0619686283,
0.0086235348,
-0.038255699,
-0.0400904417,
0.0021254383,
0.0126098897,
0.0265285745,
0.0482164398,
0.0200585034,
-0.0669152886,
0.000421144,
0.0177474376,
0.0259142276,
-0.0206792261,
-0.0428464524,
-0.0110524427,
-0.0270934887,
-0.0483455583,
-0.0639686286,
-0.0126610417,
-0.0172301326,
-0.0099825487,
-0.0208858196,
0.0250619408,
-0.0206359923,
-0.0258311797,
0.0115255117,
-0.0165274739,
0.0057885018,
-0.0135532347,
-0.0154273799,
0.0514266826,
0.0052627251,
-0.0234227777,
0.0447194092,
-0.0476298258,
-0.0030454316,
-0.0023998797,
0.0302732624,
0.0293552298,
-0.0126599437,
0.0227716435,
-0.0429951288,
0.0061897547,
-0.0384658724,
0.0080108559,
-0.0596992932,
-0.0005207049,
0.0597057827,
-0.0580744594,
0.0262610614,
-0.0089553054,
0.0130426819,
-0.0481420644,
0.0580508932,
0.0631897897,
-0.0322960168,
-0.0098458258,
-0.0087592257,
3.129e-7,
-0.0521328077,
-0.0081925336,
0.0045663822,
0.0050765765,
-0.0091470266,
0.0672724396,
-0.0254075974,
0.0592806265,
-0.0033610489,
-0.0380562656,
-0.0071990723,
0.0113370484,
0.058045283,
0.0134174824,
-0.0426681824,
0.0251948237,
-0.036477875,
0.0331260934,
-0.0504007302,
0.0310923457,
0.0928012207,
0.0127157997,
0.0104762875,
-0.0381933041,
-0.016573038,
-0.0735442713,
-0.0287337098,
-0.0290194359,
0.0132618742,
-0.0239745695,
-0.0049252626,
0.0617553666,
-0.0251246393,
-0.0030914827,
-0.0033557215,
-0.051344499,
-0.0417689718,
0.0363540985,
0.0073050307,
-0.0109620914,
-0.0121939015,
-0.0810287818,
-0.0384255387,
-0.016329648,
0.0412758291,
0.0142593607,
-0.0314601846,
0.0068781362,
0.0295903925,
-0.0370752513,
-0.0121121546,
-0.0532214791,
0.0200890321,
-0.0206980892,
0.0115978383,
-0.0045914552,
0.0000577474,
-0.0216667615,
-0.0130088339,
0.0214148238,
-0.0152072124,
-0.0514806658,
-0.004946603,
0.0569261946,
0.0137101812,
-0.0103979353,
0.0208584871,
2.612642953e-34,
-0.035695333,
-0.0159507543,
0.0082255052,
0.0094868634,
0.0186158232,
0.0510008298,
0.0144137433,
0.0188174602,
0.0053919228,
-0.0200786274,
-0.0287022069
] |
http://arxiv.org/abs/2405.15896v1
|
http://arxiv.org/pdf/2405.15896v1
|
Enhancing Augmentative and Alternative Communication with Card Prediction and Colourful Semantics
| 2024-05-24T19:33:34 | 2024-05-24T19:33:34 |
Jayr Pereira, Francisco Rodrigues, Jaylton Pereira, Cleber Zanchettin, Robson Fidalgo
|
cs.CL
|
cs.CL
|
This paper presents an approach to enhancing Augmentative and Alternative
Communication (AAC) systems by integrating Colourful Semantics (CS) with
transformer-based language models specifically tailored for Brazilian
Portuguese. We introduce an adapted BERT model, BERTptCS, which incorporates
the CS framework for improved prediction of communication cards. The primary
aim is to enhance the accuracy and contextual relevance of communication card
predictions, which are essential in AAC systems for individuals with complex
communication needs (CCN). We compared BERTptCS with a baseline model,
BERTptAAC, which lacks CS integration. Our results demonstrate that BERTptCS
significantly outperforms BERTptAAC in various metrics, including top-k
accuracy, Mean Reciprocal Rank (MRR), and Entropy@K. Integrating CS into the
language model improves prediction accuracy and offers a more intuitive and
contextual understanding of user inputs, facilitating more effective
communication.
|
Enhancing Augmentative and Alternative Communication with Card Prediction and Colourful Semantics: This paper presents an approach to enhancing Augmentative and Alternative
Communication (AAC) systems by integrating Colourful Semantics (CS) with
transformer-based language models specifically tailored for Brazilian
Portuguese. We introduce an adapted BERT model, BERTptCS, which incorporates
the CS framework for improved prediction of communication cards. The primary
aim is to enhance the accuracy and contextual relevance of communication card
predictions, which are essential in AAC systems for individuals with complex
communication needs (CCN). We compared BERTptCS with a baseline model,
BERTptAAC, which lacks CS integration. Our results demonstrate that BERTptCS
significantly outperforms BERTptAAC in various metrics, including top-k
accuracy, Mean Reciprocal Rank (MRR), and Entropy@K. Integrating CS into the
language model improves prediction accuracy and offers a more intuitive and
contextual understanding of user inputs, facilitating more effective
communication.
|
[
0.048071593,
0.0018197354,
-0.0449960679,
0.0275335666,
0.0228779241,
-0.0155828418,
-0.0270540006,
0.0354200453,
-0.0162400585,
-0.0223996565,
-0.038220685,
-0.0306890439,
0.0410337299,
0.031216424,
0.0145095447,
-0.0819219723,
0.0222701132,
-0.0164471455,
-0.0198663026,
0.0107643092,
0.0240538698,
0.0155801484,
0.0228065476,
0.0563916788,
0.0027398202,
0.0105958488,
-0.0124047138,
0.0111627309,
-0.0013576146,
0.0101506086,
0.0347327217,
0.0745523423,
-0.0155257965,
0.0376865268,
0.0000019934,
-0.0522265248,
0.0187710021,
-0.0230324287,
0.0109161716,
-0.0181358159,
0.0390684269,
0.0082911365,
-0.0076996107,
0.0456320755,
-0.0409857929,
-0.0113595715,
0.0723238811,
0.0350810289,
0.0380838253,
0.0055812271,
-0.0170794874,
-0.0001639793,
0.0364082754,
-0.0170255657,
0.0445699021,
-0.02184885,
0.0122181615,
-0.0342582762,
0.0223235767,
-0.0344145633,
0.0148092136,
0.0767480507,
0.0186488014,
-0.0390275046,
0.0218408126,
0.0018299411,
-0.0132976985,
-0.0171462912,
-0.027577091,
0.0317689627,
0.0055767987,
-0.0174378138,
-0.012610645,
0.0280728489,
0.0144646224,
-0.0038698921,
-0.0673797131,
-0.017577596,
0.0229085833,
-0.0132063134,
0.0049082339,
0.0405439809,
-0.0106862132,
0.0292589702,
-0.0205756016,
0.0355929174,
0.0241708439,
-0.0518429503,
-0.0282150321,
-0.0109441038,
0.0370038673,
-0.0491873287,
0.0606557392,
0.0144796912,
-0.0120589966,
0.0279679596,
0.0001841738,
-0.0827188715,
0.0116024734,
-0.086517036,
0.030326793,
0.0573295951,
-0.0172752794,
-0.0065113348,
-0.0041463152,
0.0155023728,
-0.0036790033,
0.0229798723,
-0.0227477569,
-0.0186105557,
-0.0219619069,
-0.0609651022,
-0.0657003224,
0.0330748223,
-0.072902821,
-0.0162000451,
-0.057934206,
0.0436001718,
0.0160239227,
-0.0146119967,
0.0403407924,
0.0086664874,
-0.0412646756,
-0.0090683093,
-0.0215796549,
0.0369514078,
-0.0173583385,
-0.0144944079,
0.0085430006,
-0.047946725,
0.0552198924,
0.0004042775,
-0.0225476399,
-0.0325642824,
-0.029604584,
0.0607035235,
-0.0283579621,
-0.0275359564,
0.0194706954,
0.0109614683,
-0.0171848405,
-0.0101947328,
0.0012575443,
-0.0415194407,
-0.0099387169,
0.0381230116,
0.0141093144,
-0.0226012655,
0.0086988695,
0.0225570854,
0.0037506735,
0.0541876517,
-0.0473569594,
0.0146573856,
0.0543748215,
0.0319368392,
0.0363747217,
0.0684923977,
0.0413402133,
0.0007316218,
0.0377498716,
-0.0450033769,
0.0039435951,
0.0765710175,
-0.0215779115,
-0.0160135813,
-0.0213105362,
-0.0332665741,
-0.0054024961,
0.008456816,
-0.0124548115,
0.0704340041,
-0.0215660911,
-0.0009477351,
0.0739767626,
0.01282633,
0.040320538,
0.0416166708,
0.0147061814,
0.0158045366,
-0.0096033625,
-0.0208123177,
-0.0031423829,
0.0305033531,
-0.0547266528,
0.0068609985,
0.0288046934,
0.030559184,
-0.0181661025,
-0.0444641598,
0.0061425408,
-0.0442227125,
0.0564300194,
-0.039177876,
0.044480931,
0.0168833118,
-0.0282397792,
0.0489489548,
0.0233965721,
-0.0629715398,
-0.0245462488,
-0.0562284961,
-0.0329631902,
0.1172467619,
0.0273957178,
0.0083539113,
-0.0069310092,
-0.0124314539,
-0.0559238382,
0.0604838915,
-0.0140890842,
0.0034900317,
-0.003435628,
-0.0741549879,
0.00539289,
0.0488768891,
-0.0289276373,
0.0357831195,
-0.0181504562,
0.0249457415,
0.02886994,
0.0091349389,
0.0092879832,
0.0050756633,
-0.0036600477,
-0.0068061333,
-0.0028542443,
-0.0018567623,
-0.000106896,
0.0442903191,
-0.0008115485,
0.0143026775,
-0.0029443027,
-0.0298973322,
0.0015743967,
0.0447986722,
-0.0671371073,
-0.0034939421,
0.0234069042,
-0.0222908985,
-0.0078858919,
0.0533818267,
0.0399811268,
-0.0308810771,
0.0424689651,
-0.0545939617,
0.1063971892,
-0.041379787,
0.0568317324,
-0.0128803812,
0.0277848858,
0.0387525111,
0.0052242763,
-0.0295967069,
-0.0123001831,
0.0013977601,
-0.0073172534,
0.009014627,
-0.0771017298,
-0.0281473994,
0.0213826541,
-0.0194401201,
-0.0005054274,
0.0240942743,
0.0493975356,
-0.0152376024,
-0.0351585075,
-0.1120581031,
-0.0524314828,
-0.0109469937,
-0.0162114631,
0.0135266297,
0.0438913368,
-0.0142872138,
-0.015417872,
-0.0153303165,
-0.0082323672,
-0.0060003777,
0.0050364719,
-0.0076623769,
0.0248705298,
-0.0455012359,
-0.0282857269,
-0.0296580475,
-0.0001091779,
0.0536881126,
0.0050752764,
-0.1163860038,
-0.054516457,
-0.0489753708,
-0.0302999355,
0.0123072797,
-0.0015214392,
-0.0623431653,
0.0094657606,
-0.0087130181,
0.0601018034,
-0.0088699255,
0.0438115261,
0.0050495896,
-0.0137320664,
-0.0130632576,
-0.0653936118,
-0.0068326932,
-0.0340793543,
-0.0094480971,
-0.0244496576,
-0.0710751414,
0.0440282971,
0.1011746749,
-0.0159846917,
0.0428618826,
-0.05784918,
-0.0464409627,
0.0100615798,
-0.0541377589,
-0.045075126,
-0.0299138166,
-0.057035163,
0.0220651943,
0.0159143377,
-0.038932167,
0.0787032619,
-0.0319073536,
-0.0499473214,
-0.0368315354,
-0.0107251266,
0.0134399254,
0.0206135064,
-0.0269049332,
0.0147715211,
-0.0141354473,
-0.04140535,
-0.0071580047,
-0.0225883219,
-0.0279476158,
-0.0419312045,
0.0179426577,
-0.0209080391,
-0.0241099242,
0.0356274098,
-0.0571099631,
0.0005861977,
0.0131043801,
-0.0382592231,
0.014252021,
0.054443676,
-0.016770022,
-0.023747446,
0.0462629199,
0.0009425797,
0.047630012,
-0.0106002558,
-0.0189099927,
-0.0087003196,
-0.0153462095,
0.0360319093,
-0.0116195632,
-0.0099698715,
-0.0429144315,
0.0055088941,
-0.0074520605,
-0.0037773431,
-0.0313487984,
0.0051241526,
-0.0200124551,
0.0590464063,
0.0330740511,
-0.0574513897,
0.022541739,
0.0090489136,
-0.0120692961,
0.0069223465,
0.0919416845,
0.0332066193,
-0.0262151752,
-0.0045603462,
-0.0658668801,
0.036741294,
-0.0085206693,
0.062201377,
-0.0836254731,
-0.0064008068,
-0.0088652801,
0.0296370815,
-0.0059095561,
0.013332583,
0.0052108113,
-0.0940305665,
-0.001622856,
-0.007481528,
-0.0337494574,
-0.0231345091,
0.0117034931,
-0.0190490261,
0.047300905,
0.0307000317,
-0.0100475578,
0.0102966884,
-0.0411422178,
0.0267788824,
-0.0948121771,
-0.0316983685,
-0.0073151579,
0.0272960532,
0.0359345451,
0.0911459103,
0.0243880674,
-0.0189629104,
-0.0843677968,
0.0187074058,
0.0032879722,
0.0212024301,
-0.0342627019,
0.0128553389,
-0.0183834899,
-0.0041478034,
-0.0552702509,
-0.0057207979,
0.0354190804,
-0.0133289853,
0.0014524609,
0.040685676,
-0.0087715909,
-0.0113491882,
-0.0198401213,
0.0191112012,
0.039229732,
0.0354096815,
0.0309999622,
-0.0907786936,
0.0167042632,
-0.0629256591,
0.0309355464,
-0.0102329021,
-0.0090294639,
-0.0008964515,
-0.0176121313,
-0.0885731056,
-0.0560639165,
-0.0189046469,
0.0179884844,
-0.0047693742,
-0.0010182632,
0.0237434376,
-0.0010874268,
0.0167860501,
-0.0041855448,
-0.0051386394,
0.0179671329,
0.0174860861,
-0.0682962462,
-0.0671648458,
-0.0457449667,
-0.0632116348,
0.0115285041,
-0.0200966895,
0.0054288544,
-0.0500630029,
0.0257465784,
0.060627304,
0.0310020801,
-0.0006573615,
-0.0175595675,
-0.1359822303,
-0.02532598,
0.0865312815,
-0.0036302775,
0.0088512786,
0.0363773443,
0.0168208648,
0.0210298393,
0.0193559099,
-0.0243063886,
0.0277088787,
-0.0313693136,
0.0444956571,
-0.0270985309,
-0.0231064279,
-0.0247884132,
0.0112596732,
0.0526453629,
-0.0476031452,
-0.0406864583,
-0.0091140419,
0.0218755715,
-0.0312892757,
0.0152097223,
-0.0229616109,
-0.0108874757,
-0.0251428112,
0.0459473021,
0.0188544262,
-0.0093745552,
-0.0571847968,
-0.0342484303,
-0.0229588151,
0.063284196,
-0.0224742312,
0.0253607891,
0.0206006467,
0.0322966091,
-0.0094670299,
-0.0060894629,
0.0311862268,
0.041811347,
-0.0119168907,
0.0742706656,
0.006816477,
-0.0341238789,
0.0354555659,
0.0454899706,
0.0167153273,
-0.0033591278,
-0.0262952074,
0.0765725449,
0.0160797667,
0.0283416063,
0.0179359335,
-0.052278772,
0.0184909161,
-0.0167224761,
-0.0004981586,
0.0153932599,
-0.0336973779,
0.0235357545,
-0.0085839331,
-0.0161418729,
-0.0126652326,
0.0082720695,
-0.0580166988,
-0.0109317088,
-0.0288453177,
-0.0135856066,
-0.0227921717,
0.0116818585,
0.0671128556,
0.01261108,
-0.0424234718,
-0.0140091795,
-0.0719338208,
-0.060799025,
0.0501073711,
0.0192468073,
0.0025429558,
-0.0162449963,
0.0522622615,
0.03082137,
0.0105616199,
-0.0124238795,
-0.0167105403,
0.0000226883,
-0.0181501284,
0.0091380449,
-0.0150956707,
-6.128526491e-33,
-0.0299815163,
-0.0137025919,
0.0067024142,
-0.0068737646,
0.0242362414,
0.0108000664,
-0.0517486297,
0.0101422127,
0.0081488499,
0.0043291235,
-0.0188413747,
-0.0131414253,
0.0242721904,
0.0285972655,
-0.0083192205,
0.0149259819,
0.023768032,
0.0310460869,
-0.0098973997,
0.0088652382,
0.0574551411,
0.0383742414,
0.0775121748,
-0.0943245441,
0.0194288511,
-0.0108799078,
-0.0092114694,
0.0015039205,
0.064453274,
0.0593696162,
-0.0049945894,
0.0574444346,
-0.0079457406,
-0.0041116928,
-0.0018555984,
0.0971659645,
-0.0295103379,
-0.0379452556,
0.0646497309,
0.0306297876,
-0.0152911097,
-0.0016015777,
0.0339369103,
-0.0117914882,
-0.0547452867,
0.0565923862,
0.0165576506,
-0.0260415077,
-0.0059275809,
-0.0296112616,
-0.0436320193,
0.0091332765,
0.0056420816,
-0.000530596,
-0.0049549928,
0.0282712746,
-0.0182519108,
-0.0416033976,
-0.0407959372,
0.0634278655,
0.0419780388,
-0.00153628,
0.0679070801,
0.0577288084,
0.018392412,
0.0315014571,
0.0207261518,
-0.0141093219,
-0.0218476783,
0.0197308585,
0.0664392784,
0.0109151434,
0.0527790114,
0.0248653833,
0.0019988003,
-0.0307873581,
0.016090909,
0.0291940253,
0.0430795439,
-0.000354652,
0.0492069311,
0.01803427,
-0.0624820665,
-0.0184322726,
-0.0164169502,
0.072810784,
-0.0082805008,
-0.0588029437,
0.0308370404,
-0.0361818708,
-0.0154586323,
0.0391289517,
0.031192081,
-0.0375305861,
-0.0239038523,
-0.0352860801,
-0.0032603892,
0.0581803992,
-0.0196093861,
0.0264509227,
-0.028605653,
-0.0206627492,
0.0443749465,
0.0232347175,
0.0045141596,
0.0287141614,
0.0225943141,
-0.0080458587,
-0.0670804307,
0.0045218449,
0.013399479,
0.0375593789,
0.0167017225,
-0.0481882431,
0.0243330095,
-0.0117560048,
-0.0458768979,
0.0239825845,
0.0024260196,
0.0215069242,
-0.049705334,
-0.0618687794,
0.0200850945,
0.0197938569,
-0.0543332547,
0.0054024607,
-0.0699076876,
0.027644366,
0.0366712287,
-0.0100940689,
-0.0131919822,
-0.0000735469,
2.803e-7,
0.0338250473,
0.0585875474,
-0.0125391111,
-0.0132122934,
0.0061351433,
-0.0029550956,
0.0232996531,
0.0064530303,
0.0369794071,
0.035220895,
-0.0147288842,
0.0044599017,
0.047692176,
0.0132659189,
-0.0953159407,
-0.0337549634,
-0.093726337,
0.0131302318,
-0.0512318499,
0.0653348491,
0.0485953316,
0.0896697566,
0.0140589345,
-0.0057056136,
-0.0069839521,
-0.0023396972,
0.0480259433,
-0.0574628673,
0.0251884572,
-0.0046886723,
0.017835917,
0.0082494412,
0.0054713679,
0.0031921254,
-0.008646504,
0.0127427587,
0.0185180828,
0.0156162735,
-0.0374357738,
0.0364343077,
0.0298039448,
-0.0159675609,
-0.04537341,
-0.0465273857,
0.0144285625,
-0.0359217748,
-0.016955018,
0.0168813821,
-0.0073671429,
-0.0073095616,
-0.0141227478,
-0.0511547737,
-0.057161279,
0.0138136977,
-0.0039944346,
-0.0287584327,
-0.0031794235,
-0.0582040511,
-0.0130205629,
0.0023727485,
-0.0055979746,
-0.066591695,
-0.0699331984,
0.007274589,
0.0022358922,
-0.0403073505,
0.0575567856,
2.329532715e-34,
0.0219018627,
0.0260659978,
-0.0055784723,
0.0190937072,
0.0547401235,
-0.0115287723,
0.0790447146,
0.0670084804,
0.0522591174,
-0.060604386,
-0.0313661583
] |
http://arxiv.org/abs/2405.15877v1
|
http://arxiv.org/pdf/2405.15877v1
|
Basis Selection: Low-Rank Decomposition of Pretrained Large Language Models for Target Applications
| 2024-05-24T18:40:20 | 2024-05-24T18:40:20 |
Yang Li, Changsheng Zhao, Hyungtak Lee, Ernie Chang, Yangyang Shi, Vikas Chandra
|
cs.LG
|
cs.LG, cs.AR, cs.CL
|
Large language models (LLMs) significantly enhance the performance of various
applications, but they are computationally intensive and energy-demanding. This
makes it challenging to deploy them on devices with limited resources, such as
personal computers and mobile/wearable devices, and results in substantial
inference costs in resource-rich environments like cloud servers. To extend the
use of LLMs, we introduce a low-rank decomposition approach to effectively
compress these models, tailored to the requirements of specific applications.
We observe that LLMs pretrained on general datasets contain many redundant
components not needed for particular applications. Our method focuses on
identifying and removing these redundant parts, retaining only the necessary
elements for the target applications. Specifically, we represent the weight
matrices of LLMs as a linear combination of base components. We then prune the
irrelevant bases and enhance the model with new bases beneficial for specific
applications. Deep compression results on the Llama 2-7b and -13B models,
conducted on target applications including mathematical reasoning and code
generation, show that our method significantly reduces model size while
maintaining comparable accuracy to state-of-the-art low-rank compression
techniques.
|
Basis Selection: Low-Rank Decomposition of Pretrained Large Language Models for Target Applications: Large language models (LLMs) significantly enhance the performance of various
applications, but they are computationally intensive and energy-demanding. This
makes it challenging to deploy them on devices with limited resources, such as
personal computers and mobile/wearable devices, and results in substantial
inference costs in resource-rich environments like cloud servers. To extend the
use of LLMs, we introduce a low-rank decomposition approach to effectively
compress these models, tailored to the requirements of specific applications.
We observe that LLMs pretrained on general datasets contain many redundant
components not needed for particular applications. Our method focuses on
identifying and removing these redundant parts, retaining only the necessary
elements for the target applications. Specifically, we represent the weight
matrices of LLMs as a linear combination of base components. We then prune the
irrelevant bases and enhance the model with new bases beneficial for specific
applications. Deep compression results on the Llama 2-7b and -13B models,
conducted on target applications including mathematical reasoning and code
generation, show that our method significantly reduces model size while
maintaining comparable accuracy to state-of-the-art low-rank compression
techniques.
|
[
-0.0278134514,
0.1079017818,
-0.0222350676,
0.0591354296,
-0.0050429632,
0.0281349663,
0.0300352424,
0.0264146943,
-0.0049593453,
-0.0025530024,
-0.0243713856,
-0.0006448204,
0.0163086839,
-0.0088936361,
0.0323585421,
-0.0073204995,
0.0084013809,
-0.0344927683,
-0.0398787893,
-0.0036179754,
0.0045442176,
0.0220046844,
0.0326361135,
0.0399774536,
0.0168113336,
-0.0238813404,
-0.0070486744,
-0.0118006859,
-0.0242536571,
0.0040323297,
-0.0092766043,
0.0591401421,
0.0523204356,
0.0191395152,
0.0000021918,
-0.0022383791,
-0.0206707008,
0.0070763212,
-0.0756436586,
0.0658266544,
0.0390584543,
-0.000795052,
-0.006355315,
0.0166330971,
-0.0219073575,
-0.0490283333,
0.0551235974,
0.0046750051,
0.0747670457,
0.0493480004,
-0.0076982724,
-0.0441131443,
-0.0251814984,
-0.0284879953,
-0.0222043265,
-0.0345524922,
0.0089659784,
-0.0158004221,
-0.0228744373,
-0.0085009094,
-0.0041173678,
-0.0023394986,
0.0332900211,
0.0359190367,
-0.0273547992,
0.030322412,
-0.0370808728,
-0.0203505028,
0.0210261066,
0.0880804732,
-0.0634816661,
-0.0353481434,
0.0213270281,
-0.0002950027,
-0.0212147199,
-0.0556757711,
-0.0377303809,
0.0311287101,
0.0404031426,
0.0464506559,
0.0075603263,
-0.013189109,
0.0540976003,
-0.0019086926,
-0.0601251721,
0.0401493013,
0.0265527181,
-0.0070201112,
-0.0185027029,
-0.0070954184,
-0.0299845487,
-0.0495319478,
0.0282242559,
0.0059975134,
0.0052858926,
0.0072424193,
-0.0763718411,
-0.0564874001,
-0.0130602345,
-0.0462662466,
0.0671134591,
0.043104887,
0.0124286758,
0.0343853496,
-0.0348515436,
0.0386258252,
-0.0335703082,
-0.0281680413,
-0.0304131452,
0.0081922598,
0.0092503559,
0.0053708316,
-0.0427223854,
0.0617730506,
-0.033357881,
-0.0188265759,
-0.0294230971,
0.0133422837,
-0.0033748189,
-0.0073400992,
0.0153068723,
0.0109446142,
0.0034064874,
-0.0120976968,
0.033054933,
0.0265953429,
-0.0121633094,
0.0239879116,
-0.0344474874,
-0.0752179325,
-0.01054681,
0.0297771227,
0.0089116823,
-0.016041711,
0.046438694,
0.0259908028,
-0.0096878484,
-0.0207316577,
-0.0273012016,
-0.005199837,
0.0399613678,
-0.0582586899,
-0.0116407927,
-0.073034361,
0.0478329957,
-0.0259757023,
0.0218080394,
-0.04944003,
0.0160426367,
0.0321784355,
-0.0352658629,
0.0449056104,
-0.0143048866,
0.0047672079,
0.0153094754,
0.0044241189,
0.0307198968,
0.0369318649,
0.0193169732,
0.0214818195,
0.0049739522,
-0.0648568198,
-0.0221650563,
0.0557308234,
0.0271556396,
-0.0241579488,
-0.0234427955,
0.0042615766,
-0.0193132572,
0.0098079452,
0.0421678573,
0.0606474839,
-0.0438144058,
-0.0048776269,
0.0261724461,
0.0758149326,
0.0527995713,
0.0926601514,
0.0336429179,
0.007072513,
0.0175167192,
-0.017199656,
-0.015985053,
0.0111583117,
-0.0458784401,
-0.0202809293,
0.0326906666,
0.0530288927,
-0.0359755754,
-0.0429881699,
-0.0215886012,
-0.0455740988,
0.0122543769,
0.0308584068,
0.0386412479,
-0.0218828041,
-0.0289439764,
0.0947527215,
0.0008813868,
-0.0759104118,
-0.0321813673,
-0.0543763153,
0.0022214365,
0.0156694129,
0.021415785,
-0.0041838349,
-0.0010682886,
0.012160358,
0.0112184267,
-0.0404495969,
0.0183780771,
-0.0078617008,
0.0400613025,
-0.078873843,
-0.0083346749,
0.0223362483,
0.0133780818,
-0.0122896098,
0.0149322134,
-0.0036567994,
0.0317235552,
-0.0182625167,
-0.0279134288,
0.0319962315,
-0.0149231879,
-0.0164548811,
0.0553093068,
0.010171406,
0.0450010151,
-0.0340005644,
-0.0160585288,
-0.0082019046,
0.0569989644,
-0.0675493926,
0.0274085999,
-0.0514242761,
-0.0362735167,
-0.0314768776,
0.0296266917,
-0.006308117,
0.0236395542,
0.0914396197,
-0.0283952467,
-0.0351117402,
0.0292571355,
-0.0596707836,
0.016216062,
-0.0424099565,
0.0011060449,
-0.0156877246,
0.0567576699,
0.0041162209,
0.0578325912,
-0.0125232097,
0.0160866063,
0.0645229518,
-0.0107921846,
0.0204728507,
-0.0234641936,
-0.0529709607,
0.0208263732,
-0.0360907577,
0.03926282,
0.00677943,
0.0224710554,
0.0160602275,
-0.0125078941,
-0.1003812104,
0.0319142267,
-0.031276118,
0.0208895802,
0.0181838218,
0.0245234612,
0.0230342168,
-0.020553967,
-0.005227874,
-0.0054599904,
-0.0017487039,
0.0108315013,
-0.0092959506,
-0.025285447,
-0.0224867053,
-0.0244266018,
-0.0453999713,
0.0141567308,
0.0856001228,
-0.0545036048,
0.0128093297,
-0.0340464637,
0.0658277944,
-0.0175701734,
0.0306902714,
-0.0022400098,
-0.0480532274,
-0.0081404028,
-0.0275094192,
0.0360669605,
0.0597769544,
-0.0225102734,
0.0443569124,
0.0932131037,
0.033967264,
-0.0253200121,
0.0009352863,
0.0073430622,
-0.0007397106,
-0.0121444548,
-0.0354463421,
-0.0164041724,
0.084914051,
-0.0070189852,
-0.0170561429,
-0.0489592701,
-0.0222365055,
0.002417316,
-0.0032465854,
-0.0535079278,
0.0355689526,
0.0011027762,
0.0479933321,
-0.0007475137,
-0.0564121939,
0.042075932,
-0.0437885188,
-0.0136312461,
-0.0277417433,
-0.0350342095,
-0.0228392556,
0.0271445513,
0.0002080812,
0.012421133,
-0.0366602466,
-0.0071407249,
-0.0260139443,
0.0264532734,
-0.0132060368,
-0.0442435108,
0.0127114281,
-0.0391015299,
0.0242228452,
-0.0110312477,
-0.0495408028,
-0.0008986392,
-0.0185297374,
-0.0239900164,
0.0175100528,
0.0215443335,
-0.002840711,
0.0023151301,
0.0125847477,
0.0091754198,
0.05175649,
0.033619754,
0.0265740398,
-0.0056085745,
0.0226679258,
0.01832835,
0.013874365,
0.0245883055,
-0.0269334223,
0.0122188423,
0.0124495318,
0.0443121605,
0.0297572836,
-0.0680653602,
0.0100465436,
0.0460756868,
0.0168536641,
-0.0704328641,
0.0557940789,
-0.0093698818,
-0.0556873679,
-0.0031335754,
0.0935703963,
0.0113315396,
-0.056604173,
0.0165580064,
0.0097463531,
0.0539804548,
0.0146603305,
0.0455711745,
-0.0698202103,
-0.0331961736,
0.0150432112,
-0.0204097293,
0.0347203687,
-0.0053537306,
-0.0083154133,
-0.0940191597,
0.015812099,
0.023043612,
0.0247654822,
-0.0133738853,
-0.0652859956,
0.0080616251,
-0.0134048164,
-0.0171679389,
0.0192004442,
-0.0149304327,
-0.0209680144,
0.0076261456,
-0.1001464203,
-0.0688798279,
0.0255350545,
0.0142428959,
-0.0012016802,
0.0370456502,
0.0326080993,
-0.0061288266,
-0.023094967,
0.0209467877,
-0.015865447,
-0.0185773522,
-0.0352970399,
-0.0026290158,
0.0426880009,
-0.0198840369,
0.0157542545,
0.0455610529,
-0.0385425873,
-0.0113353664,
-0.0111562405,
0.1244604588,
-0.0210143328,
-0.0287987124,
0.0035415101,
-0.0469706953,
-0.0221444983,
0.0613089278,
0.0393531621,
-0.0369233266,
-0.0424537919,
0.0137716318,
-0.004831057,
-0.0213784184,
-0.0393803306,
0.0005925411,
0.0049328506,
-0.0353265367,
-0.0150338123,
0.0274261348,
0.0363066494,
0.0176748205,
-0.026709659,
0.0057782019,
0.0198579393,
0.0152325761,
-0.060691569,
0.0323754549,
-0.0289732944,
0.0073148273,
-0.0233428031,
0.0157787763,
0.0009174796,
-0.0011439344,
-0.0114752948,
-0.0125801582,
-0.1007929742,
-0.0478293598,
-0.0130227739,
0.0346874595,
0.0205808114,
0.0202868711,
0.0018248273,
-0.0854448378,
0.0174090359,
0.0018713883,
-0.0261549763,
0.0701941103,
0.0268162079,
-0.0737667903,
0.0069069127,
0.0448817872,
-0.0294013508,
0.0248445645,
0.0062063965,
0.0086210184,
-0.0276630856,
-0.0294303037,
-0.0029887862,
0.0540052578,
0.0271847025,
-0.0026459501,
-0.0124965571,
-0.0104150269,
0.0313425548,
0.0050437916,
0.0125222206,
-0.0497038588,
0.0204207134,
0.0172069557,
0.0248311274,
0.009478407,
0.0005438052,
0.0454939492,
0.0137699358,
-0.0087455269,
0.0465976559,
-0.0110313976,
0.021174781,
0.06407585,
0.0000859752,
-0.0263222158,
-0.008370786,
-0.0049496153,
-0.0109169763,
-0.0680360645,
-0.0167834312,
-0.0153538147,
0.0382809527,
-0.0603304915,
0.0384136848,
-0.0024109394,
-0.0844930261,
-0.0099867703,
-0.0088837575,
0.0420166291,
0.0040717483,
0.0222158004,
-0.0117405765,
-0.0087150838,
-0.043769028,
0.031656716,
0.0019469097,
-0.0050555901,
0.0113264155,
0.0161338616,
-0.0302074347,
0.0094886934,
0.0934404135,
0.029870104,
0.0372202583,
-0.0074907551,
0.0045289085,
-0.0547421612,
-0.0199162439,
-0.018339619,
0.0359472297,
-0.0252520032,
-0.0274481252,
-0.0040135887,
-0.0657719001,
0.0093048783,
-0.0098530902,
-0.0293594468,
-0.0104597807,
0.04433088,
0.0762216076,
-0.012498281,
0.0219453257,
0.0195984617,
-0.0079197753,
-0.0267220456,
-0.0393188,
-0.0090512959,
-6.356075014e-33,
0.0108441133,
0.0320081413,
0.0126900002,
-0.0363038443,
-0.0372361988,
0.02275615,
-0.0185565371,
-0.0075748595,
-0.0233796593,
-0.0113147916,
-0.0238417853,
0.0175372679,
-0.0043487316,
0.0296223648,
-0.0229252577,
0.0184490662,
0.0576942749,
-0.0042423168,
-0.0197444241,
0.0454336591,
0.02323322,
0.0345102213,
0.0294691473,
-0.0402054265,
-0.0735565498,
-0.0161487292,
-0.0015680445,
-0.0112904822,
-0.0200353675,
0.0540677831,
-0.0660503805,
0.0357449725,
0.021755185,
-0.0234863609,
-0.020164391,
0.0490053743,
-0.0337690674,
-0.0275865439,
0.023443969,
-0.0067355148,
-0.0085883718,
-0.0709880367,
0.0666197464,
0.0340396874,
-0.0306535102,
-0.0177923497,
0.0058443057,
-0.0324501731,
-0.0259237774,
-0.015407932,
-0.0125595676,
0.0012909711,
-0.0118397335,
0.0004726363,
0.0381730758,
0.0299103688,
-0.0739794001,
0.037516918,
-0.0254896581,
-0.0004585846,
0.0186181329,
0.088091284,
0.0619265586,
-0.0071989633,
0.0331002474,
0.0422193334,
0.0023851229,
-0.0415970124,
-0.0438145958,
0.0033386417,
0.0627386346,
0.0585704073,
0.0495105274,
0.0258282106,
0.0663068146,
-0.0164269805,
-0.1088453382,
-0.0230830703,
0.0367358103,
0.0144542921,
0.0188216846,
0.0178971477,
0.0115452921,
-0.0461399555,
-0.0348203108,
0.0087072151,
-0.0312717557,
-0.0267253518,
0.0228586681,
0.0087195188,
-0.0486628413,
0.0041778767,
0.0506282635,
-0.0536205918,
-0.0068998425,
0.0517424345,
-0.0087079545,
0.0103002544,
0.0395049825,
0.0312820524,
-0.0475588292,
-0.0376544446,
-0.0062572984,
-0.0395995378,
0.0629085973,
-0.0225724503,
0.0031496065,
0.0438758247,
-0.0522640161,
-0.0301525835,
-0.033935789,
-0.0358245596,
0.0179449171,
0.0767887607,
0.043184597,
0.016491415,
0.0273236018,
0.0513024181,
-0.0259437971,
0.0113024535,
-0.0307574552,
-0.0408860557,
0.0485674553,
0.0426316559,
-0.021038292,
-0.0271248259,
-0.0750642493,
0.031900838,
0.0607777052,
-0.0578100905,
-0.0422199368,
0.0324754491,
2.864e-7,
0.009186361,
0.0031425655,
-0.0266548861,
-0.0488130748,
0.0145964716,
0.0273231845,
-0.0007373624,
0.01758877,
-0.0311562773,
-0.0117986808,
0.0171544272,
0.0259439871,
0.0603822246,
0.0072388449,
-0.0494673736,
0.0081901727,
-0.0764946491,
0.0055696624,
-0.0626258552,
-0.0346583873,
0.021874195,
0.1391692162,
0.0000840985,
-0.0309830587,
-0.0087868283,
0.0383454449,
0.024002159,
-0.0286016744,
0.0025133782,
-0.0188514777,
0.0123823276,
0.019961752,
-0.0442438684,
-0.0193133764,
-0.0033624326,
0.0249972101,
-0.0065001869,
0.0889956206,
-0.0044861133,
0.0422482006,
0.0285835955,
-0.024546925,
-0.0522894412,
-0.012184184,
0.0823174119,
-0.0153030595,
-0.0171031523,
-0.0449320115,
-0.0212318916,
-0.028050255,
0.0038259379,
-0.0236125365,
0.0465454757,
0.0170208011,
0.0135403229,
-0.0404749848,
-0.0241282787,
-0.0854966789,
0.0093898829,
0.0129264425,
-0.0727770403,
-0.0307724811,
-0.0590836033,
0.0008039848,
0.0754871741,
-0.0532140993,
0.0371908136,
2.95953478e-34,
-0.0131398952,
-0.00892381,
0.0234287735,
0.0577696972,
0.0007024532,
-0.0508496314,
0.0034123927,
0.0242964253,
-0.0328850709,
-0.1065615267,
-0.0137065202
] |
http://arxiv.org/abs/2406.00030v1
|
http://arxiv.org/pdf/2406.00030v1
|
Large Language Model Pruning
| 2024-05-24T18:22:15 | 2024-05-24T18:22:15 |
Hanjuan Huang, Hao-Jia Song, Hsing-Kuo Pao
|
cs.CL
|
cs.CL, cs.AI, cs.LG
|
We surely enjoy the larger the better models for their superior performance
in the last couple of years when both the hardware and software support the
birth of such extremely huge models. The applied fields include text mining and
others. In particular, the success of LLMs on text understanding and text
generation draws attention from researchers who have worked on NLP and related
areas for years or even decades. On the side, LLMs may suffer from problems
like model overfitting, hallucination, and device limitation to name a few. In
this work, we suggest a model pruning technique specifically focused on LLMs.
The proposed methodology emphasizes the explainability of deep learning models.
By having the theoretical foundation, we obtain a trustworthy deep model so
that huge models with a massive number of model parameters become not quite
necessary. A mutual information-based estimation is adopted to find neurons
with redundancy to eliminate. Moreover, an estimator with well-tuned parameters
helps to find precise estimation to guide the pruning procedure. At the same
time, we also explore the difference between pruning on large-scale models vs.
pruning on small-scale models. The choice of pruning criteria is sensitive in
small models but not for large-scale models. It is a novel finding through this
work. Overall, we demonstrate the superiority of the proposed model to the
state-of-the-art models.
|
Large Language Model Pruning: We surely enjoy the larger the better models for their superior performance
in the last couple of years when both the hardware and software support the
birth of such extremely huge models. The applied fields include text mining and
others. In particular, the success of LLMs on text understanding and text
generation draws attention from researchers who have worked on NLP and related
areas for years or even decades. On the side, LLMs may suffer from problems
like model overfitting, hallucination, and device limitation to name a few. In
this work, we suggest a model pruning technique specifically focused on LLMs.
The proposed methodology emphasizes the explainability of deep learning models.
By having the theoretical foundation, we obtain a trustworthy deep model so
that huge models with a massive number of model parameters become not quite
necessary. A mutual information-based estimation is adopted to find neurons
with redundancy to eliminate. Moreover, an estimator with well-tuned parameters
helps to find precise estimation to guide the pruning procedure. At the same
time, we also explore the difference between pruning on large-scale models vs.
pruning on small-scale models. The choice of pruning criteria is sensitive in
small models but not for large-scale models. It is a novel finding through this
work. Overall, we demonstrate the superiority of the proposed model to the
state-of-the-art models.
|
[
0.0025993518,
0.068816781,
0.0118267033,
0.0570395514,
-0.0220249891,
0.0150040444,
0.0263406411,
0.0286514498,
-0.0348881297,
-0.0017571722,
-0.0238318332,
-0.0480398536,
0.0035373694,
-0.0043360889,
0.0114001893,
-0.0052239094,
0.012589246,
-0.0331395492,
-0.0017962852,
0.0027035326,
-0.0090599507,
0.0209277123,
0.0331391022,
0.0469812155,
0.0119439205,
-0.028764056,
-0.0169000961,
0.0208676569,
-0.0087469118,
0.0193993095,
-0.0066452958,
0.0224231016,
0.037199419,
0.0396927893,
0.0000023408,
-0.0065745451,
-0.0536117069,
-0.0085157529,
-0.0651299208,
0.060288284,
0.0320172347,
-0.0052022971,
-0.0025606342,
0.0097269518,
-0.0110923173,
0.0091781886,
0.0336902961,
0.0852258727,
0.030485481,
0.0831163153,
0.0153485965,
-0.0285832733,
0.0070951432,
-0.0321769342,
0.0027800596,
-0.046254348,
0.0041045593,
-0.0079278136,
-0.0350777097,
-0.0100117316,
-0.000422963,
0.0133252069,
0.0344611555,
0.0080779446,
0.0002892908,
0.0168643836,
0.0252727158,
-0.0293435715,
0.0294248611,
0.0621804595,
-0.0560907014,
-0.0342789553,
0.0422713198,
-0.0277765617,
-0.0250156,
-0.0065224273,
-0.0479209162,
0.019399697,
0.0320744887,
0.0345064141,
0.0262840576,
-0.0311029032,
0.0898269564,
0.0337099284,
-0.0386584587,
0.0582909919,
0.006614638,
0.01857429,
-0.0398473032,
-0.0138960201,
-0.0156203946,
-0.0603271201,
0.0012628217,
0.0525659025,
-0.0170555897,
0.0046184389,
-0.0203241147,
0.0146208825,
-0.0186772477,
-0.0515908599,
0.0286755543,
0.017571358,
0.0272693429,
0.0237145294,
-0.0181230661,
0.0252095275,
-0.0134757692,
-0.0064241146,
-0.0378846265,
0.0097385328,
-0.0249606743,
-0.0285017658,
-0.0037880538,
0.0481457151,
-0.0451603457,
-0.0354381129,
-0.0281345788,
-0.0073756236,
-0.0046459502,
0.0001932981,
0.0091137653,
0.0265698135,
-0.0180303454,
-0.0239429735,
0.0011258319,
0.00444608,
-0.0362293236,
-0.000360627,
0.0188042317,
-0.0765662491,
-0.0164579079,
0.0435808562,
0.0324884057,
-0.0117671099,
0.0153946988,
0.050957527,
-0.0020080996,
-0.0469167195,
-0.0715846643,
-0.0487265885,
-0.0215879772,
-0.0806969479,
0.0141477231,
-0.0698848814,
0.0199524444,
-0.0335874371,
-0.0079209637,
-0.081671454,
0.0037541965,
0.0158042107,
-0.0398569219,
0.0438389145,
-0.0304458365,
0.019823717,
0.0335489698,
0.0227131452,
0.0303779002,
0.0597156473,
0.0505522676,
0.0533089526,
0.0161940847,
-0.0318962634,
0.019761363,
0.0154950907,
0.0391679406,
-0.0139203947,
-0.0387098007,
-0.0414670035,
-0.0532541461,
0.0068223658,
0.0251279306,
0.0563892201,
-0.0363597199,
0.006083929,
0.0610421449,
0.0175806805,
0.0406703241,
0.0626165867,
0.0610123985,
0.0029367241,
-0.0042467667,
0.0301278662,
-0.0185931884,
0.0233186949,
-0.0235248283,
0.0031029379,
0.0213868115,
0.0236907844,
-0.0486855395,
-0.0711036772,
-0.0189308748,
-0.0264019184,
-0.0215185136,
0.0371442959,
0.0374731794,
0.0044479929,
-0.0436519496,
0.1049681455,
0.0106797731,
-0.0450279042,
0.017516423,
-0.0623266846,
0.0434552729,
0.0606159754,
0.0371478423,
-0.0305549391,
0.0056606922,
-0.0099342735,
0.0259002727,
0.0080000199,
0.0213789288,
0.0123311346,
0.035825152,
-0.0713747591,
-0.0126655977,
-0.029319765,
0.0394574776,
-0.0039865794,
0.0057619857,
-0.0216371696,
0.0524002239,
-0.0348682553,
-0.0616573244,
0.0318723582,
-0.067381613,
-0.0441921279,
0.0220489707,
0.010095207,
0.0236311425,
0.0111931833,
-0.0099195819,
0.0215725768,
0.0432964675,
-0.0341765545,
0.0117826182,
-0.0173976962,
-0.0300306752,
-0.035864111,
0.0194486659,
-0.0062885326,
0.0501928404,
0.0449031927,
-0.0228566248,
-0.067625694,
-0.0041230135,
-0.0481404699,
0.0398009829,
-0.0320353471,
0.0387938507,
0.0033489496,
0.0426513553,
-0.0105331121,
0.0508936718,
-0.0118196579,
0.0058456534,
0.0307141989,
-0.0058869696,
-0.005627695,
-0.0486125983,
-0.0300654862,
0.0249605477,
-0.0317738466,
-0.0090996493,
0.0113226967,
-0.0038068616,
-0.0093853967,
-0.0262905713,
-0.0470470563,
0.0267441552,
-0.0380743854,
0.0206044186,
0.0310524721,
0.0107745035,
0.0383317508,
0.0152971242,
-0.0163532831,
0.0104466444,
0.0357160531,
0.0048188316,
0.0255735274,
-0.0227829814,
-0.0249542426,
-0.0195948537,
-0.0101560876,
0.0085780406,
0.0850653797,
-0.0234604813,
-0.0453199111,
-0.0531628355,
0.0617583692,
-0.040356908,
-0.0072417916,
-0.0103432508,
-0.0330766514,
-0.0175525397,
-0.0016623455,
-0.0149292601,
0.0381874964,
0.0071759345,
0.0767422095,
0.0429762118,
0.0159482453,
-0.039614439,
0.0054457141,
-0.017242182,
0.0290405471,
-0.0005623396,
-0.0751449689,
-0.0106159654,
0.0836322755,
-0.0040220949,
0.001908369,
-0.0323024765,
0.0151833808,
0.0158513673,
-0.0416690446,
0.0008981554,
0.0050885333,
-0.004072336,
0.0441150144,
-0.0340594724,
-0.0535781458,
0.0167946015,
-0.0638320819,
0.0137722632,
0.018760886,
-0.0388939641,
-0.0308020413,
0.0459759533,
0.0145550519,
0.0310832541,
0.0069887647,
0.0040652761,
-0.0305028185,
-0.0086375168,
-0.0070457393,
-0.0102304853,
0.0270899385,
-0.0464786887,
0.0212347358,
-0.0052781352,
0.0046439515,
-0.0233264863,
-0.0413628891,
-0.0199385826,
0.0236806236,
0.0353407077,
-0.0188745223,
-0.0212519262,
0.012119133,
0.0285984222,
0.0346618146,
0.0084168119,
0.0201040916,
-0.0286664963,
0.040148437,
0.072076939,
0.0152290752,
0.0536100231,
-0.0651001781,
0.0643710718,
0.0418530963,
0.0137759922,
0.0430072285,
-0.075094156,
-0.0002264801,
0.0054347804,
-0.0206597727,
-0.0342016555,
0.0325662345,
0.06789013,
-0.0525148101,
-0.0069127502,
0.1201163307,
0.0046521784,
-0.0427170992,
0.0411272496,
0.0422486886,
0.0390145555,
-0.0136974407,
0.0133872963,
-0.0255000368,
-0.0266170539,
0.0267644376,
-0.0083497353,
0.0068721701,
-0.0154947629,
-0.0342357792,
-0.0808694139,
0.0259256773,
0.0249689259,
-0.0003710066,
0.0004146917,
-0.0199995283,
-0.0290759131,
-0.0050587412,
0.0112750288,
0.0130717577,
0.01461164,
-0.0177224334,
0.0321993977,
-0.0503026545,
-0.0520161279,
0.0010367755,
-0.0047631441,
0.028967876,
0.0502577275,
0.06391453,
0.0224885885,
-0.0043278961,
0.0357598811,
0.0132764243,
0.0161669385,
-0.0321539231,
0.0564961806,
0.0279814266,
-0.012293525,
0.0133922426,
0.0409140475,
-0.0552680083,
-0.0035156023,
-0.0312477872,
0.0571477078,
-0.0256708693,
-0.0383626819,
0.0272240881,
0.0014100779,
-0.0296267122,
0.0904457942,
0.0581385046,
-0.0169374831,
-0.0569995679,
-0.0149696702,
-0.0066495473,
-0.051125966,
-0.0100421831,
0.0065344679,
0.0260866694,
-0.0438192599,
-0.0583466887,
0.0064363289,
-0.0029539536,
0.0309781097,
-0.0184124317,
0.0377193131,
-0.0067786113,
-0.0088219885,
-0.0241100509,
0.0246939473,
-0.0347028971,
-0.0053163487,
-0.042204719,
0.0210802257,
0.0187893547,
-0.0573098622,
-0.0085084513,
0.0047087977,
-0.1126330048,
-0.0804815888,
-0.0555499569,
0.0243414752,
0.0277763959,
0.0667403042,
0.0123722907,
-0.0944947973,
0.0235415716,
0.0271531586,
-0.0147049995,
0.0279264525,
0.0108380029,
-0.0152984653,
-0.0135678891,
0.0579731688,
-0.0129759274,
0.0147409728,
0.0523291416,
0.0130359139,
-0.0346228443,
-0.0221342463,
0.0034722856,
0.0356386565,
-0.0058965269,
-0.0188877564,
-0.0238540042,
-0.0360477678,
0.0033021914,
0.0000920238,
-0.0101767872,
-0.0775852501,
0.0556899868,
0.0253317263,
0.0082957214,
-0.0139837014,
-0.0208833013,
0.0555164032,
-0.0097930692,
-0.0271480177,
0.0069371401,
-0.0371720716,
-0.0025655802,
0.0167503022,
0.0008959796,
-0.0266191047,
0.0214737635,
-0.0396291278,
-0.0066261948,
-0.0012402848,
0.0308337137,
0.0067000207,
0.0197422337,
-0.0545463786,
0.0145754963,
-0.0602434762,
-0.099384442,
-0.0461434424,
0.0554088987,
0.054916013,
-0.0129915746,
0.0465509556,
0.0049577472,
-0.0000252346,
-0.046676714,
0.0772796422,
0.0450721569,
-0.017278811,
-0.0017278135,
0.0125775458,
-0.0470741503,
-0.0275297146,
0.0956506208,
0.0569127202,
0.0237437859,
0.0023264545,
-0.0211808234,
-0.0547330119,
-0.0415328257,
-0.0217619445,
0.024447605,
-0.0185295194,
0.0134013789,
-0.01094684,
-0.0616630651,
0.0084094359,
-0.0301606022,
0.0146672046,
-0.0358884819,
0.0265930817,
0.0660600439,
-0.006821895,
0.0126355691,
-0.0422914401,
-0.0140040601,
-0.0103215184,
-0.0116114141,
0.0040536956,
-6.835230756e-33,
0.0039249761,
-0.0351706408,
0.0190330148,
-0.0306772292,
-0.0243283473,
0.0183728375,
-0.0491475761,
-0.0353235453,
-0.0327501334,
0.0086619584,
-0.0134719172,
0.0196876414,
0.0081823096,
0.0271677934,
-0.0563853532,
-0.0133108757,
0.0752224475,
0.006029638,
-0.0255990475,
0.0522832312,
0.0388111211,
0.0625144467,
0.0666303858,
-0.0221927967,
-0.0196332876,
-0.0068013626,
-0.0271484312,
0.0159065444,
-0.0389243215,
0.0387471728,
-0.0347751305,
-0.0003238523,
0.0439967848,
-0.0393402018,
-0.0178156383,
0.016126506,
-0.043792028,
0.0059585641,
0.0166266747,
0.0113870539,
-0.0171330515,
-0.0751503035,
0.0864384547,
0.009788637,
0.0033563192,
-0.0104339747,
0.0255919695,
-0.0257968344,
-0.0444245636,
-0.0178701039,
-0.0304819904,
0.0399373546,
-0.0021601452,
-0.0410920419,
0.0319871418,
0.0336876176,
-0.0463155396,
0.0438075624,
-0.0446165688,
-0.0274289437,
0.0421411581,
0.07737948,
0.0839219764,
0.0182824507,
-0.0144144362,
0.0723989531,
0.0067839269,
0.0072155441,
-0.0520244613,
0.0135080516,
0.010509206,
0.0976410136,
0.0493496992,
0.0253087915,
0.0351404399,
-0.0243602935,
-0.0978670716,
0.0026094816,
-0.0037732406,
0.0159005262,
0.0540263094,
0.0200266447,
-0.0155466655,
-0.0261033271,
-0.0271653067,
-0.0416749269,
-0.0234343391,
-0.0103615597,
0.0275119506,
0.0272835083,
-0.0774777308,
-0.0272639077,
0.0320090763,
-0.0747051612,
0.0017393017,
0.0434770361,
-0.0177297462,
0.0060739401,
0.0018850912,
-0.0059207799,
-0.0442694873,
-0.0535967723,
-0.0360218212,
-0.003789247,
0.0169567708,
-0.0085839592,
0.0105148936,
0.022488473,
-0.0248104017,
-0.0148710543,
-0.0778374597,
-0.013619761,
0.014634504,
0.064497076,
0.036351189,
-0.0287542548,
-0.026351409,
0.0347559266,
0.0082923481,
0.0100148702,
0.0066798697,
-0.0101174191,
0.0031706688,
0.0432705656,
-0.0269673597,
-0.024846362,
-0.0192244593,
-0.0040417449,
0.070207946,
-0.0450567417,
-0.0339557678,
-0.0011653195,
2.878e-7,
-0.0139925582,
-0.0124711143,
0.0037141745,
-0.0032322062,
0.0265490003,
0.0322018377,
0.0146673871,
-0.0243601352,
-0.0388699919,
-0.031583488,
0.0283208992,
0.0359967314,
0.1013625711,
-0.0156762898,
-0.071646437,
-0.0052480362,
-0.078873165,
0.004055284,
-0.0196096487,
-0.0547996201,
-0.0145122856,
0.0988125131,
0.0086151492,
-0.0308533255,
0.0042131399,
-0.0269474946,
0.0479651354,
-0.0381031334,
0.0099731963,
0.0404490158,
0.0728851855,
-0.0204142425,
-0.0316572227,
-0.0247748327,
-0.0177773423,
0.0085364385,
-0.0086308504,
0.0949546099,
0.0161013007,
-0.0009569318,
0.0172209665,
-0.0082262596,
0.0097422013,
-0.0476924069,
0.0723968074,
0.017100079,
-0.0222360343,
-0.0491695702,
0.0117915701,
-0.0493565872,
-0.0164819006,
-0.021239059,
0.0486463718,
0.026963396,
-0.0288598984,
-0.0535053723,
-0.0015395286,
-0.0552896447,
-0.0394691005,
-0.0211305078,
-0.0836761817,
-0.030196324,
-0.0609034821,
-0.0150600011,
0.056549307,
-0.0585707501,
-0.0097346958,
2.840984109e-34,
0.0082730018,
-0.0194997247,
-0.015840197,
0.0511891469,
0.0476928502,
0.0012687474,
-0.017496679,
0.0027667733,
-0.0007118639,
-0.0626156852,
-0.030852722
] |
http://arxiv.org/abs/2405.15766v2
|
http://arxiv.org/pdf/2405.15766v2
|
Enhancing Adverse Drug Event Detection with Multimodal Dataset: Corpus Creation and Model Development
| 2024-05-24T17:58:42 | 2024-05-27T02:55:45 |
Pranab Sahoo, Ayush Kumar Singh, Sriparna Saha, Aman Chadha, Samrat Mondal
|
cs.AI
|
cs.AI, cs.CL, cs.CV
|
The mining of adverse drug events (ADEs) is pivotal in pharmacovigilance,
enhancing patient safety by identifying potential risks associated with
medications, facilitating early detection of adverse events, and guiding
regulatory decision-making. Traditional ADE detection methods are reliable but
slow, not easily adaptable to large-scale operations, and offer limited
information. With the exponential increase in data sources like social media
content, biomedical literature, and Electronic Medical Records (EMR),
extracting relevant ADE-related information from these unstructured texts is
imperative. Previous ADE mining studies have focused on text-based
methodologies, overlooking visual cues, limiting contextual comprehension, and
hindering accurate interpretation. To address this gap, we present a MultiModal
Adverse Drug Event (MMADE) detection dataset, merging ADE-related textual
information with visual aids. Additionally, we introduce a framework that
leverages the capabilities of LLMs and VLMs for ADE detection by generating
detailed descriptions of medical images depicting ADEs, aiding healthcare
professionals in visually identifying adverse events. Using our MMADE dataset,
we showcase the significance of integrating visual cues from images to enhance
overall performance. This approach holds promise for patient safety, ADE
awareness, and healthcare accessibility, paving the way for further exploration
in personalized healthcare.
|
Enhancing Adverse Drug Event Detection with Multimodal Dataset: Corpus Creation and Model Development: The mining of adverse drug events (ADEs) is pivotal in pharmacovigilance,
enhancing patient safety by identifying potential risks associated with
medications, facilitating early detection of adverse events, and guiding
regulatory decision-making. Traditional ADE detection methods are reliable but
slow, not easily adaptable to large-scale operations, and offer limited
information. With the exponential increase in data sources like social media
content, biomedical literature, and Electronic Medical Records (EMR),
extracting relevant ADE-related information from these unstructured texts is
imperative. Previous ADE mining studies have focused on text-based
methodologies, overlooking visual cues, limiting contextual comprehension, and
hindering accurate interpretation. To address this gap, we present a MultiModal
Adverse Drug Event (MMADE) detection dataset, merging ADE-related textual
information with visual aids. Additionally, we introduce a framework that
leverages the capabilities of LLMs and VLMs for ADE detection by generating
detailed descriptions of medical images depicting ADEs, aiding healthcare
professionals in visually identifying adverse events. Using our MMADE dataset,
we showcase the significance of integrating visual cues from images to enhance
overall performance. This approach holds promise for patient safety, ADE
awareness, and healthcare accessibility, paving the way for further exploration
in personalized healthcare.
|
[
0.0052774455,
0.0357553065,
0.0130962133,
-0.0139511609,
0.0201357957,
0.0532093123,
-0.031611748,
0.0261732582,
0.0461208485,
-0.0301136039,
0.0166255701,
-0.01793411,
-0.0461054221,
0.046057608,
0.0388903506,
-0.0428320467,
0.0121737216,
0.0549724586,
0.0076196743,
0.0428424589,
0.0377803035,
0.0296418816,
-0.0261991005,
0.0452276468,
-0.0215377659,
-0.0083331354,
0.0113556059,
0.0169294309,
0.0663058683,
-0.0929318443,
-0.0285752807,
0.0030073687,
0.0327479057,
-0.0300255325,
0.000001978,
-0.0245238952,
-0.0019759359,
0.0081415428,
-0.0006948741,
-0.0831008181,
0.0661868453,
-0.0728758425,
0.0098559512,
-0.0165068842,
-0.0285221301,
-0.0221233945,
0.0660079718,
0.0507775135,
0.0191621538,
0.0538830198,
-0.0352280177,
0.0148370275,
0.0065402463,
0.0161869116,
0.057012748,
-0.0422159247,
0.0055981944,
-0.0126351882,
0.0317763202,
-0.0250925478,
0.0041343933,
0.0299297404,
-0.0225950368,
0.0168449301,
0.0219757669,
0.0415627398,
0.0001636131,
-0.0365890712,
-0.0338758267,
0.084358409,
-0.0084057348,
0.0316014327,
0.0029229443,
-0.0195944533,
-0.0280048568,
0.0440916717,
-0.0222963374,
0.0097820098,
-0.0391737111,
0.0326978117,
-0.0410494134,
0.0040414906,
0.0211226493,
0.0022383919,
0.0371421874,
0.0028037841,
0.0105310166,
0.0039098007,
-0.0058676624,
0.0038982055,
-0.011360527,
-0.0190674607,
0.0330971442,
-0.003279197,
-0.0166765638,
0.0251043532,
-0.0013153289,
-0.0417628251,
0.086658217,
-0.0952489749,
0.0568579063,
-0.0225756411,
-0.0551156811,
0.0610012561,
0.0041766204,
-0.0581022054,
-0.0177431554,
0.025296161,
-0.0107910866,
0.0358212739,
-0.0146016544,
0.0190894809,
-0.0065377946,
-0.0324156657,
-0.0014730477,
0.0037252845,
-0.0202692356,
0.0144054145,
-0.0080158981,
0.0081247594,
0.0033469244,
0.0461756699,
-0.0332589336,
0.011166784,
-0.0200995617,
-0.0072023333,
0.0258796886,
0.0150412237,
-0.0124493688,
0.0062857536,
0.0499526039,
0.0056012534,
0.0542098731,
-0.0057753986,
-0.0040276228,
0.0885575116,
0.0055626929,
-0.0227538832,
-0.0723757967,
-0.0643606111,
-0.0132450657,
0.0238586981,
0.0426784977,
-0.0394619592,
0.0633743852,
-0.0308781005,
0.006958676,
-0.0108964844,
0.0106369611,
0.0125156688,
-0.0425532572,
0.0746376365,
-0.0635961443,
-0.0044596205,
0.0103426771,
0.0269603785,
0.0116428882,
0.0329198278,
0.0357122011,
0.0433197282,
0.0535976812,
-0.0201075841,
-0.0003772967,
0.0215961691,
-0.0197335649,
-0.0145850908,
-0.0141871152,
-0.003482583,
-0.0238929894,
0.0037089526,
-0.0221069884,
0.0410038829,
-0.0084169386,
-0.0191302001,
0.0384014063,
0.0242438521,
0.0268886704,
-0.0443268046,
0.0062778564,
-0.0274740402,
-0.0549288318,
-0.0197613798,
0.0013911144,
0.0155111607,
0.0026906929,
-0.0196680743,
0.0168710425,
0.0243252255,
-0.0101568252,
-0.0155042987,
-0.0560654849,
-0.0381797701,
0.0067922547,
-0.0046920781,
0.0353907421,
-0.026517611,
-0.0398784913,
0.0437453836,
0.0589281209,
-0.0530247092,
-0.0296103172,
-0.0085820248,
0.0027142414,
0.1103027463,
0.0398164093,
-0.0256121997,
-0.0144581236,
-0.0218528137,
0.0258188732,
-0.0084465612,
-0.0142111294,
-0.0218518265,
0.0433995724,
-0.028736975,
-0.0451548174,
0.0233272109,
0.0068780077,
-0.044042781,
-0.0368294083,
0.0178532787,
-0.0399401225,
-0.0300778486,
-0.0232903734,
0.0519026406,
0.0076116445,
-0.034970928,
0.0143392971,
0.0433080792,
-0.0009565937,
0.0443737395,
0.0248125847,
-0.0135546569,
0.083234027,
-0.0684409142,
-0.0470127761,
0.0143897459,
0.015116428,
-0.0192847885,
0.0181099195,
-0.0690232813,
-0.0011854107,
-0.0327119902,
-0.0804492384,
-0.0065118191,
0.0029948344,
-0.0619356334,
0.0057680961,
-0.008451079,
-0.0313534662,
0.0777228326,
0.0296918266,
-0.0244757794,
0.022689214,
-0.0204095468,
-0.0270167869,
-0.0200908221,
-0.042041108,
-0.0080653504,
-0.0103676906,
0.0107973078,
0.0017186161,
0.0070424234,
0.0226234253,
0.0082853697,
0.0034127957,
-0.016412355,
0.0362863541,
-0.0690053329,
-0.031187756,
0.027059054,
-0.0067392164,
0.0300590657,
0.0207333453,
-0.0184737314,
-0.0540427715,
-0.0102689518,
-0.0085302098,
-0.0042576711,
0.0019151132,
0.0478385128,
0.0529421084,
0.0187123343,
-0.0211499743,
-0.0233386084,
-0.0392400436,
0.0486842319,
0.0077540935,
-0.0626930222,
-0.0215182342,
-0.0358875208,
-0.0261562709,
0.0106348507,
0.0081254747,
0.0127310818,
0.0056560095,
-0.0091278981,
-0.0116825877,
0.0180993974,
-0.0057233199,
0.0565506667,
0.0323826298,
-0.0269841012,
0.0028439816,
0.0081868814,
-0.0679775402,
-0.0157577507,
0.0586311109,
-0.0776705742,
0.009276473,
0.0390553623,
0.0155116562,
0.0190228168,
0.0085192304,
-0.0295475163,
-0.0046613375,
0.0096250447,
-0.0510000139,
0.0058119772,
-0.0288954172,
0.0298410822,
-0.0146667976,
-0.0660824999,
0.0517689213,
0.0139657222,
0.0348257832,
-0.028528966,
-0.0266220067,
-0.0539662093,
-0.0299974568,
0.0350866392,
0.0396921672,
0.0297050271,
0.0207921732,
0.0412956327,
0.019020332,
0.013428702,
-0.0246071573,
0.0667465851,
-0.0339926779,
0.0015026428,
-0.0156060914,
-0.0279998593,
-0.018723771,
-0.020748321,
0.0072705518,
-0.0495583713,
-0.0394458249,
0.0235690698,
-0.0411600433,
0.0407973826,
-0.0477197543,
0.0121239023,
0.0466176607,
-0.0393024199,
0.0295878835,
0.0395714007,
0.0485750698,
-0.0275167283,
0.0082783829,
-0.0185232554,
0.030438561,
0.1003718227,
-0.0031701261,
0.0103715146,
-0.0242771991,
0.04583266,
-0.0256036092,
0.0064996821,
0.0116648749,
0.0323068723,
0.0330379233,
-0.0280849487,
0.0591909327,
0.0324950665,
0.0192534812,
-0.080109112,
0.0400984138,
0.0143227801,
-0.0745572522,
-0.0233649556,
-0.0257837754,
-0.0844636038,
-0.0387204923,
-0.0499085821,
0.0098086698,
-0.0519457757,
-0.0119994534,
-0.0169385653,
-0.0017041506,
0.0549127199,
-0.0005825216,
-0.0119138109,
-0.0265551247,
-0.04505964,
0.0116891004,
-0.0409148745,
-0.0028389683,
-0.0432707965,
-0.0194096752,
-0.044477433,
0.0054755365,
-0.0296521001,
-0.0874992907,
-0.0165462829,
-0.020693114,
-0.0390264131,
0.0246875156,
0.0286270287,
0.0008706613,
-0.02483337,
-0.0151335429,
0.009002856,
0.0224515758,
0.0034636084,
0.0906641036,
-0.0254869647,
-0.0067233201,
-0.0632081181,
0.0518797934,
-0.0453570224,
-0.0407529548,
-0.0185006652,
0.0786010921,
-0.0399424732,
-0.0159027558,
-0.1037819609,
0.0236236546,
0.0393132791,
0.0170352198,
-0.0068339542,
-0.0074451403,
-0.0070633707,
0.002675886,
0.0437665246,
-0.037399631,
-0.0117700137,
0.0178840533,
-0.0304839388,
-0.036353495,
-0.0728372186,
-0.0051869755,
0.0129521582,
-0.0445507988,
-0.023033103,
-0.024236571,
-0.0021490818,
-0.0131858801,
-0.0636181906,
0.0671297684,
0.053214062,
0.0305618569,
-0.0177962873,
-0.029610008,
0.0083837267,
0.0177220497,
0.0088554882,
0.0570706017,
-0.0583874844,
-0.0128764836,
-0.0017151313,
0.0152382515,
0.0120589007,
-0.0343137719,
0.0142509518,
-0.022467725,
0.0386008359,
0.08995042,
0.0023555311,
-0.0461716913,
-0.0264830049,
-0.0072517609,
-0.0186953917,
-0.0259959605,
-0.0226612538,
0.026427662,
-0.0408336371,
-0.0155376615,
-0.0249479078,
-0.010169887,
0.0092243133,
-0.0042629307,
0.0459754467,
-0.0480398312,
-0.0247291587,
-0.0489299744,
0.0074576316,
0.0171627551,
0.0153436419,
-0.0189574547,
0.0401200615,
-0.0139745874,
-0.0392683968,
-0.0321284607,
0.0431996956,
-0.0437694229,
-0.0338988863,
-0.0739933848,
0.0664836392,
0.0297709331,
-0.0031596646,
0.0259121452,
0.0041660885,
0.0211470462,
0.0188919697,
-0.0374260247,
0.0178836063,
-0.0208428353,
0.0525070317,
-0.0294076689,
0.0222031418,
-0.0143249687,
0.103531085,
-0.0154778967,
0.0103954664,
-0.0341205187,
0.022834396,
0.0950267389,
-0.0203526691,
0.043244753,
-0.0422818437,
-0.0109511958,
-0.0333198048,
0.020501649,
-0.017090736,
0.006245676,
0.0725261644,
0.0139804129,
-0.0074781226,
-0.0201371964,
0.0940596759,
-0.0143191963,
0.033060953,
-0.0037148998,
-0.0327473842,
-0.0218752604,
-0.0133164255,
0.0668356791,
0.025306318,
-0.0609231405,
0.0044749952,
-0.0609399676,
-0.0388829559,
0.0330080725,
-0.01513262,
-0.0092609879,
0.0132572958,
0.041496288,
0.021660544,
0.0266136881,
-0.0066118459,
-0.006037157,
0.0336079076,
-0.0124093089,
0.0103980768,
-0.003263999,
-5.504988084e-33,
0.0060936976,
-0.061564099,
-0.00718184,
-0.0129497396,
0.0376424752,
0.0106676575,
-0.0048403828,
-0.0069727665,
0.0452434346,
0.0328205153,
0.0206844788,
-0.0247903764,
-0.0011247138,
0.0302982889,
0.0170229767,
-0.0247860625,
-0.0008644268,
0.014970839,
-0.00703301,
0.030753294,
0.0214345846,
0.0616712756,
0.034067478,
-0.1027671695,
0.0123475827,
0.0258040801,
-0.0028993755,
-0.0214013308,
0.0194632765,
0.0770329013,
-0.0477405787,
0.0746369958,
0.0462668985,
-0.0064652753,
-0.054470174,
0.0184558164,
-0.0479381047,
-0.0106718009,
-0.0236295946,
-0.014515358,
-0.03241615,
-0.0226711445,
0.0446743369,
-0.0190037992,
-0.0143852625,
-0.0174578,
0.0471475869,
-0.0302641578,
-0.0219393373,
-0.0321287513,
-0.0167930275,
0.0586171448,
-0.0059121586,
-0.0047336556,
0.0150462184,
0.0169761255,
0.0188809484,
0.0540080369,
-0.0380928442,
0.0257355105,
-0.0557081625,
-0.0447745845,
0.0282104835,
-0.0250909999,
-0.0243596658,
0.0782358125,
0.0208804943,
-0.0093967607,
0.0015276847,
0.0410823263,
-0.0719613731,
0.0817586333,
0.0091733998,
0.0495673716,
0.0400130637,
0.0136946756,
-0.0568891503,
0.0586442277,
-0.1233625785,
0.0282823704,
0.00887434,
0.0097768744,
-0.0183766987,
0.0021151914,
0.0251635127,
0.0265921149,
-0.0211420394,
0.0023266238,
0.0007671905,
0.0002636329,
0.0022411644,
0.044602368,
-0.0317875519,
-0.0063103675,
-0.0168534648,
-0.0207721181,
0.0553469397,
0.0342907682,
-0.0149974022,
-0.0440954752,
-0.047247082,
-0.0091587007,
0.0358997211,
-0.016823465,
0.0199389383,
0.0109391548,
0.0438996553,
-0.00754468,
-0.0091107339,
0.0015109248,
0.0216656215,
-0.001667082,
0.0071863085,
0.0279394593,
0.0398431979,
-0.0002009386,
-0.020330172,
0.1015090495,
-0.0128140915,
0.0061309771,
-0.0043024826,
-0.0849893391,
-0.0080910083,
0.0285247397,
-0.0402648747,
0.0168015566,
-0.0361864865,
-0.0177853908,
0.0803018883,
-0.0817783922,
0.0070340256,
-0.0055853985,
2.77e-7,
0.0980063602,
0.0140531221,
0.0346084535,
0.0002089333,
0.039024096,
0.0241613518,
-0.0201178007,
0.0294935387,
-0.0292064585,
-0.0187058132,
0.0347667187,
0.0143646486,
0.0037724343,
0.0584580153,
0.0061210091,
-0.0216478594,
-0.0319178402,
-0.0316276699,
-0.0415124223,
0.0017652657,
0.084680602,
0.0377619304,
-0.0322449468,
-0.0550641492,
0.0474166758,
-0.0454344489,
0.0500129275,
-0.057195697,
-0.0012230468,
-0.0216067247,
0.0102681685,
0.0010165615,
-0.0216244943,
0.0208228957,
-0.041994404,
-0.0390408412,
-0.0150278527,
0.0519811325,
-0.0629599988,
0.0033812225,
0.0099309031,
-0.0273562912,
0.0004864874,
0.0162957367,
0.0633257255,
-0.0026908577,
-0.0393891335,
-0.0069447258,
0.0232858676,
0.0086847134,
0.0596703626,
-0.0439906344,
-0.0693724081,
0.0479609855,
0.0418359004,
0.017216824,
-0.0279728957,
-0.1092474833,
0.0007052834,
0.0444282331,
-0.025199417,
-0.0229571294,
0.0563503057,
-0.020885136,
-0.0238717161,
-0.0669476241,
-0.0317846052,
2.291901055e-34,
0.0311623737,
0.0197762679,
-0.0004115244,
-0.0013181552,
0.0368265919,
-0.0355482586,
-0.0026705901,
-0.0550091863,
0.0003365469,
-0.0419522561,
-0.007310024
] |
http://arxiv.org/abs/2405.15765v1
|
http://arxiv.org/pdf/2405.15765v1
|
Scaling Laws for Discriminative Classification in Large Language Models
| 2024-05-24T17:58:38 | 2024-05-24T17:58:38 |
Dean Wyatte, Fatemeh Tahmasbi, Ming Li, Thomas Markovich
|
cs.CL
|
cs.CL, cs.LG
|
Modern large language models (LLMs) represent a paradigm shift in what can
plausibly be expected of machine learning models. The fact that LLMs can
effectively generate sensible answers to a diverse range of queries suggests
that they would be useful in customer support applications. While powerful,
LLMs have been observed to be prone to hallucination which unfortunately makes
their near term use in customer support applications challenging. To address
this issue we present a system that allows us to use an LLM to augment our
customer support advocates by re-framing the language modeling task as a
discriminative classification task. In this framing, we seek to present the
top-K best template responses for a customer support advocate to use when
responding to a customer. We present the result of both offline and online
experiments where we observed offline gains and statistically significant
online lifts for our experimental system. Along the way, we present observed
scaling curves for validation loss and top-K accuracy, resulted from model
parameter ablation studies. We close by discussing the space of trade-offs with
respect to model size, latency, and accuracy as well as and suggesting future
applications to explore.
|
Scaling Laws for Discriminative Classification in Large Language Models: Modern large language models (LLMs) represent a paradigm shift in what can
plausibly be expected of machine learning models. The fact that LLMs can
effectively generate sensible answers to a diverse range of queries suggests
that they would be useful in customer support applications. While powerful,
LLMs have been observed to be prone to hallucination which unfortunately makes
their near term use in customer support applications challenging. To address
this issue we present a system that allows us to use an LLM to augment our
customer support advocates by re-framing the language modeling task as a
discriminative classification task. In this framing, we seek to present the
top-K best template responses for a customer support advocate to use when
responding to a customer. We present the result of both offline and online
experiments where we observed offline gains and statistically significant
online lifts for our experimental system. Along the way, we present observed
scaling curves for validation loss and top-K accuracy, resulted from model
parameter ablation studies. We close by discussing the space of trade-offs with
respect to model size, latency, and accuracy as well as and suggesting future
applications to explore.
|
[
0.0508162379,
0.0115070958,
-0.0177342556,
0.0320819542,
-0.0171918515,
0.037815541,
0.0374466069,
0.0393416099,
0.0053737042,
-0.078786701,
0.0039755255,
-0.0898316801,
-0.0148601923,
0.0252939574,
-0.015638452,
-0.0411876775,
-0.0073112389,
-0.031919606,
0.0404172502,
-0.0154776741,
0.0032171155,
0.010300952,
0.0057427832,
0.0234792028,
-0.0090794628,
-0.0205853153,
0.0312370453,
-0.0033166502,
-0.0220122598,
0.0020187313,
0.0481058285,
0.0430777334,
0.0029762043,
0.0568044074,
0.0000019337,
-0.071867317,
-0.0024849719,
0.0153528359,
-0.0839433149,
0.0417893194,
0.0340611227,
-0.0229231808,
0.0165551901,
0.033786241,
-0.0178418346,
0.0055459649,
0.0511651374,
0.063613236,
0.0403183773,
0.0737794712,
-0.0049587167,
-0.0683069974,
0.0191336516,
0.0040560495,
-0.021676613,
-0.0317902602,
0.0567370318,
0.022818584,
-0.0146889798,
-0.0198300779,
-0.0029859536,
0.0401307717,
-0.0125187403,
0.0093624787,
0.0002324963,
0.0378317535,
-0.0291657634,
-0.0007725383,
0.006718433,
0.0437794961,
-0.00044043,
-0.0216857772,
-0.0158423223,
-0.009231301,
-0.005886422,
0.0069329618,
-0.0321450904,
0.0044229226,
-0.0078586219,
-0.0104337772,
-0.0209605768,
-0.0137515161,
0.0286627933,
0.0207295232,
-0.0444496013,
0.0327020288,
-0.0087627834,
-0.0471126661,
0.0022427056,
0.0014570202,
-0.0432226285,
-0.0391604528,
0.024008872,
0.0136985201,
-0.0164803434,
-0.0090813115,
0.0198126882,
-0.0297424812,
0.0381071791,
-0.0621704012,
0.0718872398,
-0.0039530215,
-0.0050404873,
0.0454545431,
-0.0354017243,
0.0561845005,
-0.0088312579,
0.0044151759,
-0.039246697,
0.0097867884,
-0.0735945627,
-0.0319927558,
-0.0372266769,
0.0333904922,
-0.0058281724,
-0.0093159368,
-0.0509776138,
0.025688,
0.0053325831,
0.008353699,
0.0351007953,
0.001721209,
-0.0196374338,
-0.013767181,
-0.0283447243,
0.0449044779,
-0.0438485593,
0.0044052084,
-0.0335649103,
-0.0721653104,
0.0385250486,
0.027863048,
-0.0095960107,
-0.010643417,
0.0308094975,
0.0275180023,
-0.0113081755,
-0.0581027679,
-0.0285450388,
0.0002379261,
-0.0525316969,
-0.0528158769,
-0.0412368663,
-0.0379784256,
-0.0007883049,
0.0150209544,
-0.0065126256,
-0.0210757442,
-0.0060756272,
0.00017265,
-0.0231602881,
0.0917250142,
-0.0105411755,
0.0352654196,
0.0692949668,
-0.0124624493,
0.0042286962,
0.0607433654,
0.0199250486,
0.0173666272,
0.0813318565,
-0.030548444,
0.0537445769,
0.0178386588,
-0.0018289713,
-0.0051163496,
-0.0545001552,
0.021430403,
0.0122905541,
0.0445637219,
0.0048653642,
0.04286208,
-0.0184835568,
-0.0420446396,
0.0544782951,
0.0491320789,
0.0233625881,
0.0508936271,
-0.0037658007,
0.0329407267,
-0.0268597174,
-0.0270920172,
-0.0279834867,
-0.0452261604,
-0.0255165789,
0.0098465569,
0.0113813197,
0.0053655664,
-0.0453311838,
-0.0322089642,
-0.0609256849,
-0.0034504193,
-0.0296383165,
0.0425685756,
0.0846773237,
-0.0532898232,
-0.0001945248,
0.0621384904,
0.0064614383,
-0.0049087242,
-0.0186302271,
-0.0692087486,
0.0055513834,
0.0678815544,
0.0423883274,
-0.0383861996,
-0.0026689232,
-0.0044054878,
0.0121038398,
0.0152303539,
0.0649991035,
-0.013925585,
-0.0138309775,
-0.0537403673,
0.0022726182,
0.0486706942,
0.0098974863,
-0.0003286448,
0.0277465712,
-0.0077219298,
0.0348978378,
-0.0312432237,
-0.0570506454,
0.0295438319,
0.0065479721,
-0.0402976274,
0.0277975556,
0.0065886346,
0.0342316702,
0.0276987627,
-0.026239641,
-0.0107853878,
0.0355712846,
-0.0300090294,
-0.0157911461,
-0.0328054167,
0.0045144502,
0.01337492,
0.0402304307,
0.0308300629,
0.0262415931,
0.0209645126,
-0.0155653078,
-0.0136049176,
-0.0226509292,
-0.0315275863,
0.0544856004,
-0.0529091805,
0.0238411743,
-0.0350495204,
0.0279248543,
-0.0437497497,
0.0564111397,
0.0116070434,
0.0433323793,
-0.0384092815,
-0.0473806225,
0.0009788928,
-0.0518226847,
-0.0214597415,
0.0379247628,
-0.0137044135,
0.0344737321,
0.0105211968,
-0.0270032492,
-0.0210345481,
-0.0073885275,
-0.0851439834,
-0.0360166952,
-0.0292090233,
0.0261933263,
-0.0258584172,
0.0023905113,
-0.0332013518,
-0.0305622201,
-0.0178252868,
-0.035465952,
0.0188519433,
0.0595985726,
0.0091496203,
0.0051681823,
-0.0572497323,
0.0015744448,
-0.0373758674,
0.0093608359,
0.0483288765,
0.005210177,
-0.0526965782,
-0.0881082267,
-0.001703616,
-0.0447468981,
0.0207914617,
-0.0029183968,
-0.0102791963,
-0.0220613088,
-0.015644243,
-0.0124471299,
0.0587632805,
0.0233606789,
0.0108140167,
0.0272863172,
0.0166274868,
-0.0520972051,
-0.01170466,
-0.0357027762,
-0.0525051057,
-0.0293903947,
-0.0325021893,
0.0114842355,
0.0593467839,
-0.0147539452,
0.0146219395,
-0.064862594,
-0.0152536696,
-0.0127658388,
-0.0057044406,
-0.0720505714,
0.0097699855,
-0.0317324661,
0.0582945608,
0.0360610038,
-0.0465357192,
0.0606661588,
-0.01605995,
0.0050930479,
-0.0358750112,
-0.046972461,
-0.0220193975,
0.0493842326,
-0.02342261,
0.0299357232,
0.0180168338,
-0.0485023893,
0.0091743646,
-0.0489762276,
0.0159517732,
-0.0616391972,
0.0312366243,
-0.0534633175,
0.0431897938,
-0.0125469714,
-0.0523830988,
-0.0249123182,
-0.046971418,
-0.0562847629,
-0.002628576,
0.0087691974,
-0.0124796368,
-0.039350491,
0.0569156781,
-0.0209599342,
0.0457507446,
-0.0089078378,
-0.0220767632,
0.0211583413,
0.0305559412,
0.0118473023,
0.044288069,
0.0370115153,
-0.0298579223,
0.0204189084,
0.02477712,
0.0435401686,
0.0120446822,
-0.0119567718,
-0.0317842625,
0.0468742028,
-0.0508123301,
-0.0252407268,
0.0295021776,
0.0077618184,
-0.0223872755,
-0.009003222,
0.1081938669,
0.0231068544,
-0.0460528396,
0.0009100711,
-0.010391511,
0.0249536932,
0.0320106372,
0.0479458198,
-0.0576702319,
-0.0037838279,
0.0045377305,
0.0612328053,
-0.0087661985,
0.0068026497,
-0.0100473426,
-0.0896214619,
0.0571340472,
-0.0040208986,
-0.0135782547,
-0.0272890367,
-0.0579065606,
-0.0088689793,
-0.003656294,
0.0380311124,
0.0122894645,
-0.0191847458,
-0.0406392254,
0.0545162298,
-0.0368862152,
-0.0787782297,
0.0313093327,
-0.0270662121,
0.0284477584,
0.045924779,
0.0098548671,
0.0213957578,
-0.0440712012,
0.037103191,
-0.0101892781,
0.0155742606,
-0.0148659553,
0.0481577702,
0.0126225939,
0.0101069594,
0.0260309745,
-0.0081902584,
-0.0507698879,
0.042306032,
0.0182917658,
0.0504590571,
-0.0046112691,
-0.0247949027,
0.0101593705,
-0.0085727526,
-0.0143014276,
0.0776881352,
0.0083410097,
-0.0484898575,
-0.0558314174,
-0.0058962763,
-0.0168137718,
0.0550540313,
-0.0334444642,
0.014666127,
-0.0275978427,
0.0061544934,
-0.0503120162,
-0.012110699,
0.0060691317,
0.0528565608,
-0.0032533796,
0.0037782246,
-0.0379559249,
0.043898467,
-0.0190662052,
0.0242599305,
-0.0586950928,
0.0373550393,
-0.1089614108,
0.0556910895,
0.0375055261,
-0.0093149208,
-0.0319925323,
-0.0376483202,
-0.0422512814,
-0.0704108253,
0.0082381917,
0.0035029773,
0.0376373082,
0.0247611236,
0.0025379232,
-0.1239719987,
0.0318003744,
0.0669020489,
-0.0300814975,
0.0139348432,
0.0119792093,
-0.0101838047,
-0.0427337252,
0.0352582112,
-0.0414964817,
0.0287731271,
-0.0093189878,
0.0060735722,
-0.025644388,
-0.0201885663,
-0.0098175518,
0.0383739583,
0.0231710281,
0.0089876251,
-0.0543037727,
0.0024485115,
0.0452231355,
0.0161968991,
-0.0057689506,
-0.0002707379,
-0.0093460521,
0.0371927656,
-0.0233903881,
-0.0055779265,
-0.0181702357,
-0.0153266285,
0.0108990213,
-0.0582711212,
-0.0172105059,
-0.0011396499,
0.0133934775,
0.0498309955,
-0.0116187185,
-0.0404696949,
0.0190379098,
0.0348896384,
0.0433817394,
-0.0114667071,
0.0684064105,
-0.0064170598,
0.0237708595,
-0.0479732342,
0.0006820737,
-0.0470503531,
-0.0938343704,
-0.0346094519,
0.0241355598,
0.014582037,
-0.0178401452,
0.0202843454,
-0.0351154394,
-0.0252960194,
-0.0613971427,
-0.0074508209,
0.0114329308,
-0.0026146406,
0.0291319657,
0.0331585556,
-0.0153339952,
0.0387609452,
0.0677761659,
0.0313187949,
0.0106395129,
-0.0261474494,
-0.0355793275,
-0.0543057956,
-0.0368294865,
-0.0014293302,
0.0220354851,
-0.0171319563,
-0.0182859581,
-0.0870520696,
-0.0551733077,
0.0006061451,
-0.034402851,
0.0134012923,
-0.022020435,
0.032577429,
0.0776785091,
0.0136255855,
-0.005941153,
-0.0374449007,
0.0211868472,
-0.0163487792,
0.0177405812,
-0.0349824838,
-6.28293649e-33,
0.0082243374,
-0.0078138513,
-0.0067734192,
-0.015839003,
0.0075242892,
0.0494140871,
0.0428395607,
0.0015225888,
0.0086680986,
-0.0082843164,
-0.0285308287,
0.0217354782,
0.0286146887,
0.0301874559,
-0.0079794144,
-0.0297439694,
0.0588890128,
-0.0288474262,
-0.0139667336,
0.034285143,
0.0454213358,
0.0116689587,
0.0271821842,
-0.0920748338,
0.0332351886,
-0.0073898807,
0.0016904582,
0.0545826703,
0.0105378143,
0.0681002364,
-0.0467496216,
0.0898461342,
0.0306471679,
-0.0475974679,
0.0255037453,
0.011063247,
-0.0561918132,
0.02463625,
0.0028084582,
0.017227089,
0.016887296,
0.0157823153,
0.0335314125,
-0.013096191,
-0.0224070363,
0.0159259252,
0.0338897146,
-0.0384947434,
-0.0596355684,
-0.0451159962,
-0.0135046635,
0.0327258781,
-0.0094002308,
-0.0085826013,
0.0595646352,
0.0468016677,
0.0023646725,
0.0466251895,
-0.0373258963,
0.0147578605,
0.0457467809,
0.013366038,
0.0554144941,
0.0391063727,
-0.025802467,
0.0387729555,
-0.0188340172,
-0.0004473477,
0.008317071,
-0.0577990375,
0.0114212856,
0.0237228945,
0.0302991588,
0.0444520749,
0.0228947513,
-0.0204403941,
-0.0226645842,
0.0034267108,
0.0607689396,
0.0103938868,
0.0603433549,
0.0123252803,
-0.0868700519,
-0.0610896647,
-0.014644417,
-0.0056894929,
-0.0527728423,
-0.0003782218,
0.0347775891,
0.0365954489,
-0.0296875555,
-0.007024874,
0.0136544919,
-0.0406743772,
-0.0524438471,
-0.0021621631,
0.0383908525,
0.0273096953,
0.0070083672,
-0.0231602248,
-0.051461149,
-0.0273600966,
-0.0381458737,
-0.0270236079,
-0.0079610627,
-0.0349310711,
0.0139172226,
-0.016856363,
-0.0645023063,
0.0095035639,
0.0100663872,
-0.0118286479,
0.0236325152,
0.0512775034,
0.022616623,
0.0032506499,
0.0181233864,
0.0719955638,
0.0379813388,
-0.0118061835,
0.0058670477,
0.0033350685,
0.0030308247,
0.024786327,
-0.0431915708,
0.0114723081,
0.0057566259,
-0.0087350449,
0.0533642322,
-0.035375651,
-0.013375015,
0.0092396168,
2.645e-7,
0.0370357297,
0.0755847469,
0.0052657411,
-0.0276380908,
0.0268997326,
0.0450789593,
0.0366526768,
-0.0089524044,
0.0404452048,
-0.0075287111,
0.025742745,
-0.0116039803,
0.0648122802,
0.0152382916,
-0.0713043287,
-0.0120912297,
-0.0797044411,
0.0361042768,
-0.0642547458,
0.009597268,
0.0319383591,
0.0946899727,
0.0030002997,
-0.0418835543,
-0.0256569777,
-0.0300858077,
0.0529740043,
-0.0710323676,
0.040442273,
0.0285334028,
0.0696866512,
0.0001330666,
0.0348509438,
0.0562743656,
0.0134060597,
0.0325726382,
0.0250642616,
0.0701488927,
0.002455666,
0.015711803,
-0.0161132943,
-0.0067040697,
0.0156261064,
-0.0902853832,
0.0574323162,
-0.0565582886,
-0.0157493986,
-0.0305115506,
0.0051592807,
-0.0698384866,
0.0378308184,
-0.0216621757,
-0.0037152597,
0.0407103114,
0.0109646199,
-0.0318269394,
0.003108789,
-0.0907877088,
0.0174614526,
0.0463853963,
-0.0395289287,
-0.0244221259,
-0.0030709009,
-0.0047353473,
0.0075973999,
-0.0327752046,
0.0267943945,
2.549739312e-34,
0.0465390608,
0.0161507335,
0.0106707206,
0.0532686822,
0.0173845235,
0.0063614543,
0.0322351754,
0.0108564831,
0.0168721769,
-0.1271323413,
0.0183836091
] |
http://arxiv.org/abs/2405.15760v1
|
http://arxiv.org/pdf/2405.15760v1
|
GPT is Not an Annotator: The Necessity of Human Annotation in Fairness Benchmark Construction
| 2024-05-24T17:56:03 | 2024-05-24T17:56:03 |
Virginia K. Felkner, Jennifer A. Thompson, Jonathan May
|
cs.CL
|
cs.CL, cs.CY, I.2.7; K.4.2
|
Social biases in LLMs are usually measured via bias benchmark datasets.
Current benchmarks have limitations in scope, grounding, quality, and human
effort required. Previous work has shown success with a community-sourced,
rather than crowd-sourced, approach to benchmark development. However, this
work still required considerable effort from annotators with relevant lived
experience. This paper explores whether an LLM (specifically, GPT-3.5-Turbo)
can assist with the task of developing a bias benchmark dataset from responses
to an open-ended community survey. We also extend the previous work to a new
community and set of biases: the Jewish community and antisemitism. Our
analysis shows that GPT-3.5-Turbo has poor performance on this annotation task
and produces unacceptable quality issues in its output. Thus, we conclude that
GPT-3.5-Turbo is not an appropriate substitute for human annotation in
sensitive tasks related to social biases, and that its use actually negates
many of the benefits of community-sourcing bias benchmarks.
|
GPT is Not an Annotator: The Necessity of Human Annotation in Fairness Benchmark Construction: Social biases in LLMs are usually measured via bias benchmark datasets.
Current benchmarks have limitations in scope, grounding, quality, and human
effort required. Previous work has shown success with a community-sourced,
rather than crowd-sourced, approach to benchmark development. However, this
work still required considerable effort from annotators with relevant lived
experience. This paper explores whether an LLM (specifically, GPT-3.5-Turbo)
can assist with the task of developing a bias benchmark dataset from responses
to an open-ended community survey. We also extend the previous work to a new
community and set of biases: the Jewish community and antisemitism. Our
analysis shows that GPT-3.5-Turbo has poor performance on this annotation task
and produces unacceptable quality issues in its output. Thus, we conclude that
GPT-3.5-Turbo is not an appropriate substitute for human annotation in
sensitive tasks related to social biases, and that its use actually negates
many of the benefits of community-sourcing bias benchmarks.
|
[
0.0097749,
0.0571751408,
-0.0340454355,
0.0427736863,
-0.0136752995,
0.0454139709,
0.0421863012,
0.000015605,
0.0266595464,
-0.0443254188,
0.0092264088,
0.0278009605,
-0.0088052414,
-0.0706220567,
0.0108329225,
-0.0334316939,
0.0302242674,
0.0777835026,
0.012365167,
-0.0192180965,
0.0448887758,
0.022385193,
0.021622194,
0.0506598428,
-0.0509993248,
0.0194499772,
0.051155258,
0.0033414827,
0.0779361799,
-0.0222241245,
0.0103054726,
0.0426530764,
-0.0342561342,
0.0256549492,
0.0000018672,
-0.0456820615,
-0.0108092735,
0.0254203733,
-0.0021930449,
-0.0119811995,
0.0486799702,
0.0068593943,
-0.004445049,
0.0257749297,
-0.0401047133,
0.0144583704,
0.0385270454,
-0.0351026691,
0.0439281128,
0.0255524814,
-0.0008486721,
0.0089388136,
-0.0450858735,
-0.0026721728,
0.0346836187,
0.0222713761,
0.0196532086,
-0.0240513664,
-0.0661008954,
-0.0400021225,
-0.0100578079,
-0.005153155,
0.0235347282,
0.0325827077,
0.0261974465,
-0.0249205008,
-0.0517039783,
-0.0238873307,
0.0158544518,
0.0414675362,
0.0527063757,
0.0401917212,
0.0391466022,
-0.0129326899,
0.0242209509,
0.014945562,
-0.019878041,
0.0177028421,
-0.041864641,
0.0157908741,
-0.0194950886,
0.0262938384,
0.030378921,
0.0129279559,
0.0263242405,
0.0831913874,
0.0079043908,
0.0218629148,
0.0102752689,
0.0358214602,
0.0241780914,
-0.0256315377,
0.0247026943,
-0.0170348566,
0.000019521,
0.008119856,
0.0075151715,
0.0043450743,
0.0590616502,
-0.0444702283,
0.0377384871,
0.0646295026,
-0.029396262,
0.0130606517,
0.047404401,
-0.0050690044,
0.0000131315,
0.0180851109,
0.0092319883,
0.0161517076,
-0.0148091959,
0.0142304208,
0.0364662297,
0.0164931491,
-0.0184421465,
0.1008848175,
-0.0650874972,
-0.0115064122,
-0.0578109436,
0.0285924654,
0.0044642962,
-0.04243806,
-0.0024204114,
0.0211161748,
-0.0457259119,
-0.00596717,
-0.1049485654,
0.0170122571,
-0.0349314325,
-0.1025484279,
0.0215897243,
0.0326665528,
-0.0073662931,
-0.0448940843,
0.0010399688,
0.0321393982,
0.0533538312,
-0.014832926,
-0.0370381251,
-0.0129934233,
-0.0063155647,
-0.0399692841,
0.0147947641,
-0.0420161337,
0.0496257395,
0.0081048133,
-0.0017505507,
-0.0407430083,
-0.0223924909,
0.027359264,
-0.0136844469,
0.069894433,
0.0062719821,
-0.0137441568,
0.0502477661,
0.0567706227,
-0.0425374284,
0.0614434518,
0.0170302782,
-0.0232758801,
0.0417917147,
-0.0244861282,
0.0099530667,
0.0212839898,
0.0223128647,
0.0301170088,
-0.0258796234,
0.0517396741,
-0.0253485497,
-0.0106360177,
0.0332599916,
0.0552880764,
-0.0280484091,
-0.056579195,
0.0226088818,
0.0743370652,
-0.0109165153,
0.0326546319,
-0.0483059511,
0.0112097031,
-0.0155016892,
0.0039953529,
-0.0965410694,
0.0280978978,
-0.0143183023,
-0.0378155224,
0.001492364,
0.0492240563,
-0.0215418339,
-0.0180593859,
-0.0260914881,
0.0369327627,
0.0358964093,
-0.0128195938,
-0.0171688125,
-0.0318383314,
0.0002627452,
0.0195091553,
0.0229053423,
-0.0264085252,
0.0239956193,
-0.0561151393,
0.0325513743,
-0.0073528122,
-0.0148188388,
-0.0534828044,
0.0149212927,
-0.0440346226,
-0.0172245707,
0.045596756,
0.073717773,
-0.0071966313,
-0.0443814471,
-0.0224747341,
-0.0213745479,
0.0210158732,
0.0160304066,
0.0653244853,
0.0465781018,
0.0082058953,
0.0106022432,
-0.0412892923,
0.0048890654,
0.0394567847,
0.0324260183,
0.020143427,
-0.0304459017,
-0.0215029605,
0.0476609878,
0.0293576121,
-0.012966875,
-0.0286037251,
-0.0404243171,
0.0133349244,
-0.0028655468,
-0.008389947,
-0.0322247632,
0.0232187565,
0.0164068975,
-0.090577282,
-0.0300346892,
0.0602701828,
-0.0238631144,
0.0432860032,
0.0140930004,
0.0327277742,
-0.0092025092,
-0.0348683,
-0.0384413823,
-0.0448281877,
-0.0090417815,
-0.0362866819,
0.0418658257,
-0.0090829674,
-0.0033956044,
-0.0034333533,
-0.06781587,
0.037345212,
-0.0212244503,
-0.0366958678,
0.0040437537,
-0.0098019363,
0.0011474476,
0.0185145829,
-0.0267298538,
-0.0388796255,
0.0127116283,
-0.0870108381,
0.0073384321,
-0.0073362915,
0.0023382825,
-0.0492334813,
-0.0068953945,
-0.002122825,
-0.0100178644,
-0.0447764471,
-0.0102183241,
0.0165370163,
0.042900566,
0.017063044,
-0.02578732,
0.002951972,
-0.062647149,
-0.0119094616,
-0.0434926338,
0.0127846524,
0.0702116415,
-0.0193570331,
0.006341666,
-0.0535226502,
-0.0712722018,
-0.0048751603,
-0.0081610624,
-0.0085451258,
0.0485695004,
0.0740963966,
-0.006909214,
0.0224596467,
0.0397731848,
0.0291159041,
0.0235515237,
0.0006477394,
-0.0275823027,
-0.0149866156,
-0.0036705877,
0.0079526622,
-0.0198629741,
-0.0606543906,
-0.0325093679,
0.0285054371,
0.0445629247,
-0.0046568001,
-0.0665713921,
0.0669894516,
-0.0876980275,
-0.0433594882,
-0.0403756872,
-0.0533490479,
-0.0307413191,
0.0122776367,
-0.0458906814,
-0.0304701682,
0.0159579292,
0.0300917737,
0.0228862427,
0.0170897469,
0.0259972811,
-0.0501984283,
0.0391627774,
-0.0032770829,
0.0060615502,
-0.0730781704,
0.0100498516,
-0.0369203836,
-0.0258075036,
0.0614292845,
-0.0651538298,
0.0027901705,
-0.004244437,
0.0004305588,
-0.0660059378,
-0.0518604144,
-0.0385373309,
-0.0275333282,
-0.028668493,
-0.0171158928,
0.0267461147,
0.0224588923,
0.0331114084,
0.0116078528,
-0.0112928804,
0.0095823808,
0.0372427963,
-0.0077412091,
0.0154962214,
-0.0649481267,
-0.0473793335,
0.0198599324,
0.0625982508,
-0.0044233608,
-0.0342537872,
0.0501283556,
0.0400117375,
-0.0108314417,
-0.0146216853,
0.018454928,
0.0139671909,
-0.0377402939,
0.0041489936,
-0.0064807395,
0.0262625832,
-0.0040781968,
0.0551801883,
0.0836599842,
-0.0058133025,
-0.060740035,
-0.0082006911,
-0.0019977298,
-0.0063710618,
0.0431777388,
-0.0120995697,
-0.1702666432,
0.0269928761,
0.0055800234,
0.0562602989,
-0.0917634889,
-0.0964457244,
0.023564687,
-0.0738674104,
0.0505180322,
0.0152878799,
-0.0477653444,
0.0070217033,
-0.0339267254,
0.0295259468,
-0.041716814,
0.0119995959,
-0.0170501787,
0.0672114789,
0.0251101702,
0.0087295612,
-0.0741703585,
-0.0097360322,
0.0263117347,
-0.0591270886,
0.0766226277,
0.0229270756,
0.0429523773,
-0.0113328183,
-0.002231349,
0.0490262397,
-0.0242823511,
-0.0420872308,
-0.0123795252,
0.043503657,
-0.0109917717,
-0.0328209326,
-0.0181554221,
0.0257769767,
-0.0002386328,
0.0174168982,
0.0286218729,
0.0221332684,
0.0276224948,
0.0009209759,
-0.0078095291,
-0.0006332647,
-0.0274577197,
0.0897687003,
0.0189411882,
0.0080074826,
0.0305828825,
-0.0586795099,
-0.0062073031,
-0.0186137706,
-0.0836618096,
0.0195249636,
0.0236442443,
-0.0089538405,
-0.0157779902,
-0.0535116047,
0.0069936425,
0.0248051733,
0.0233746078,
-0.0666471422,
0.001006116,
-0.0112169217,
-0.054970827,
0.0230001248,
-0.0226760246,
0.0199243911,
0.0436405018,
-0.0121322088,
-0.043128714,
0.0015086608,
-0.0254100095,
0.0166658722,
-0.0379993655,
-0.0128252357,
0.090030767,
0.0498587787,
0.0324109383,
0.0156538654,
0.0047570998,
-0.0101211928,
0.0008437107,
0.056951765,
-0.0277723633,
0.0275649186,
0.0031961459,
-0.005398632,
-0.0107389875,
-0.0271880124,
-0.0166146681,
0.0375719704,
0.0480376333,
-0.0327314697,
-0.030312242,
-0.018795548,
0.0013212626,
-0.0124496361,
0.0454396643,
-0.021736661,
-0.0134521136,
-0.0389573053,
-0.0284474473,
-0.0006733969,
-0.0270148385,
-0.0214946903,
-0.0745044872,
0.0214827973,
0.004346035,
-0.0627730489,
0.0090586459,
-0.0510971472,
0.0269033834,
-0.0703796893,
0.0588946454,
0.036433652,
0.0372599438,
-0.0394945368,
-0.0086282687,
-0.0108212149,
0.0178327952,
-0.006818437,
0.0492031164,
0.0190282539,
0.0092958771,
0.0140069984,
0.0091888504,
0.0627999157,
-0.0195062831,
-0.0204473473,
-0.001820433,
-0.0385080464,
0.013492614,
-0.0199695565,
0.0203374065,
0.0036113134,
0.0192507133,
-0.0285756793,
-0.0085158031,
-0.0144862505,
-0.0041628834,
-0.0000268309,
0.0400380827,
-0.0036746943,
-0.0154430633,
-0.0213347543,
-0.0023687184,
-0.0082467087,
0.0431842804,
-0.031391155,
-0.0395356305,
-0.0195618067,
-0.0521670878,
0.0040212292,
-0.0063205496,
-0.0062875194,
-0.0607172661,
-0.1149087176,
-0.036491394,
-0.0169124436,
0.0204334557,
0.048986569,
-0.0113606611,
0.0027604166,
0.0240220055,
-0.0330941118,
-0.0173954163,
0.0061419592,
0.0012223814,
-0.0704429075,
-0.0103458781,
0.032391049,
-6.096339251e-33,
-0.0413650982,
-0.0381530859,
0.0056175203,
0.0128951604,
-0.0275022779,
-0.0080424501,
0.025747722,
0.0154613303,
-0.045254685,
-0.0035660395,
0.0067021493,
0.0336419828,
0.018307019,
-0.0230548959,
0.0106797535,
-0.0637506396,
0.0328632444,
-0.0064064455,
-0.0376436189,
-0.0212033838,
-0.0751011223,
-0.0216645189,
0.0388622954,
-0.0664168373,
0.0564602464,
-0.0082735922,
-0.0170643814,
0.0097501166,
0.0202187672,
0.0273854025,
-0.0091112843,
0.0562314056,
0.0344463512,
-0.0181435458,
0.0294770505,
0.0258046445,
-0.0403818004,
-0.0206995979,
0.0361058004,
0.0011621766,
0.0014033475,
-0.0125439493,
0.0019393981,
0.0245882124,
-0.0809024498,
-0.0128848096,
0.0570801012,
-0.0249268766,
0.0412019305,
0.0019817555,
-0.0266162902,
0.0427694246,
-0.0333100893,
0.0697415024,
-0.0296575036,
0.014495872,
0.0376821272,
0.0589347742,
-0.0517194867,
-0.0065339752,
-0.021295879,
-0.0495789796,
0.0260800067,
0.0301061012,
-0.0037396194,
0.0549471416,
0.042457357,
0.0808599442,
-0.0202490687,
-0.0472890139,
0.0222270861,
-0.0214899927,
0.0575322323,
0.1461219043,
0.0075499699,
-0.0021644563,
-0.0234731697,
0.0299498811,
0.0690594465,
-0.0259948708,
0.0088646775,
-0.0021944947,
-0.0538016669,
-0.0192573965,
-0.0578031577,
-0.0934821069,
-0.0356165357,
0.0093352599,
0.0570642054,
0.0056180237,
-0.0207154397,
0.0643348843,
0.0022086906,
-0.0028094291,
0.0319517106,
0.0045612981,
-0.0076666949,
-0.0148374196,
0.0088707814,
0.0001094501,
-0.0306224581,
-0.000577433,
-0.0227860678,
0.0340237729,
0.0643459857,
-0.0089965649,
0.0040145996,
-0.0038886124,
-0.057352487,
-0.0016098286,
0.0038360693,
0.02903709,
-0.0234012324,
-0.0366748646,
0.0158744883,
0.013450956,
0.0230642147,
-0.0037161137,
0.0318066031,
-0.0016811355,
-0.0299975388,
-0.0126692168,
-0.0099963518,
0.0061095827,
-0.0028300507,
-0.0297498722,
0.0754084736,
0.0252843387,
0.0260973684,
-0.0420667492,
0.0031248233,
-0.054095272,
2.434e-7,
0.0351383388,
0.0137329716,
-0.0133566419,
-0.0134605523,
-0.017851023,
0.0015527979,
0.0372311249,
0.0469400547,
0.0334484316,
-0.0299397726,
0.0096833482,
0.0349703021,
0.0473806299,
0.0396360233,
0.0029281531,
-0.0481151976,
-0.0302228369,
0.0319225974,
-0.0534924045,
0.0430283882,
-0.0073397136,
0.02449077,
0.0623901486,
0.0087825507,
-0.0203363374,
0.0125401234,
0.0192502905,
-0.0707841888,
-0.0113934586,
0.0091417208,
0.0488690138,
0.0086972984,
0.0368723907,
0.0399892144,
0.0002681335,
-0.0036235135,
0.0228875801,
0.0351868421,
0.0321870074,
-0.0013621396,
-0.0239792652,
0.0002984493,
0.0047982591,
-0.054192625,
0.028627798,
0.0206445102,
-0.0081538288,
-0.0348816961,
-0.0570730045,
-0.0469573401,
0.0462108441,
0.0297815371,
0.0001895025,
-0.0172673073,
0.032969065,
0.0236674249,
-0.0077977879,
-0.0434710495,
-0.0128842546,
-0.0285695866,
-0.0308829118,
-0.0934979618,
0.0088489735,
0.0195186324,
-0.0463613868,
0.0306709092,
0.0182656404,
1.830745782e-34,
0.0189182162,
-0.008898492,
0.0196587406,
-0.0492675304,
0.0631517693,
0.0403217636,
-0.038853921,
-0.0293015614,
0.0194369443,
0.0035002357,
0.0184623711
] |
http://arxiv.org/abs/2405.15750v1
|
http://arxiv.org/pdf/2405.15750v1
|
Filtered Corpus Training (FiCT) Shows that Language Models can Generalize from Indirect Evidence
| 2024-05-24T17:47:20 | 2024-05-24T17:47:20 |
Abhinav Patil, Jaap Jumelet, Yu Ying Chiu, Andy Lapastora, Peter Shen, Lexie Wang, Clevis Willrich, Shane Steinert-Threlkeld
|
cs.CL
|
cs.CL, cs.AI, cs.LG
|
This paper introduces Filtered Corpus Training, a method that trains language
models (LMs) on corpora with certain linguistic constructions filtered out from
the training data, and uses it to measure the ability of LMs to perform
linguistic generalization on the basis of indirect evidence. We apply the
method to both LSTM and Transformer LMs (of roughly comparable size),
developing filtered corpora that target a wide range of linguistic phenomena.
Our results show that while transformers are better qua LMs (as measured by
perplexity), both models perform equally and surprisingly well on linguistic
generalization measures, suggesting that they are capable of generalizing from
indirect evidence.
|
Filtered Corpus Training (FiCT) Shows that Language Models can Generalize from Indirect Evidence: This paper introduces Filtered Corpus Training, a method that trains language
models (LMs) on corpora with certain linguistic constructions filtered out from
the training data, and uses it to measure the ability of LMs to perform
linguistic generalization on the basis of indirect evidence. We apply the
method to both LSTM and Transformer LMs (of roughly comparable size),
developing filtered corpora that target a wide range of linguistic phenomena.
Our results show that while transformers are better qua LMs (as measured by
perplexity), both models perform equally and surprisingly well on linguistic
generalization measures, suggesting that they are capable of generalizing from
indirect evidence.
|
[
0.0421425737,
0.0682951808,
-0.009503454,
0.0270935092,
-0.0240041446,
0.0207517724,
-0.0178270806,
0.0291860104,
-0.0319900326,
-0.0590024702,
-0.0423913002,
-0.0584708564,
-0.0163397305,
0.0050975196,
0.0464233831,
-0.0703640804,
0.0366707332,
0.010066947,
-0.0160838123,
0.0036525463,
0.0047517195,
0.0075382758,
0.0175412893,
0.0234269481,
0.0044957632,
0.0276694428,
0.0130310543,
-0.0244305152,
0.0052971886,
0.0083002001,
0.0096440716,
-0.0061770519,
0.0375284366,
0.0429209694,
0.0000018034,
-0.0445261337,
-0.0062800702,
-0.0263116173,
0.0055653783,
-0.0187430009,
0.0496428199,
-0.0250909589,
0.0074925162,
0.0266760066,
-0.0673609897,
-0.0322088525,
0.0539304428,
0.0498990975,
0.0774888843,
0.0879514888,
-0.018123528,
-0.0537949353,
0.0310045611,
-0.0497032478,
0.0116795655,
0.0050438824,
0.0118352035,
-0.0261959452,
-0.0078853276,
-0.0100141801,
-0.0271912571,
0.0249386933,
0.0140720848,
0.0237152092,
-0.0177800264,
0.0259308517,
-0.0184883215,
-0.0296989586,
0.0208730102,
0.0351008624,
-0.0184167977,
-0.0262886379,
-0.0146298436,
0.0095622698,
0.0057497555,
0.0418556258,
-0.0350209065,
0.0346975066,
0.0109857023,
0.0541031286,
0.0282031167,
-0.027391864,
0.0378472023,
0.0279842597,
-0.0638245046,
0.0811001807,
0.0120979911,
-0.0520316586,
0.0265230145,
0.0349612236,
-0.0265305322,
-0.0484121181,
-0.0019191939,
0.0084551796,
-0.0023105489,
0.0304648299,
-0.0009786802,
-0.0234731473,
0.0096830502,
-0.027534252,
0.0048662601,
0.0153062791,
-0.0207925681,
0.0362390317,
-0.0505187809,
0.0263058059,
-0.0464560054,
0.0599499121,
-0.0505555458,
-0.0425130017,
-0.0847700238,
0.0039773914,
-0.0108420877,
0.0332646966,
0.0025262337,
0.0122740716,
-0.0557244308,
0.0091437269,
0.0145336241,
-0.0155786909,
0.0105853947,
0.0134983519,
-0.0764793679,
-0.0117691103,
-0.0342311747,
0.0051629622,
-0.0185985044,
0.0651036873,
-0.0151928775,
-0.0709988922,
-0.0256182253,
0.0302324072,
0.0356669649,
-0.0421497785,
0.0073905746,
0.0311680697,
-0.0049668513,
-0.0643722564,
-0.0548037849,
0.032625597,
-0.0238953922,
-0.0628993884,
0.017474303,
-0.0290364381,
-0.0092806434,
-0.0191592798,
0.0342394151,
-0.0202353727,
-0.0178888105,
0.0104646282,
-0.0300333351,
0.07065624,
0.0027636259,
-0.0068461699,
0.0335117765,
-0.0122614438,
0.0000436854,
0.0609446019,
0.0479658283,
-0.0032111953,
0.0031174896,
-0.0266861003,
0.0021316267,
0.0324042812,
0.0014527186,
0.0244752076,
-0.0327528715,
-0.0206026398,
-0.0503941067,
-0.0038611256,
0.0361820273,
0.0790447295,
-0.0460715219,
0.0031171776,
0.0572743416,
0.0393585339,
0.0625837296,
0.0408261977,
0.0548263974,
-0.0016708635,
0.0059030047,
0.0469102561,
-0.0484393798,
-0.039644517,
0.0161140282,
0.0172482152,
-0.0326946266,
0.0058933049,
-0.0512859412,
-0.0410511047,
-0.0319317244,
-0.051257927,
0.0331276208,
-0.0237297229,
0.0646730736,
0.0667742118,
-0.0030218055,
0.0848881379,
0.0489500761,
-0.0655992851,
0.0092707882,
-0.0647174791,
0.0311378874,
0.0095837405,
0.0332825258,
-0.0354510471,
-0.0188254248,
-0.0292444807,
-0.0278064031,
0.0480508506,
0.0247723516,
0.0151635474,
0.0460603498,
-0.0446639918,
-0.0069542546,
0.0058897599,
0.0269958917,
0.0264820047,
0.0080200117,
-0.0080478443,
0.01305361,
-0.0500301085,
-0.0217113625,
0.0676956698,
-0.0194806103,
-0.0222507771,
0.0343408212,
-0.0015505082,
0.0590158775,
-0.0114274668,
-0.0211941488,
-0.0084005976,
0.0311262961,
-0.0401398912,
-0.0458021238,
0.0493929908,
-0.0513224266,
-0.0458247922,
0.0261646565,
-0.0490835644,
0.0043818187,
0.0759098828,
-0.0507632978,
-0.0365371443,
0.0274935123,
-0.0400985032,
0.0675022751,
-0.0557362773,
0.0264164135,
-0.003957029,
0.0102090351,
-0.0236724541,
0.0489899069,
-0.0058356482,
0.0095375245,
-0.015990993,
-0.0364171229,
0.0214415994,
-0.0307666119,
-0.0293145813,
-0.0015588492,
0.0097904764,
-0.0017624351,
0.0402074233,
-0.0023880196,
-0.0312545262,
-0.0230225641,
-0.0794388503,
-0.0337693244,
-0.0310812779,
0.0012320601,
0.011899678,
0.0053850175,
0.015549385,
-0.0243962277,
-0.0395699404,
-0.0165089518,
0.0106030004,
0.0231639184,
0.0617025569,
0.012066613,
-0.0192643162,
-0.0705051646,
-0.0008367712,
-0.0280770306,
0.08342655,
0.0019515154,
-0.0205643512,
-0.0565599464,
0.0239355024,
-0.0290425941,
0.0295839794,
-0.0093638105,
-0.0454539098,
0.0123347584,
-0.0109632295,
-0.0098190336,
0.0415052772,
0.0262731947,
-0.0338333063,
0.04801175,
0.0121791959,
-0.0584057234,
0.0135723455,
-0.0500423424,
-0.0078176036,
-0.055986613,
-0.0192277916,
0.0183632746,
0.0747806728,
-0.0010553726,
-0.0153853483,
-0.0498592965,
0.0029579552,
-0.0085834516,
0.0018070489,
-0.0041363346,
-0.0226473473,
-0.0432493985,
0.0178738609,
-0.0188490562,
-0.0645225048,
0.0708984956,
-0.031390436,
-0.0157046933,
-0.0661389232,
-0.0358318761,
-0.042911429,
0.0235243198,
0.020420732,
0.0018744067,
0.0343042761,
-0.0254662205,
-0.0525978468,
0.0065235118,
0.0163514391,
-0.0103490371,
0.061155349,
-0.0287502166,
0.0162373818,
0.0152932927,
-0.0334224515,
-0.061803326,
-0.0518518314,
-0.0238226391,
-0.0007197508,
0.0146274138,
-0.0168204904,
-0.0516528301,
0.0084865307,
0.0188178048,
0.0547419898,
0.0192525685,
-0.0183874518,
-0.0281464811,
-0.0005095415,
0.0364603773,
0.0131837623,
-0.0045695617,
-0.0186929833,
0.0263504386,
0.0355420075,
0.0398510806,
-0.0099816527,
-0.0569474734,
0.0108182831,
0.0053584203,
-0.0290129185,
-0.0277668256,
0.0124029499,
0.0149400542,
-0.0139504876,
-0.0196261927,
0.1295312941,
0.0302413069,
-0.0439225361,
-0.002441006,
0.0263445731,
0.024175914,
0.0234997161,
-0.0140634459,
-0.0964386091,
-0.0339695737,
-0.0168770403,
-0.0133653153,
0.0012651815,
-0.0141371461,
-0.0165749006,
-0.090123035,
-0.0030981533,
0.0040258281,
-0.0095608504,
-0.0142510664,
-0.0426863506,
0.0078922752,
-0.0239403788,
0.0169492848,
-0.0316105708,
-0.0219689608,
-0.017541457,
0.0252447538,
-0.0626833811,
-0.0509059951,
-0.0091641862,
-0.0515683107,
-0.0051479526,
0.0566734672,
0.0250462145,
-0.0152574424,
-0.0281204712,
0.0130947949,
-0.0199273136,
0.0240908228,
0.0001480428,
0.0651577562,
-0.0203957856,
-0.0147512853,
0.0434674285,
0.0027295111,
-0.0473901294,
0.0134864999,
-0.0083711529,
0.0413841046,
-0.0158279147,
-0.0342615955,
0.007437964,
-0.0100235548,
-0.0038560873,
0.0749970302,
0.0400048196,
-0.0481448472,
0.0114146154,
-0.0373233631,
-0.0056281355,
0.0149468649,
-0.0370662808,
0.0047858255,
0.0072162398,
-0.0374758653,
-0.1055084392,
-0.0401543789,
-0.0018564744,
0.0052203177,
0.0070170984,
-0.0055507305,
0.0285579488,
-0.0129608279,
-0.0013492048,
0.038312126,
-0.0237396378,
-0.0162727386,
-0.0432791524,
-0.0129773049,
0.0092200264,
-0.0073541501,
-0.0436081514,
0.0223426577,
-0.0832666233,
-0.0818331391,
0.0374375023,
0.0269124378,
0.033252731,
0.0504497997,
-0.0019725498,
-0.0425303727,
0.0406439565,
0.0136585161,
-0.0238421708,
0.0227679629,
0.0762836337,
-0.0361826643,
0.0138120661,
0.0330365449,
-0.034649834,
0.0487616584,
0.0377994813,
0.0181975793,
-0.0591852702,
0.0130433487,
0.0257819574,
-0.0159776248,
0.0826303139,
-0.012987324,
0.0051261936,
-0.0394365191,
-0.0041401535,
0.0063441079,
-0.0032345299,
-0.0474334285,
0.0159944054,
0.0598736964,
0.0122339223,
-0.0288862847,
0.0091575198,
-0.0585173406,
0.0608936511,
-0.0221242364,
-0.0028387255,
0.0026104962,
0.0046398225,
-0.0167631097,
0.0120461565,
0.0068491343,
0.0013924798,
0.0066070487,
0.0164160691,
-0.0072957454,
0.020624483,
0.0072778706,
0.0345025845,
-0.0164003223,
0.0099884765,
-0.1048050374,
0.0014074564,
-0.0257588923,
-0.0177402087,
0.0578232855,
-0.0302626509,
0.0615032762,
-0.0192121696,
0.0231492631,
0.0078874938,
0.0217533447,
-0.0057750791,
0.0074533257,
0.0366986692,
-0.0039627249,
-0.0512976758,
0.024593113,
0.080818817,
-0.0094473893,
0.0467696562,
-0.018898176,
0.0019499452,
-0.0418752134,
-0.012887259,
-0.0015751965,
0.0217731241,
0.0081815133,
0.0210128166,
-0.0158940777,
-0.0731262714,
0.0513467826,
-0.0507247858,
0.0190204065,
-0.0398858562,
0.0442479886,
0.0611328892,
-0.0080759479,
0.0088920966,
-0.0184160043,
-0.0432404615,
0.0772952363,
0.0052508633,
-0.0724251494,
-6.425870725e-33,
-0.002730852,
0.0322513878,
-0.0046996144,
0.003986666,
0.0166089106,
-0.006922958,
-0.0088847028,
-0.0053138849,
-0.0498558693,
0.0207613856,
-0.0012014961,
0.0727039427,
0.0221098252,
0.0231083985,
-0.010576088,
0.0007744795,
0.0534457974,
-0.0031033678,
-0.0226961654,
-0.0208329484,
0.0332986563,
0.0549419783,
0.1189142391,
-0.0446863659,
0.0185799785,
0.0196428262,
-0.0150557328,
0.0251596365,
-0.005897895,
0.051480189,
-0.0175025743,
0.0560052432,
0.0312298518,
-0.0550981164,
-0.0158968959,
0.0366804227,
-0.042577602,
-0.014483233,
0.0419169702,
0.0006450113,
-0.0303653348,
-0.0223613661,
0.0528232642,
-0.0139502557,
0.0095429737,
0.0090349065,
0.038854003,
-0.0201909263,
0.0184355266,
-0.0002496505,
-0.012788577,
0.0393131673,
-0.0035984111,
-0.0036717101,
0.0042909742,
0.0448013023,
-0.018621603,
0.0413898341,
-0.0410199501,
0.026095124,
-0.0094870841,
0.047147762,
0.0102617471,
-0.0219887942,
-0.0258254167,
0.0787522644,
0.0096161077,
-0.028048113,
-0.0286580734,
-0.0079975538,
0.0509277657,
0.0396954939,
0.0460140146,
0.0493571311,
0.0632292926,
-0.0093227802,
-0.0814235434,
0.0030475894,
-0.0075679505,
0.0789851993,
0.0133767938,
0.0111397188,
-0.0318671763,
-0.0173256025,
-0.0243557636,
0.00472891,
-0.0147079052,
0.0005305039,
0.0092053898,
0.0400159545,
-0.0092062801,
-0.0139425974,
0.0107754311,
-0.0259942953,
0.005331005,
-0.016049305,
-0.0234836955,
-0.0019503997,
0.0031627363,
0.01714883,
-0.0613225698,
-0.0379267298,
0.0321620889,
0.0195236001,
0.0053275917,
-0.0107830139,
0.0611537024,
0.045631621,
-0.0669747964,
-0.0027483031,
-0.0249996409,
-0.0061276704,
-0.0169244297,
0.009928382,
0.029437948,
0.0636947304,
-0.0112333186,
0.0534654222,
-0.0126266154,
-0.0270051695,
-0.0317222215,
-0.0251051225,
0.0283942912,
0.0536485687,
-0.0159625839,
-0.0327118561,
0.0076552695,
0.0058649979,
0.0710319802,
-0.0581217781,
-0.0164278038,
0.0230880659,
2.614e-7,
0.0580598414,
0.0291037392,
0.0131569505,
0.0228097998,
0.0158379357,
0.0164861307,
-0.0010116622,
0.0174658857,
0.0400943533,
-0.0087541193,
-0.0126142874,
0.0177065879,
0.0609951764,
0.0351275578,
-0.0998480245,
-0.0467946045,
-0.0868333578,
-0.0094918236,
0.0000273084,
0.0124967918,
0.0172427669,
0.1542906463,
0.0334034637,
-0.0351614617,
0.0071830368,
-0.0963227674,
0.0802912638,
-0.0321675353,
-0.005432786,
0.0337263197,
0.0375258178,
0.0218202472,
0.0118268486,
0.0280457232,
-0.0147094037,
0.0615151115,
-0.0040734648,
0.0131019875,
0.0001165922,
-0.0345942304,
0.0000111563,
0.0123251099,
-0.0088046845,
-0.0275265351,
0.0540017672,
-0.0744763985,
-0.0197646562,
0.0062151304,
-0.0231186319,
-0.0118979514,
0.0377819352,
-0.0053035188,
-0.0445890948,
0.0099960342,
0.0290864296,
0.0125009799,
-0.0220554378,
-0.0883353353,
-0.0304883886,
-0.0240011234,
-0.0659800172,
-0.0480364151,
-0.0102426605,
-0.0463252999,
0.0162894316,
-0.0529766046,
0.0188370757,
2.267129578e-34,
0.0413054787,
0.0174505301,
0.0108761704,
0.0408319756,
0.0495569333,
-0.0025730792,
0.0022092997,
-0.0047865678,
0.0230638981,
-0.0769287497,
0.0032059669
] |
http://arxiv.org/abs/2405.15729v2
|
http://arxiv.org/pdf/2405.15729v2
|
Optimizing Large Language Models for OpenAPI Code Completion
| 2024-05-24T17:19:03 | 2024-06-10T21:58:24 |
Bohdan Petryshyn, Mantas Lukoševičius
|
cs.SE
|
cs.SE, cs.CL, cs.LG, 68T07, 68T50, 68T05, I.2.2; I.2.6; I.2.7; D.1.2; D.2.1; D.2.3; D.2.6
|
Recent advancements in Large Language Models (LLMs) and their utilization in
code generation tasks have significantly reshaped the field of software
development. Despite the remarkable efficacy of code completion solutions in
mainstream programming languages, their performance lags when applied to less
ubiquitous formats such as OpenAPI definitions. This study evaluates the
OpenAPI completion performance of GitHub Copilot, a prevalent commercial code
completion tool, and proposes a set of task-specific optimizations leveraging
Meta's open-source model Code Llama. A semantics-aware OpenAPI completion
benchmark proposed in this research is used to perform a series of experiments
through which the impact of various prompt-engineering and fine-tuning
techniques on the Code Llama model's performance is analyzed. The fine-tuned
Code Llama model reaches a peak correctness improvement of 55.2% over GitHub
Copilot despite utilizing 25 times fewer parameters than the commercial
solution's underlying Codex model. Additionally, this research proposes an
enhancement to a widely used code infilling training technique, addressing the
issue of underperformance when the model is prompted with context sizes smaller
than those used during training. The dataset, the benchmark, and the model
fine-tuning code are made publicly available.
|
Optimizing Large Language Models for OpenAPI Code Completion: Recent advancements in Large Language Models (LLMs) and their utilization in
code generation tasks have significantly reshaped the field of software
development. Despite the remarkable efficacy of code completion solutions in
mainstream programming languages, their performance lags when applied to less
ubiquitous formats such as OpenAPI definitions. This study evaluates the
OpenAPI completion performance of GitHub Copilot, a prevalent commercial code
completion tool, and proposes a set of task-specific optimizations leveraging
Meta's open-source model Code Llama. A semantics-aware OpenAPI completion
benchmark proposed in this research is used to perform a series of experiments
through which the impact of various prompt-engineering and fine-tuning
techniques on the Code Llama model's performance is analyzed. The fine-tuned
Code Llama model reaches a peak correctness improvement of 55.2% over GitHub
Copilot despite utilizing 25 times fewer parameters than the commercial
solution's underlying Codex model. Additionally, this research proposes an
enhancement to a widely used code infilling training technique, addressing the
issue of underperformance when the model is prompted with context sizes smaller
than those used during training. The dataset, the benchmark, and the model
fine-tuning code are made publicly available.
|
[
0.0001020734,
0.0548767075,
-0.0413476713,
-0.0172469877,
0.0039582583,
-0.0175116286,
-0.0322269872,
0.0007612217,
0.0020506722,
0.011560021,
-0.0222810395,
0.0277143251,
0.0332714617,
0.0345593207,
0.0164329205,
0.0005769767,
0.0511003062,
-0.0537723191,
-0.0457143746,
0.0155006899,
-0.0318842679,
0.013621157,
0.0181394611,
0.0847173184,
0.0343960933,
-0.037846595,
0.0117130363,
-0.0014613259,
-0.0518663637,
-0.0062525785,
0.0195080228,
0.0439074077,
-0.004766813,
0.0060987929,
0.0000024353,
-0.0569413975,
0.0183192063,
0.0006474194,
-0.0057007796,
0.06740392,
0.004694473,
0.018081069,
0.0248220842,
0.0433665365,
0.0020264436,
0.0118579669,
0.0538078994,
-0.0214512777,
0.0340009332,
0.000733271,
0.0237921104,
0.002858974,
0.0457250848,
-0.020400444,
0.0084240101,
-0.0008037645,
0.0042907782,
0.0221982468,
0.0618490167,
-0.0047989124,
0.023044968,
0.0230872855,
0.0284981504,
0.0193038583,
0.0299854726,
0.0370107815,
-0.0132683711,
-0.033820767,
0.0221919846,
0.0261504613,
-0.0196746215,
-0.0396311879,
-0.0168966074,
-0.0331797004,
-0.0391277485,
-0.0321479812,
-0.038728077,
0.0208444111,
-0.0064966143,
-0.0105255935,
0.0141721163,
-0.0283129308,
0.009416406,
0.0214397199,
-0.0236376822,
0.0728979707,
0.0126948506,
-0.0040733484,
-0.0043315529,
-0.0267597511,
-0.0270747505,
-0.0556457974,
0.0052272864,
0.0452276804,
-0.034032423,
0.0097493557,
-0.033315856,
-0.0285693239,
0.0504314974,
-0.0381486602,
0.0399805233,
0.0769269541,
-0.058510378,
-0.0116138067,
-0.0311720055,
0.0655743927,
-0.0223083515,
-0.0034893439,
-0.0478588454,
-0.0012793553,
-0.0373083316,
0.0008399585,
-0.0754305348,
0.0543634072,
-0.0247094929,
0.0147887785,
-0.0570211969,
0.0474346764,
0.0097527737,
0.0713339075,
-0.0018863517,
-0.029675832,
-0.0475122854,
0.0001321706,
0.0217540618,
-0.0113914134,
-0.0217114445,
-0.0175444987,
-0.0169139709,
-0.0450334735,
0.0055759833,
0.0107004326,
-0.0464151911,
-0.0420624316,
0.022069905,
0.0397734083,
0.0086724879,
0.0027106183,
-0.0430500805,
-0.04851586,
0.0235781875,
0.0001887372,
-0.0561723597,
-0.1016925499,
-0.0522850454,
0.0448718481,
-0.0117047215,
-0.0799839869,
-0.0171778947,
0.0202754904,
-0.0050392356,
0.0111891441,
-0.0589830652,
0.026054278,
0.0472018868,
0.0321070962,
-0.0378916487,
0.0607759282,
0.0227455553,
0.0449275114,
-0.0070782467,
-0.0427534878,
0.0225119833,
0.0439064465,
0.0001268846,
-0.0009103987,
0.0027267502,
0.0384998582,
0.0757632256,
-0.0156930573,
0.0162145793,
0.0194680709,
0.0055502206,
0.0482435189,
0.0291084815,
0.0076365499,
0.0149274468,
-0.0118886437,
0.0472228304,
0.023959199,
0.0159733258,
-0.0445619747,
-0.0484239236,
0.0489743352,
-0.0077629816,
0.0245887432,
0.0240168516,
0.0625484511,
-0.0031390199,
-0.0607371368,
0.0040638126,
-0.0497864634,
-0.0645889044,
-0.0010992858,
0.0430852994,
-0.0067707449,
-0.0571023375,
0.0094436733,
0.054919634,
-0.0545614995,
-0.0088871717,
-0.0345339142,
0.0396225415,
0.0356558599,
0.0609779023,
-0.0313304216,
-0.0275646914,
-0.0288040619,
0.0142625505,
-0.0278688222,
0.0094540361,
0.0009037626,
-0.0231208205,
-0.0712957829,
-0.0195165128,
-0.0235022139,
0.0023267074,
0.0388521329,
0.0268723425,
-0.0067383596,
0.0302298088,
-0.0015574613,
-0.0505742095,
0.041490566,
-0.0176230129,
-0.0015863528,
0.0033623485,
0.0159937628,
0.0429834798,
-0.0286662485,
-0.0263714995,
0.0181288049,
0.0227457564,
-0.0226946305,
-0.0324010998,
-0.0253246296,
-0.0132344449,
0.0445243157,
0.0181017406,
0.0013642198,
-0.0145202046,
0.0079223793,
0.0009068034,
-0.0213239938,
0.0317811891,
-0.0128493663,
0.0493649729,
-0.0083009154,
0.0311231297,
-0.0390525721,
0.0038041542,
-0.0260199234,
0.0252686664,
0.0073565864,
0.0248447601,
0.013151668,
0.0837061554,
0.0340316147,
0.0262594465,
-0.0367536768,
0.0280588455,
-0.041600734,
0.030558642,
0.0060978923,
0.0295063592,
-0.0026556998,
-0.0561399199,
-0.0989650339,
0.0390300378,
0.001188168,
-0.0047629117,
-0.0079003554,
0.0640440136,
-0.0066876994,
0.0121805165,
0.0317857079,
-0.0077808327,
-0.0196398776,
0.0142031396,
0.0129902335,
0.0197095554,
-0.0795102194,
-0.037708737,
-0.0028162652,
-0.0047173197,
0.0161580127,
-0.0207851231,
-0.0171402693,
-0.0431234315,
0.0923825428,
0.0002681991,
0.0164486002,
-0.0186082181,
-0.0511442535,
0.0102111809,
-0.0111283734,
0.0761444122,
0.0248333029,
-0.041492898,
-0.0448275171,
0.0586230569,
0.017589625,
-0.0559613258,
0.025820177,
0.0402213857,
-0.0966038629,
0.0187192466,
-0.0156126767,
0.046016749,
0.0697534084,
-0.014228153,
-0.0140395518,
-0.0472909287,
-0.0257849935,
-0.0439859517,
0.0077798967,
0.013319972,
-0.0196383409,
-0.0128450021,
0.083828114,
-0.0299153291,
-0.1089791656,
0.0328029245,
-0.0539009981,
-0.022089608,
-0.0246629156,
-0.013060811,
-0.0580037795,
0.0325082727,
-0.0014213149,
0.0138375079,
0.0081319539,
-0.0307906494,
-0.0373558626,
0.0064507639,
-0.0230455212,
0.0046417098,
0.0496408716,
-0.0396454819,
0.0114182094,
-0.0067125494,
-0.0132754985,
-0.0155499149,
0.0127045857,
-0.0445169061,
-0.0051072245,
0.015613202,
0.0172620695,
0.0347264074,
-0.0065209013,
-0.0126969041,
0.0454876386,
0.0358931869,
-0.0324594639,
-0.0310549084,
-0.0663240626,
0.0083332742,
-0.0721454546,
0.0008466183,
-0.0140477819,
-0.0361945033,
-0.0119391829,
0.0611808598,
0.0428347029,
-0.0598880202,
-0.0252720695,
0.0648976862,
-0.0483133905,
-0.0761137679,
0.1092823893,
0.0252126455,
-0.0303679891,
0.0040171696,
0.0806450397,
0.0500742272,
-0.0622110032,
0.0044397865,
-0.0250331219,
0.0553327985,
0.0647589117,
0.0307727791,
-0.0893736482,
-0.0168800559,
0.0071114274,
0.0267448686,
-0.0219391529,
0.003808351,
-0.0107747307,
-0.0893590152,
0.0372925699,
0.002178472,
-0.0329381004,
-0.030607393,
-0.0499694608,
0.0274287015,
0.0243281703,
-0.0361123495,
0.0384127721,
0.0011595748,
0.027863076,
-0.0095361099,
-0.0236596297,
-0.0047768904,
0.0254716258,
-0.0447090603,
0.0493376069,
0.1015569493,
-0.0248373747,
-0.0154148499,
0.0033909944,
0.0608485118,
-0.0262666289,
-0.0008607259,
-0.0408113077,
0.024940243,
0.026760282,
0.0147436708,
0.015615345,
0.0146078803,
-0.0499150679,
-0.0378682278,
0.0425055176,
0.0655005947,
-0.001041314,
0.0071677985,
-0.0208651479,
-0.0160655212,
-0.0153387608,
0.0117226522,
0.0495411865,
-0.0238845181,
-0.0179341938,
-0.0405317955,
0.0058111125,
-0.0380934551,
0.0554747768,
-0.0061118025,
0.006923038,
-0.0116287423,
-0.0170151088,
0.0027901081,
0.0413889289,
0.0325159766,
0.0354835838,
0.0073933247,
0.0091070719,
-0.016352592,
-0.0594317429,
0.0158057883,
-0.0395245366,
0.0298685916,
-0.077463299,
-0.0077527543,
0.037749961,
-0.0437403023,
-0.0131804012,
0.031612698,
-0.0704211518,
-0.0341262296,
-0.0298861805,
0.081905134,
0.0302839931,
-0.0121070901,
-0.0484677441,
-0.094191581,
-0.0355438553,
0.0098133767,
-0.0271347836,
0.0380765572,
0.0049963244,
0.0178139433,
0.0330094509,
-0.0110828765,
-0.067723237,
-0.0233783461,
0.0134201553,
-0.0262712948,
0.0523432754,
-0.0128564918,
0.0342793688,
0.037514396,
-0.0101911537,
0.0042974306,
-0.0191708393,
-0.028876733,
0.0075538014,
0.0066449302,
-0.009627304,
-0.0339051969,
0.003053698,
0.0169858523,
0.0080757141,
0.0456009135,
-0.0069087064,
-0.0535239093,
0.0671574846,
-0.0085043544,
0.0490684398,
-0.0183660649,
0.0566997565,
0.0841882154,
0.005326225,
-0.0335570276,
-0.012815454,
-0.0019847618,
0.0211511739,
-0.0340411998,
0.050601501,
-0.012042434,
0.0495981351,
-0.0008569523,
-0.0003548834,
-0.016233936,
-0.0633531436,
-0.0065153213,
0.0322777219,
0.0342279896,
0.0371398404,
-0.0003421499,
-0.0247748811,
0.0158862602,
-0.0095778834,
0.062482398,
0.0022344973,
-0.0394476019,
0.0074299769,
0.0075288122,
-0.0172629934,
-0.058581017,
0.1199963614,
-0.0323648229,
0.0448925085,
-0.0343845822,
0.0147984792,
-0.012161117,
0.0007169351,
0.0420345925,
0.0346820503,
-0.0449995734,
-0.0276299343,
-0.0348566249,
-0.0424950644,
0.0059760613,
0.0062676864,
-0.0272584409,
-0.0015250001,
0.0270569175,
0.0887827203,
0.0181435551,
-0.0432562158,
-0.0010032177,
-0.0323720723,
-0.0074365023,
-0.042696923,
-0.0066417553,
-5.841086898e-33,
-0.0479258299,
-0.0141127128,
0.0254945513,
0.0369633697,
0.0132807652,
-0.0232697669,
-0.0358134247,
0.0144961681,
0.0084393378,
-0.0703445524,
0.0169747099,
-0.0069936421,
-0.0094057303,
-0.0259828027,
0.066669181,
0.0560299791,
0.0433668643,
0.0296809636,
-0.0474856012,
0.0160251688,
-0.0660586581,
0.031734243,
-0.0533391386,
-0.0250968523,
-0.0165727064,
-0.0126319723,
0.0287551936,
-0.0232104156,
0.0018520664,
0.0610836297,
0.0112194093,
0.0233237613,
0.0043495554,
0.0034785036,
-0.0239016246,
-0.0184739903,
-0.0059232591,
-0.0484485812,
0.0488253087,
0.0079344781,
0.0180709884,
-0.0764859542,
0.0454527363,
-0.0072704284,
-0.0386294052,
0.0148692755,
0.0177661739,
-0.0060051843,
-0.0303950515,
0.027740106,
-0.0691993237,
0.0164903533,
-0.0208022911,
-0.009642425,
-0.0070847673,
0.0231530517,
0.0001514898,
0.0591933317,
0.0448669419,
-0.0207718052,
0.0274410117,
0.0633554906,
0.0528791957,
0.0090746721,
-0.0120318271,
-0.0298792347,
0.0502863601,
0.0288690906,
-0.0289365016,
-0.0010932616,
0.0321460254,
0.0530408882,
-0.0001829939,
0.0306952596,
0.0015605339,
-0.0635680035,
-0.0225238502,
0.0299530942,
-0.0064497152,
0.0279982109,
0.0062974081,
0.0285238344,
-0.0482869297,
-0.0177446809,
-0.0697102845,
-0.0312606022,
-0.0417860448,
-0.0381207988,
0.0429286286,
-0.0022426539,
-0.0160637647,
-0.002812854,
0.008148103,
0.0003044583,
0.0351275057,
0.0020860382,
-0.0566106364,
0.072380662,
-0.0088780625,
0.0410513245,
-0.0323085859,
-0.018831864,
-0.0039494252,
-0.0248424336,
0.0716480166,
0.0241245553,
0.0136511931,
0.0133427009,
-0.0637034178,
0.0248263236,
-0.0455473848,
-0.0166288167,
0.0254559238,
-0.0071863178,
0.019514557,
-0.007949302,
0.0217632595,
0.0539570116,
-0.0100845704,
-0.0282440688,
-0.0723237321,
-0.0440591425,
0.0111534586,
-0.0214643646,
-0.0274460968,
-0.0136068398,
-0.0334906727,
0.0515801162,
0.0380005836,
0.0093649607,
-0.0631570742,
0.0093724504,
3.081e-7,
0.0088110305,
0.0482129455,
-0.0291253403,
0.0548062474,
-0.0302568655,
0.0102254013,
-0.0197690222,
-0.0280409642,
0.0268875919,
0.0138925128,
-0.030792864,
-0.0058634193,
0.029840976,
0.0183256343,
-0.0365299769,
0.0154263275,
0.0196446553,
0.026192192,
-0.1079204977,
-0.0101505015,
0.0255119111,
0.0480762757,
-0.0044903024,
0.0076012537,
-0.003454726,
-0.0196313336,
0.0282421019,
-0.0238822717,
0.0055543776,
0.0604329072,
-0.0705615133,
0.0364376307,
-0.002137989,
-0.0228546411,
0.0132637164,
-0.0266172327,
0.0167184211,
0.0533484258,
0.0059343739,
0.0768003911,
0.0581847504,
-0.0186469443,
-0.02234588,
-0.0986698866,
0.032554321,
-0.0416577794,
-0.013111759,
-0.0296819676,
-0.0261901654,
-0.0104901902,
-0.0006080955,
-0.0206307471,
0.0534074381,
-0.000187028,
-0.0147943124,
-0.0154206688,
-0.0433697626,
-0.0100483056,
-0.0434843488,
0.011784533,
-0.0142229553,
-0.0469030477,
-0.0804420263,
0.0013175025,
0.021249637,
-0.0299506094,
0.0550885275,
3.077507807e-34,
0.022299435,
-0.0533200912,
0.0331236683,
0.016876258,
0.039704904,
-0.0321070999,
-0.0045165797,
0.0484969504,
-0.0187430903,
-0.017620245,
-0.0152554447
] |
http://arxiv.org/abs/2405.15708v1
|
http://arxiv.org/pdf/2405.15708v1
|
EmpathicStories++: A Multimodal Dataset for Empathy towards Personal Experiences
| 2024-05-24T16:57:18 | 2024-05-24T16:57:18 |
Jocelyn Shen, Yubin Kim, Mohit Hulse, Wazeer Zulfikar, Sharifa Alghowinem, Cynthia Breazeal, Hae Won Park
|
cs.CL
|
cs.CL
|
Modeling empathy is a complex endeavor that is rooted in interpersonal and
experiential dimensions of human interaction, and remains an open problem
within AI. Existing empathy datasets fall short in capturing the richness of
empathy responses, often being confined to in-lab or acted scenarios, lacking
longitudinal data, and missing self-reported labels. We introduce a new
multimodal dataset for empathy during personal experience sharing: the
EmpathicStories++ dataset
(https://mitmedialab.github.io/empathic-stories-multimodal/) containing 53
hours of video, audio, and text data of 41 participants sharing vulnerable
experiences and reading empathically resonant stories with an AI agent.
EmpathicStories++ is the first longitudinal dataset on empathy, collected over
a month-long deployment of social robots in participants' homes, as
participants engage in natural, empathic storytelling interactions with AI
agents. We then introduce a novel task of predicting individuals' empathy
toward others' stories based on their personal experiences, evaluated in two
contexts: participants' own personal shared story context and their reflections
on stories they read. We benchmark this task using state-of-the-art models to
pave the way for future improvements in contextualized and longitudinal empathy
modeling. Our work provides a valuable resource for further research in
developing empathetic AI systems and understanding the intricacies of human
empathy within genuine, real-world settings.
|
EmpathicStories++: A Multimodal Dataset for Empathy towards Personal Experiences: Modeling empathy is a complex endeavor that is rooted in interpersonal and
experiential dimensions of human interaction, and remains an open problem
within AI. Existing empathy datasets fall short in capturing the richness of
empathy responses, often being confined to in-lab or acted scenarios, lacking
longitudinal data, and missing self-reported labels. We introduce a new
multimodal dataset for empathy during personal experience sharing: the
EmpathicStories++ dataset
(https://mitmedialab.github.io/empathic-stories-multimodal/) containing 53
hours of video, audio, and text data of 41 participants sharing vulnerable
experiences and reading empathically resonant stories with an AI agent.
EmpathicStories++ is the first longitudinal dataset on empathy, collected over
a month-long deployment of social robots in participants' homes, as
participants engage in natural, empathic storytelling interactions with AI
agents. We then introduce a novel task of predicting individuals' empathy
toward others' stories based on their personal experiences, evaluated in two
contexts: participants' own personal shared story context and their reflections
on stories they read. We benchmark this task using state-of-the-art models to
pave the way for future improvements in contextualized and longitudinal empathy
modeling. Our work provides a valuable resource for further research in
developing empathetic AI systems and understanding the intricacies of human
empathy within genuine, real-world settings.
|
[
0.0114076715,
0.028609423,
-0.0363359749,
0.0112649444,
-0.0240097847,
0.0281164609,
0.0480828211,
-0.0270757228,
0.0271434989,
-0.007863123,
-0.0173735842,
0.0070343763,
-0.0177206937,
0.0150478277,
-0.0024029301,
-0.0627848133,
0.0228162743,
0.0675558373,
0.0323094092,
0.0239454582,
-0.0077218711,
0.0400857776,
-0.027374452,
0.0536538474,
-0.0397978686,
-0.0296659451,
-0.0059817773,
0.0172331464,
0.0504633524,
-0.0469966717,
0.0006213118,
0.0014347724,
0.0344108529,
-0.0253123455,
0.0000020002,
-0.0208397899,
0.0230366178,
-0.0271803122,
0.0253830645,
-0.04326104,
0.0585056804,
-0.0337119401,
0.0320915431,
0.0197534505,
-0.0508498326,
-0.034756925,
0.0161864348,
0.0203087237,
0.0303025749,
0.0497035943,
-0.0463302173,
0.0156815462,
-0.0062548011,
-0.0026146006,
0.0154822087,
0.0205646995,
-0.0382853299,
-0.0097508831,
0.0336931124,
-0.0634331033,
-0.0536540933,
0.0240478851,
0.0364222191,
0.0220691804,
0.0385383852,
0.0348622128,
0.0352694578,
-0.056609638,
-0.0129122874,
0.1055585891,
0.041760534,
0.0428481847,
0.0050149863,
-0.0095637981,
0.0212693997,
-0.0267433338,
-0.0324832648,
0.0109950658,
-0.0183818378,
0.005845686,
-0.0476964675,
0.0386319645,
0.0139079699,
0.0033056678,
0.0730076209,
0.0045998846,
-0.0139498906,
0.0383687653,
-0.0022828204,
-0.0090192305,
0.0086501231,
-0.0342887789,
0.0108638695,
0.0021447763,
-0.0029634952,
0.0023739897,
-0.062284708,
0.0472636409,
-0.0239249133,
0.0236758385,
0.0244686883,
0.0352251269,
-0.0163890049,
0.0097710611,
0.0279385727,
0.005930115,
0.0123158796,
0.0246479735,
-0.0353455804,
0.0084969411,
-0.0408197306,
0.0342090651,
0.0210931357,
0.0005350186,
-0.036894422,
-0.0318378992,
-0.0211306196,
0.0279819444,
-0.0130989384,
0.0109198475,
-0.0582615472,
0.0174408648,
-0.0013136872,
0.0080041047,
-0.0085778581,
0.029004246,
-0.0637698695,
0.0402028933,
0.0220096521,
-0.0474036671,
-0.0047823335,
-0.0085144974,
0.0332759097,
-0.034361124,
0.0272706654,
0.0250108186,
0.0018071879,
-0.0731965899,
-0.010446324,
0.0038713701,
0.0201821197,
-0.0316423811,
0.016032761,
0.0032450282,
0.0733442158,
0.0064553958,
0.0346165858,
-0.0232949294,
-0.00246883,
-0.0168260988,
-0.0301482361,
0.0695601627,
0.0161560457,
0.0235658474,
0.0032340765,
0.03207222,
0.0214154832,
-0.0121993246,
0.0329332575,
0.017515311,
0.0530713238,
-0.0167824626,
0.0404005349,
-0.0057308497,
-0.0315042399,
-0.0053389706,
0.0202588122,
-0.0379812121,
-0.0301296506,
0.084381938,
0.0400050208,
0.006219002,
-0.0129166311,
-0.0596455894,
0.062516138,
0.1147155166,
-0.0166334007,
0.0516928211,
-0.0006044472,
-0.0020320581,
-0.0030094415,
-0.0222970434,
-0.0414786562,
0.0759046674,
0.0428241044,
-0.0428038798,
-0.0151376612,
0.0050248825,
-0.0391138829,
-0.0122637367,
-0.0067549311,
-0.0344202146,
0.0436357558,
-0.0035380677,
-0.0173779652,
-0.0177105982,
-0.0262974035,
-0.0560726672,
0.018790422,
-0.0822320804,
0.0180157702,
-0.0205940288,
0.032384403,
0.0922947899,
0.0402686745,
-0.0499166586,
-0.0238585379,
0.0038295945,
-0.0110663688,
-0.0234369859,
0.0156753697,
0.0027727352,
0.085415788,
-0.0268825796,
-0.002970255,
-0.014842988,
0.0130847255,
0.0160912797,
-0.0523339845,
0.0297721308,
-0.0226616245,
-0.032914035,
0.0072743171,
0.0511400998,
-0.0859439075,
0.0249025207,
0.0083594145,
0.0210319273,
0.0095210802,
0.0267753843,
-0.0040939376,
-0.0396214984,
0.0229970943,
-0.0206101369,
-0.021495115,
0.1193042323,
-0.0551696494,
-0.0402201898,
0.0114620384,
-0.1138341129,
-0.0009487096,
0.0630425587,
0.01850136,
0.0097727282,
0.0427297316,
-0.0273867436,
0.0136129297,
-0.0288540833,
0.0017037874,
-0.0063318345,
0.0161671732,
0.0253833774,
-0.0100303199,
-0.0305590369,
-0.0070853219,
-0.0606371164,
0.0068653431,
-0.0156770032,
-0.0016193866,
-0.0143943261,
0.0027209255,
0.0164086614,
-0.0140645131,
-0.0121014975,
-0.0337499008,
-0.0792144388,
-0.0075642103,
-0.1361325383,
-0.0057568122,
-0.0020069866,
-0.0452017784,
0.0076908115,
-0.010866615,
-0.0370229855,
-0.0195232611,
-0.0382084027,
0.0652886927,
-0.0078411447,
0.0084756901,
0.0237468537,
0.0104060937,
0.0283018611,
-0.0806387067,
0.002819919,
-0.0304028969,
0.0386660919,
0.0109730838,
0.0283938069,
-0.0082914103,
0.0009834358,
-0.0437719785,
0.0270517915,
-0.0295954999,
-0.0548189729,
0.0023087407,
-0.0233124867,
0.0091447523,
0.0347550213,
0.0105843712,
0.0670348704,
0.0496923029,
0.0012474828,
-0.0295475405,
0.0262924973,
-0.0050746952,
0.0406592563,
-0.0541173369,
-0.049538102,
-0.0085331332,
0.0264960416,
0.0096289078,
0.0264500063,
-0.0341277122,
-0.033558473,
0.0127696861,
-0.058674261,
0.0211017188,
0.012487377,
0.0074922782,
-0.0085919658,
-0.0197065119,
-0.0517366193,
0.0566434525,
0.0282009058,
-0.0374572165,
0.0132572791,
-0.000373159,
-0.0060701459,
0.0953220427,
-0.0313159898,
0.0251833647,
-0.0223044939,
0.0246783867,
-0.0343633257,
0.0073918845,
0.0231345147,
-0.0025825677,
-0.026680477,
-0.0200771131,
-0.0528440736,
0.0296116658,
-0.0259454679,
-0.0143600916,
-0.0699950904,
-0.0196855273,
-0.039156653,
-0.0218632836,
-0.0113418596,
-0.0363514125,
-0.0395172015,
0.0475683063,
0.0355534591,
-0.0066730375,
-0.0197837055,
-0.0169790499,
-0.0011008667,
0.0570321418,
0.0091642225,
0.020960236,
-0.0212570466,
0.0896555558,
0.0304953437,
0.0767651051,
0.0245650448,
-0.0027533725,
0.0355911776,
-0.0256067812,
0.0361948237,
0.0123862177,
-0.0205879342,
0.0503940396,
0.0048478097,
0.0088841058,
0.0999408439,
0.0020844953,
-0.0939263254,
0.0307000261,
-0.0218497328,
-0.0425116979,
0.0283956099,
-0.0330413245,
-0.0335509144,
-0.0146876788,
0.0035620993,
0.0254923981,
-0.0034535325,
-0.0722022802,
0.0007664417,
-0.0656859875,
0.0413253829,
-0.0229999814,
0.0293337274,
-0.0577819459,
-0.0272473171,
0.0126767382,
0.0155564668,
0.0109482938,
-0.0219527315,
0.0024946101,
-0.0129935537,
-0.0185197406,
-0.0075746085,
-0.0094219781,
0.0417545773,
-0.0552576371,
0.0688721612,
0.0096164411,
0.083161518,
0.0061049829,
0.0056863888,
-0.0669515356,
0.0516560189,
0.0205694921,
-0.0009989997,
0.0595835783,
-0.0435578227,
0.0096186502,
0.0015216054,
0.0893811584,
-0.040211428,
-0.0064760381,
0.0270888135,
0.0646127462,
-0.0358816683,
-0.0339238383,
-0.0115611833,
0.0274320878,
0.0267043486,
0.0356675684,
0.0125861466,
-0.0577991046,
-0.0023838694,
0.0538471378,
0.0239052735,
0.0207898766,
0.0065248157,
0.0125512453,
-0.0039275219,
-0.001801696,
-0.0392348245,
-0.0839638487,
0.0194715373,
0.0051884074,
-0.003589296,
-0.0705648288,
-0.0006147111,
0.023481803,
-0.0846297219,
0.0344985463,
0.025813017,
0.0385298282,
0.0144495834,
0.0409727022,
0.0344748981,
-0.0203269497,
0.0237888712,
0.0035702677,
-0.0123823043,
-0.0184473228,
0.0084387055,
0.0323641151,
0.007164896,
0.0501270778,
0.0181895494,
-0.0330432579,
0.0193738043,
0.018932905,
-0.074789986,
0.0520252548,
0.0119917244,
0.0230962951,
0.0047291229,
0.015351559,
-0.0142171821,
0.0044779736,
-0.0047557992,
-0.0140732462,
0.0249113813,
-0.040313527,
-0.0173020195,
-0.0210354328,
0.0140509205,
-0.0190839749,
0.0006890055,
-0.045260977,
-0.0267666113,
-0.0288499128,
0.0394168831,
-0.0384478644,
-0.0046969787,
0.0313255601,
0.0232345052,
-0.0409047157,
-0.0517931245,
-0.0267918855,
0.0065925396,
-0.0683519393,
-0.0024320579,
-0.0125562083,
-0.011230059,
-0.0404081307,
0.0385120623,
0.0204796772,
-0.0028871833,
-0.0463607796,
-0.0123005304,
0.0013200863,
0.0599364303,
0.0154845743,
-0.0072303326,
-0.023500232,
0.048225645,
-0.0118610524,
0.0101401722,
-0.0052536861,
0.0308498573,
0.042594254,
0.0052403836,
0.0298357699,
-0.02154948,
0.028445363,
-0.0294834971,
0.0414074585,
-0.0331670158,
-0.0429738425,
-0.0467067361,
0.0081098331,
-0.0137437433,
0.0321879499,
0.0160039235,
-0.0562795214,
0.0599444509,
-0.0060978602,
-0.0148686692,
-0.0309809092,
0.017830288,
0.0676266402,
0.0029324787,
0.0030435042,
0.0159601085,
-0.0606852956,
-0.0471559428,
0.0315396413,
-0.0182240922,
-0.0270301048,
-0.0317512229,
0.0128276953,
0.0363491364,
-0.009774914,
0.0073930118,
-0.0233676676,
0.0091554699,
-0.058912538,
-0.0101631079,
0.0110873934,
-5.201426953e-33,
0.0204425268,
-0.0062925904,
-0.0506659336,
-0.01717842,
-0.0424564518,
0.0305975843,
-0.0365870781,
0.0080203526,
0.0205711611,
0.0501093306,
-0.0067785135,
0.0526251309,
0.0001606053,
-0.0100075193,
-0.001351816,
-0.0033745272,
-0.0028564946,
-0.0189340021,
-0.0122252591,
0.053081695,
-0.0036636402,
0.0652732477,
0.0360179767,
-0.0674854368,
-0.025999805,
0.0004066753,
-0.0265805814,
-0.0186472423,
0.0489696339,
0.0133516481,
-0.0224446226,
0.0427575931,
0.0409293324,
0.0232574493,
0.005120418,
0.0269774757,
-0.0713133886,
-0.0002757862,
-0.0217191949,
-0.0059229098,
-0.1039274931,
-0.0278456993,
0.0123775722,
0.0148605471,
-0.0276732855,
-0.0458875187,
0.0391385928,
0.0041503375,
0.0529858768,
0.020006394,
-0.0221305359,
-0.0432122536,
-0.044987116,
-0.0607233979,
-0.0076575666,
0.0423250124,
-0.0037356964,
0.0397104882,
-0.0268506948,
0.000611538,
-0.1002513543,
-0.0362598822,
0.012360977,
-0.0491455942,
0.0124999797,
0.1017244384,
0.0352939293,
0.0002615569,
-0.0236003157,
0.0680877715,
0.0423997864,
0.0392307937,
0.0119862324,
0.0733679906,
-0.0554303825,
-0.0450688973,
-0.007964923,
0.0613045655,
0.0187217742,
-0.0027380877,
0.0306142643,
-0.0011445225,
-0.006723145,
0.0324709751,
-0.0449829064,
-0.077882871,
-0.0116959754,
-0.0147418501,
0.0136496294,
-0.0274533853,
-0.0162388794,
-0.0233376641,
0.0037354247,
-0.0379652642,
0.0194832738,
0.0148615614,
-0.045214802,
0.0467278026,
0.015265855,
0.0343488641,
0.0084784338,
-0.013337736,
0.0343874581,
0.0164993014,
0.0803810507,
-0.0218790807,
-0.0101426365,
0.0548950396,
-0.0049956515,
-0.0345800892,
-0.0326622277,
0.0356563255,
0.0348219499,
-0.017655367,
-0.0280459579,
0.0129201086,
-0.034412507,
-0.0022911837,
0.0041928436,
0.0703905001,
-0.0566793904,
-0.0712261721,
-0.0259235203,
0.0191867799,
-0.0165740773,
-0.0285888426,
-0.0453476459,
-0.0687129125,
0.0725447014,
-0.0204686802,
-0.0212193839,
-0.0171797406,
2.553e-7,
-0.0657814518,
0.0230308622,
0.0297776405,
-0.0478913859,
0.0219215341,
0.0544244498,
0.0007102526,
0.118855238,
0.0013487708,
-0.0464299545,
0.0582569391,
0.0173411909,
0.0581150129,
0.0035477136,
-0.0302282497,
0.0401942059,
-0.0190974306,
-0.0089617996,
-0.0336220562,
0.003674095,
0.0574044175,
0.045718383,
0.0736567304,
-0.0225033481,
0.0570770651,
0.0085330699,
0.0058641746,
-0.0885867402,
-0.0172613505,
-0.0188621506,
-0.0256251544,
-0.0100855129,
0.0153132025,
-0.043922279,
-0.0002525014,
-0.0058335601,
0.0108823497,
0.0633229315,
-0.0266687293,
-0.0968683958,
0.027157655,
0.049104847,
-0.0381635688,
-0.016600363,
0.0295795556,
-0.0508152097,
-0.0361994542,
0.0159153529,
0.0050187372,
0.0013119647,
0.00934863,
-0.0208235532,
-0.034879081,
0.0256605949,
0.0133837834,
-0.0163487308,
0.0016552608,
-0.0749651343,
0.017268436,
-0.0338090584,
-0.0652308911,
-0.0477249622,
0.002265909,
0.0040625199,
-0.0047301315,
-0.035025388,
0.0037067817,
1.607013506e-34,
-0.0246192385,
0.0107366852,
-0.0202618185,
0.008514598,
0.0247986652,
-0.0135986954,
-0.0010412189,
-0.0123127215,
0.0309320092,
-0.0385759249,
-0.0188098699
] |
http://arxiv.org/abs/2406.00029v1
|
http://arxiv.org/pdf/2406.00029v1
|
Clustered Retrieved Augmented Generation (CRAG)
| 2024-05-24T16:36:47 | 2024-05-24T16:36:47 |
Simon Akesson, Frances A. Santos
|
cs.CL
|
cs.CL, cs.AI
|
Providing external knowledge to Large Language Models (LLMs) is a key point
for using these models in real-world applications for several reasons, such as
incorporating up-to-date content in a real-time manner, providing access to
domain-specific knowledge, and contributing to hallucination prevention. The
vector database-based Retrieval Augmented Generation (RAG) approach has been
widely adopted to this end. Thus, any part of external knowledge can be
retrieved and provided to some LLM as the input context. Despite RAG approach's
success, it still might be unfeasible for some applications, because the
context retrieved can demand a longer context window than the size supported by
LLM. Even when the context retrieved fits into the context window size, the
number of tokens might be expressive and, consequently, impact costs and
processing time, becoming impractical for most applications. To address these,
we propose CRAG, a novel approach able to effectively reduce the number of
prompting tokens without degrading the quality of the response generated
compared to a solution using RAG. Through our experiments, we show that CRAG
can reduce the number of tokens by at least 46\%, achieving more than 90\% in
some cases, compared to RAG. Moreover, the number of tokens with CRAG does not
increase considerably when the number of reviews analyzed is higher, unlike
RAG, where the number of tokens is almost 9x higher when there are 75 reviews
compared to 4 reviews.
|
Clustered Retrieved Augmented Generation (CRAG): Providing external knowledge to Large Language Models (LLMs) is a key point
for using these models in real-world applications for several reasons, such as
incorporating up-to-date content in a real-time manner, providing access to
domain-specific knowledge, and contributing to hallucination prevention. The
vector database-based Retrieval Augmented Generation (RAG) approach has been
widely adopted to this end. Thus, any part of external knowledge can be
retrieved and provided to some LLM as the input context. Despite RAG approach's
success, it still might be unfeasible for some applications, because the
context retrieved can demand a longer context window than the size supported by
LLM. Even when the context retrieved fits into the context window size, the
number of tokens might be expressive and, consequently, impact costs and
processing time, becoming impractical for most applications. To address these,
we propose CRAG, a novel approach able to effectively reduce the number of
prompting tokens without degrading the quality of the response generated
compared to a solution using RAG. Through our experiments, we show that CRAG
can reduce the number of tokens by at least 46\%, achieving more than 90\% in
some cases, compared to RAG. Moreover, the number of tokens with CRAG does not
increase considerably when the number of reviews analyzed is higher, unlike
RAG, where the number of tokens is almost 9x higher when there are 75 reviews
compared to 4 reviews.
|
[
0.0489515923,
-0.0030889781,
-0.0232546907,
0.0276005436,
-0.025796663,
-0.0262682922,
0.0099846451,
0.0089512374,
-0.0020785821,
-0.0708743483,
-0.0047096787,
-0.0525577776,
-0.0147288879,
0.0105965836,
0.0041302047,
-0.0031680432,
0.0659356564,
-0.0195238367,
0.0088857105,
-0.022040002,
-0.0022907499,
0.030334793,
0.0161804482,
0.0429782644,
-0.0143191777,
-0.001975972,
-0.0134147275,
-0.0052259988,
0.0091230432,
-0.0124524022,
0.0042784228,
0.0482869893,
0.0019532256,
0.0504091159,
0.0000022128,
-0.0608919524,
-0.0525158122,
0.0166045036,
-0.0235511735,
0.02676378,
0.0515408032,
0.0371887982,
-0.0066391011,
0.0440247692,
-0.0260757245,
-0.0425612256,
0.0263330024,
0.0406776182,
0.0304353014,
0.0442945734,
-0.0120719001,
-0.0272736121,
-0.0029407507,
-0.0034394965,
-0.001140269,
-0.0154201267,
0.0226615369,
0.0048597441,
0.056421563,
0.0342297666,
0.0332253054,
0.0387028828,
0.0125436224,
-0.0277668182,
0.0097084763,
0.0234986991,
-0.0743660033,
0.0165849496,
-0.0087238131,
0.0282974634,
0.0363152623,
-0.0355835855,
0.0214054026,
0.0079214629,
-0.0107717747,
0.0227591284,
-0.0666824654,
0.005195661,
-0.0167763382,
0.010257219,
-0.0017103422,
-0.0189157072,
0.0060386392,
-0.0067170989,
-0.0357974991,
0.0529542938,
0.0255127829,
-0.0222536698,
-0.0329250917,
0.0008374688,
0.0046973396,
-0.0346048363,
-0.0202178508,
0.019342266,
-0.0576275811,
0.0301466864,
-0.0019932953,
-0.0524610542,
0.0651560128,
-0.0461560562,
0.0355423875,
0.0267768595,
-0.0271719098,
0.0286355559,
-0.0474450178,
0.0235524792,
-0.0060831537,
-0.0176247135,
-0.0379295088,
-0.0031796293,
-0.0510287397,
-0.0283505954,
-0.0407440029,
0.0682946667,
-0.0677120462,
-0.0107604265,
-0.0358877964,
0.0590123609,
0.0196217336,
0.0164584219,
-0.0100980531,
0.0048903851,
-0.00801514,
0.019051455,
-0.0136061879,
0.0259622205,
-0.0169132054,
-0.009077738,
0.0068047792,
-0.0898776874,
0.0093384655,
0.0079379538,
0.0228215959,
-0.0034446144,
-0.0274381582,
0.0424365811,
0.0036215428,
-0.0282144211,
-0.0392507277,
-0.0861866549,
-0.0298072305,
-0.0428605154,
-0.0152668348,
-0.1039103419,
0.0337671302,
0.0433302186,
-0.0362064056,
0.0033907969,
0.0029644577,
0.0240410306,
-0.0255202241,
0.0027775532,
-0.0465805158,
0.0365826264,
0.0229084026,
0.0078964923,
0.0293348059,
0.0923087448,
-0.0441911481,
0.0129485652,
0.0325271152,
0.0002130192,
0.0090309922,
0.0315695107,
0.0330588371,
0.0346810818,
-0.0455842428,
0.0278174058,
-0.0125505449,
-0.0364512615,
-0.0105587076,
0.0501659587,
-0.0307276603,
0.0377135724,
0.0033641658,
0.0190835651,
0.0378535241,
0.0086061656,
0.0255417172,
0.0362065546,
-0.0106937904,
-0.0002051996,
-0.0174902994,
0.0197758786,
-0.0499790199,
0.0293068346,
0.0755168572,
0.0152256554,
-0.0288730524,
-0.0239349771,
-0.006087034,
0.0176410936,
-0.0173704885,
0.0042596725,
0.0540736355,
0.0183591079,
-0.010479439,
0.0315353274,
0.0355015174,
-0.0586605705,
0.0077042407,
-0.0516089462,
0.0371492244,
0.0837193206,
0.1062821448,
0.0054783793,
-0.0338494629,
0.024994893,
-0.0080796247,
0.0318365134,
0.0500623323,
-0.0159418676,
-0.0411878973,
-0.0366775393,
-0.0346087404,
0.0162159204,
0.0103665115,
0.0325931534,
0.0366155431,
0.0114692189,
0.0281133149,
-0.0126090739,
-0.0402717441,
-0.0087960791,
0.013434818,
-0.01729583,
0.0036274691,
-0.0033131032,
0.0372768417,
0.0635940433,
-0.0140249925,
0.0352657512,
0.0572503023,
-0.0160211064,
-0.0358339138,
0.01729762,
-0.0124334944,
-0.0039281952,
0.0250006504,
0.0258081071,
0.0132551324,
-0.0351700671,
-0.0018524863,
-0.0528992563,
0.0122496942,
-0.0286007728,
0.1086917669,
-0.0468228981,
0.0314066708,
0.0008031207,
-0.0128439935,
-0.0195146203,
0.0645961314,
0.0066914554,
0.0197426621,
0.0342096463,
-0.0073039038,
0.0155705949,
-0.0394333079,
-0.0232584644,
0.0084693003,
-0.0170470439,
0.0255962331,
0.0460507683,
-0.0081836032,
-0.0416557491,
0.0070440057,
-0.1205514148,
-0.0081436373,
-0.009945062,
0.0088564605,
0.0271307491,
0.0482224375,
-0.0148744974,
0.0000622844,
-0.0428845771,
-0.0122556202,
0.0343968719,
0.0585415475,
-0.0243765097,
0.0474030785,
-0.0249516778,
-0.0194315445,
-0.0059412108,
-0.0162745751,
0.0360892899,
-0.0025221363,
-0.0278572794,
-0.0679342672,
0.0498561822,
-0.043878179,
0.0224382877,
-0.0174777359,
-0.0212842952,
0.0192119293,
-0.063081257,
0.0337528735,
-0.0331107564,
0.0132123139,
-0.0376292244,
0.0450096093,
-0.0396113694,
-0.1058222204,
0.0025171582,
-0.0515180379,
-0.0319801383,
0.0103434091,
0.0053779706,
0.0318269283,
0.0124345897,
0.0217406098,
-0.0133290691,
-0.0249009356,
-0.0125869196,
-0.0111481473,
-0.0256685391,
-0.0461237878,
-0.043110095,
-0.0106634861,
0.0608720519,
-0.009262803,
-0.0342127755,
0.0612110198,
-0.042587243,
0.0067321137,
-0.0799423084,
0.0235562809,
-0.0138599006,
0.0398279428,
-0.0229114965,
0.0444921032,
-0.0802844092,
-0.0027219283,
-0.0144360904,
0.006875094,
-0.0399185047,
-0.0617473051,
0.0618647151,
-0.0282961484,
0.0329016037,
-0.0040209871,
-0.0170103349,
-0.0603327006,
-0.0040509016,
-0.0058663618,
-0.0172788259,
0.0138709098,
-0.0076837372,
0.011354954,
0.0414493456,
0.0300320759,
0.0479523577,
-0.0227355752,
-0.014274844,
0.0025218858,
0.0248483457,
0.0655927584,
-0.0207253676,
-0.0076763914,
-0.0024126645,
0.0197972916,
0.0787797421,
0.0730560198,
-0.0389293134,
-0.0066452632,
-0.0374590345,
0.0380490124,
-0.0069543305,
-0.0162640978,
0.023976963,
0.0119929677,
-0.0656463504,
-0.0038844445,
0.0996900052,
0.0549308285,
-0.0637957603,
-0.0205954202,
-0.0459484085,
0.0005330361,
0.0216427855,
0.0659442768,
-0.1050349697,
0.0080497423,
0.00960356,
0.0448588021,
-0.0079016155,
0.0042107566,
-0.0182904024,
-0.1564056724,
0.0552259982,
0.0011404298,
-0.0546789803,
-0.0319004506,
-0.0305768009,
0.0078189429,
0.0558042824,
0.0216819234,
0.0619726442,
0.0096604563,
-0.0006312786,
-0.0029941068,
0.0041789729,
-0.0550165102,
0.0095212273,
-0.0180613734,
0.0636873469,
0.0664822385,
0.0144613283,
0.0314944983,
-0.0215158537,
0.0372180007,
-0.0342946835,
0.0330655053,
-0.0464297719,
0.0026509694,
0.0182096977,
0.0051521892,
-0.0092264628,
0.0018066091,
0.0021459698,
-0.0179396011,
0.0136631168,
0.0663997754,
-0.0267616324,
0.0190911777,
-0.0499249436,
-0.0007511986,
-0.0047929543,
0.0572583526,
0.0197700951,
-0.0133349588,
-0.0122705717,
-0.0110485591,
-0.0049935775,
-0.0268978123,
-0.0110908495,
0.0076419115,
-0.021805916,
-0.0254615769,
-0.0706956536,
-0.0096895667,
0.0348800644,
0.0418886766,
0.0205504168,
0.0472044535,
0.0146883577,
-0.0172091592,
-0.0713865161,
0.0094321994,
-0.0372225232,
-0.0198874958,
-0.0353308693,
-0.0081241261,
0.0366931669,
-0.0230659004,
-0.0141384229,
0.0238333903,
-0.0043004267,
-0.0861168504,
-0.0090304101,
0.0346259251,
0.0293543097,
0.0647693723,
-0.027860878,
-0.0238609463,
0.0439796075,
0.0471973903,
-0.0663968697,
-0.0003464808,
-0.0202299636,
0.0288077053,
-0.0113066612,
0.0024311519,
-0.0989247337,
0.0111415712,
-0.0203560665,
0.0258271825,
-0.0025729667,
-0.0043408247,
0.0085652154,
0.0713139325,
0.0341864415,
-0.0075374488,
0.0088880667,
-0.0196423456,
0.0479484759,
-0.0690646544,
-0.0246179327,
-0.0168412514,
0.022816211,
-0.0569371805,
-0.0255857315,
0.019596681,
0.0094280746,
-0.0301634632,
0.0633058771,
-0.0147443907,
-0.0059465785,
-0.0139320698,
-0.0014723142,
0.0583852045,
-0.0096462881,
-0.029072063,
-0.0199015867,
0.0279636066,
0.0295031369,
-0.056531027,
0.0397870466,
0.009460479,
-0.0275441501,
0.0024623161,
0.0344736911,
-0.054105781,
-0.019550385,
-0.0253522545,
0.0215875804,
0.0066000521,
0.0008322482,
0.068355225,
-0.0264257658,
0.0097208871,
-0.0367548764,
-0.002426065,
0.0376425199,
-0.0433026068,
0.0183093138,
0.0397978984,
-0.0246757269,
0.0179340169,
0.0699764267,
-0.0607674345,
0.019104138,
-0.0398539156,
-0.0268638711,
-0.0500343964,
-0.0294720847,
-0.0104194721,
0.0381290279,
-0.0256431513,
-0.0331291184,
-0.0616440102,
-0.0516020432,
0.00005679,
0.0072115269,
-0.00284541,
-0.0101927035,
0.0273224097,
0.0780971497,
0.0351850316,
-0.0273648016,
-0.0521573834,
0.0219698064,
-0.0449534394,
-0.0284044519,
0.0092608826,
-6.746242895e-33,
-0.0271422099,
-0.0261024944,
0.025402965,
-0.0392854512,
-0.0102279121,
0.0041634333,
-0.047061801,
-0.0121092079,
0.0268801674,
-0.0477079563,
-0.0226580091,
0.0006902176,
0.0279346425,
0.0094562247,
0.0170620065,
0.0190349314,
0.0422106124,
-0.002110604,
0.0015888961,
0.0561615936,
0.0864945427,
0.0096986201,
0.0144095533,
-0.0479719006,
0.0190203171,
-0.0301713757,
0.0147237424,
0.0103189638,
0.0182861648,
0.0751831681,
-0.0224162322,
0.0548928007,
0.0199134629,
-0.0355979912,
-0.0168832596,
0.0118780686,
-0.0512805469,
-0.0144038917,
0.051218044,
-0.0017155021,
-0.0073946989,
-0.0157693885,
0.0101004625,
-0.0169446617,
-0.0115092006,
0.0691855475,
0.0033346326,
-0.0322575495,
-0.0251190607,
-0.0628492162,
-0.0475051366,
0.0141738988,
0.0251561832,
0.0396622904,
0.0045009539,
0.0743780211,
0.0167229455,
0.0279598571,
-0.0674894527,
0.0524643548,
0.0116103776,
0.0690593496,
0.0588053316,
0.0077807736,
0.0101012681,
0.0245513469,
-0.007612363,
-0.0272871125,
-0.0299889874,
0.0179317798,
0.032561399,
0.0239127278,
-0.00737637,
0.059123721,
0.0237789471,
-0.0378619954,
-0.0306128468,
0.0418400094,
0.0084655592,
0.0300127659,
0.0893876106,
-0.0197569784,
-0.0534132868,
-0.0334776677,
-0.0162009578,
0.0132182641,
-0.0661982074,
-0.0556728803,
0.0478245914,
0.0368229188,
-0.0510426387,
-0.0396034755,
0.0333278589,
-0.0357257724,
-0.0342323072,
0.0387405455,
-0.0366008244,
0.00242039,
-0.0269261058,
0.0080481824,
-0.000630837,
-0.0324593671,
0.0398942195,
-0.0022674003,
-0.0260846391,
0.0266616642,
0.0186686646,
0.0244124718,
-0.0849244222,
0.0248906612,
-0.0326740257,
0.0584145598,
-0.005229997,
-0.0175301041,
0.0270631518,
0.0108587127,
0.0109917251,
0.0517884977,
0.0397512801,
0.0273322482,
-0.0069056586,
-0.0116844885,
0.0055200192,
0.0093629612,
-0.0220834687,
0.0282517448,
0.001706932,
0.0192123391,
0.0391184948,
-0.0054557007,
-0.0179787725,
0.0173272919,
3.016e-7,
0.0150583377,
0.0421491265,
0.0266560707,
-0.0229387097,
0.0113495858,
-0.0057640127,
0.028045699,
-0.0038954206,
0.025942022,
0.0078293113,
-0.0175236613,
-0.0120674875,
0.0275843199,
0.0473542064,
-0.1263224185,
0.0267978385,
-0.0708054528,
-0.0078765368,
-0.0702249184,
-0.0121970735,
0.0342750624,
0.0724602118,
-0.0234295577,
-0.0082670553,
-0.0145580117,
-0.0370962955,
0.0130019551,
0.016084902,
-0.0444233231,
0.0189467575,
0.0460248478,
0.0350613706,
0.0184032433,
-0.0154223703,
-0.0185407829,
0.0378665924,
-0.0558999367,
0.0235113055,
0.0193114709,
-0.0624979287,
0.0493519045,
-0.0383350588,
0.0002927717,
-0.0471554622,
0.0801726952,
-0.0174881946,
0.0163121689,
-0.0484746136,
-0.0222988408,
-0.0162320919,
-0.0406789482,
-0.016872745,
-0.019525338,
0.0399871878,
-0.0043723327,
-0.0349466987,
0.0075294143,
-0.083627373,
-0.0150103085,
0.0070853722,
-0.0270155687,
-0.0704343095,
-0.0185325984,
0.0314170644,
-0.028397372,
-0.0373842828,
0.0248777755,
3.175464368e-34,
0.0519793071,
0.0235727075,
-0.0040517421,
0.0188505966,
0.0131737161,
-0.021137068,
-0.0429068506,
0.0264152605,
0.0138145732,
-0.0832075849,
-0.0442692377
] |
http://arxiv.org/abs/2405.15683v1
|
http://arxiv.org/pdf/2405.15683v1
|
VDGD: Mitigating LVLM Hallucinations in Cognitive Prompts by Bridging the Visual Perception Gap
| 2024-05-24T16:21:59 | 2024-05-24T16:21:59 |
Sreyan Ghosh, Chandra Kiran Reddy Evuru, Sonal Kumar, Utkarsh Tyagi, Oriol Nieto, Zeyu Jin, Dinesh Manocha
|
cs.CV
|
cs.CV, cs.AI, cs.CL
|
Recent interest in Large Vision-Language Models (LVLMs) for practical
applications is moderated by the significant challenge of hallucination or the
inconsistency between the factual information and the generated text. In this
paper, we first perform an in-depth analysis of hallucinations and discover
several novel insights about how and when LVLMs hallucinate. From our analysis,
we show that: (1) The community's efforts have been primarily targeted towards
reducing hallucinations related to visual recognition (VR) prompts (e.g.,
prompts that only require describing the image), thereby ignoring
hallucinations for cognitive prompts (e.g., prompts that require additional
skills like reasoning on contents of the image). (2) LVLMs lack visual
perception, i.e., they can see but not necessarily understand or perceive the
input image. We analyze responses to cognitive prompts and show that LVLMs
hallucinate due to a perception gap: although LVLMs accurately recognize visual
elements in the input image and possess sufficient cognitive skills, they
struggle to respond accurately and hallucinate. To overcome this shortcoming,
we propose Visual Description Grounded Decoding (VDGD), a simple, robust, and
training-free method for alleviating hallucinations. Specifically, we first
describe the image and add it as a prefix to the instruction. Next, during
auto-regressive decoding, we sample from the plausible candidates according to
their KL-Divergence (KLD) to the description, where lower KLD is given higher
preference. Experimental results on several benchmarks and LVLMs show that VDGD
improves significantly over other baselines in reducing hallucinations. We also
propose VaLLu, a benchmark for the comprehensive evaluation of the cognitive
capabilities of LVLMs.
|
VDGD: Mitigating LVLM Hallucinations in Cognitive Prompts by Bridging the Visual Perception Gap: Recent interest in Large Vision-Language Models (LVLMs) for practical
applications is moderated by the significant challenge of hallucination or the
inconsistency between the factual information and the generated text. In this
paper, we first perform an in-depth analysis of hallucinations and discover
several novel insights about how and when LVLMs hallucinate. From our analysis,
we show that: (1) The community's efforts have been primarily targeted towards
reducing hallucinations related to visual recognition (VR) prompts (e.g.,
prompts that only require describing the image), thereby ignoring
hallucinations for cognitive prompts (e.g., prompts that require additional
skills like reasoning on contents of the image). (2) LVLMs lack visual
perception, i.e., they can see but not necessarily understand or perceive the
input image. We analyze responses to cognitive prompts and show that LVLMs
hallucinate due to a perception gap: although LVLMs accurately recognize visual
elements in the input image and possess sufficient cognitive skills, they
struggle to respond accurately and hallucinate. To overcome this shortcoming,
we propose Visual Description Grounded Decoding (VDGD), a simple, robust, and
training-free method for alleviating hallucinations. Specifically, we first
describe the image and add it as a prefix to the instruction. Next, during
auto-regressive decoding, we sample from the plausible candidates according to
their KL-Divergence (KLD) to the description, where lower KLD is given higher
preference. Experimental results on several benchmarks and LVLMs show that VDGD
improves significantly over other baselines in reducing hallucinations. We also
propose VaLLu, a benchmark for the comprehensive evaluation of the cognitive
capabilities of LVLMs.
|
[
0.0046603093,
0.041426897,
0.0071374942,
0.0489105061,
-0.0185524616,
0.0411475971,
-0.0125535065,
0.0311265495,
0.0264336746,
-0.0476213731,
-0.0086518098,
-0.0517609529,
0.0282043833,
0.0205253325,
0.0079525346,
-0.0627742037,
0.0201224592,
-0.0738566518,
-0.0136624807,
-0.0069205225,
-0.0073325285,
-0.0140691027,
-0.0030631241,
0.0257161558,
-0.0019451939,
-0.0522337295,
-0.011974927,
0.0195565447,
0.011843754,
-0.0187764801,
-0.0539385937,
0.0564857088,
0.0042826775,
0.0004894123,
0.0000021864,
-0.0364233218,
0.0407940224,
-0.0160000715,
-0.0074398997,
0.0167532805,
0.0081677781,
-0.0497599877,
0.006496178,
-0.0135652432,
-0.0098358216,
-0.0434046052,
0.0892998874,
0.1041802019,
0.0522992238,
0.0218930896,
-0.0108082294,
-0.0212502386,
0.0333636813,
0.0216542035,
0.0606968701,
0.0129480604,
0.0303754825,
0.0269383155,
-0.0195590779,
-0.0237862151,
-0.0352393687,
0.0424994752,
0.0195207447,
0.0037239327,
0.0275650565,
0.0490035601,
-0.0347482674,
-0.0449752398,
-0.0347376093,
0.0630603731,
-0.0576306358,
-0.0323020145,
0.0064286171,
0.0057849507,
0.0283883717,
-0.0160604548,
-0.0190197621,
0.0490524247,
-0.0419129767,
0.0079690265,
-0.0255554412,
0.0191841405,
0.0213227235,
-0.0107140616,
-0.0321964137,
-0.0047043054,
0.0152291507,
-0.0135386856,
0.0195834134,
0.0120601077,
-0.0349696651,
-0.0569785796,
0.0035561535,
-0.004509592,
-0.0253662113,
-0.0315187946,
-0.0547138639,
-0.0268311054,
-0.0097903358,
-0.0545154922,
0.0386424623,
0.0109154396,
-0.0886732861,
0.0591230057,
-0.0314673968,
0.0872170031,
-0.0111608431,
0.0464128815,
-0.0071691051,
0.0276169367,
-0.0145436078,
-0.0190904364,
-0.0579881892,
-0.0191874821,
-0.0302389748,
-0.0396858528,
0.0054555917,
0.0177570153,
-0.0356722251,
0.020258395,
-0.0325328112,
0.0298560373,
0.0014026905,
-0.0376817472,
-0.0217901208,
0.0157867819,
0.0260363761,
-0.0345973372,
0.0040254886,
-0.0865569413,
-0.0045343055,
0.0524540581,
0.0003195021,
-0.0019100588,
-0.0292909592,
0.0325364321,
-0.013072134,
-0.0036262064,
-0.0279300902,
0.003591334,
-0.0130732078,
-0.0322763324,
0.020674644,
-0.049665805,
0.0039613638,
-0.0091229677,
0.0164544955,
-0.0346181169,
0.0287992489,
0.0167582333,
-0.0292095318,
0.1085440591,
-0.042408295,
-0.0065633426,
0.0156298634,
0.0420085229,
-0.0113738198,
0.0518559553,
-0.0199308842,
0.0568145141,
-0.010810799,
-0.0115955807,
0.0658624396,
0.0212829877,
-0.0214416534,
0.0013068089,
-0.0553356111,
-0.0100386254,
0.02377031,
0.0165914223,
0.0544805862,
0.0707581341,
-0.0397358947,
-0.0377500504,
0.0827408582,
0.0582523793,
-0.0237280391,
-0.0444103591,
0.0154237226,
0.0276502836,
0.0009226767,
0.0252956692,
-0.0763757452,
0.0435770005,
-0.0167402141,
0.0014879599,
-0.0067535937,
0.0144297574,
-0.0403745882,
-0.028394755,
-0.0068478365,
-0.0421079062,
0.0332968459,
0.028688265,
0.01313945,
0.0088159507,
-0.0626449436,
0.0069309217,
0.001260602,
-0.0210298002,
0.008031494,
-0.0317080952,
0.0209383275,
0.0897771493,
0.0506531745,
0.0450248048,
0.0403694697,
-0.0286666732,
0.0484823696,
0.0437958278,
0.0139571447,
0.0310223196,
0.0108952839,
-0.063852638,
-0.0451607555,
0.0165777188,
0.0105366465,
0.0122499084,
-0.0371064506,
-0.0103041278,
-0.0087733772,
-0.0135834944,
-0.0367451049,
0.0404936373,
0.0480731614,
-0.0241337102,
-0.0327671021,
0.0119767925,
-0.0034868533,
-0.009240048,
-0.0020303177,
0.0274559278,
0.0177115984,
-0.0407590754,
0.0156069873,
0.0196586028,
-0.0696799532,
-0.003731474,
0.018234089,
-0.0286774524,
0.0063504539,
0.0197791737,
0.0149536589,
-0.0049251285,
-0.0042313472,
-0.0197793506,
0.0012927747,
-0.0464730151,
0.0218306091,
0.0329534747,
0.0314626768,
-0.0087003522,
0.0745208189,
-0.0503114425,
-0.0203624032,
0.0247841701,
0.0224149432,
-0.0079219928,
-0.0386615396,
-0.0345915742,
0.0268184077,
-0.0014564204,
0.0132304663,
-0.0181129146,
-0.013839555,
-0.0417902023,
0.0405414104,
-0.0566810369,
-0.070724383,
-0.0459008403,
-0.0629238561,
-0.001561492,
-0.0189047642,
-0.0076146531,
-0.0460388362,
-0.0319816433,
0.0616725832,
0.0053298851,
0.0206975602,
-0.005005863,
-0.0062518017,
-0.0233507324,
0.0272997264,
-0.0052321889,
0.0414126217,
0.0613770969,
0.0089170672,
-0.0667983368,
-0.0758719146,
0.0484372117,
-0.0030996643,
0.0372487083,
-0.0174760241,
0.0234432723,
-0.009485174,
-0.0145722907,
0.0442134552,
0.0913030058,
0.0192925818,
0.0069571901,
0.0507953018,
0.0013874306,
-0.0202396847,
-0.0338025242,
0.016071301,
-0.0231974013,
0.0022806691,
-0.0306823067,
0.0022077826,
0.0336611941,
0.0063295048,
0.0470375307,
-0.0974456519,
0.0070073581,
-0.011888369,
-0.0786454007,
-0.0735444352,
-0.0091592809,
-0.0040103132,
0.0623857342,
-0.0345456824,
-0.0880677253,
0.0535219945,
-0.0344003402,
-0.0074290694,
0.0411592498,
0.0295630489,
0.0437333845,
0.0612641424,
0.0239516925,
0.0151362242,
0.0319386683,
-0.0150838476,
-0.0285662133,
-0.0153642725,
-0.0279167369,
-0.0239760168,
0.0638057441,
-0.0592976175,
0.0006043135,
-0.0270998701,
-0.0108465655,
-0.0679117143,
-0.0018430349,
-0.0134836212,
-0.0369664915,
-0.0800451562,
0.0170406532,
-0.0343183167,
0.0543470383,
0.0033226067,
-0.0057543609,
-0.0176376421,
-0.0231058821,
-0.0156831648,
0.0092922784,
0.0724270716,
0.0044527217,
0.023063641,
-0.0484686568,
0.0348227844,
0.0971308127,
0.0198314041,
0.0413436703,
-0.0205745064,
0.030242838,
0.0264337659,
0.0548440143,
-0.0216067266,
0.0609142482,
0.0423386022,
-0.0525037907,
-0.0205093678,
0.0758862197,
0.0211374424,
-0.0830578282,
-0.0164462999,
-0.0254607014,
0.0525218137,
0.0078001684,
-0.0140569201,
-0.0877090767,
-0.0220668819,
-0.0194875635,
-0.0044142762,
0.0193217248,
-0.0009916836,
-0.02680072,
-0.047166843,
0.079125382,
0.0426561572,
0.0520930141,
-0.0768371001,
0.0228875186,
-0.0193193723,
0.0190265216,
0.0080174161,
0.0000527527,
-0.0003234317,
-0.00974721,
0.0063705812,
0.0069556413,
-0.0127609707,
0.0352923125,
-0.0322679505,
0.0139328027,
0.0898472592,
-0.0287014581,
0.0309785455,
-0.1073421463,
-0.0029697786,
-0.017337516,
0.0131112114,
0.0015788885,
0.03752755,
-0.0285678767,
-0.0192474034,
0.0307844076,
0.0261675045,
-0.047261361,
-0.0375354066,
0.0155052757,
0.0320770107,
-0.0330666155,
-0.0347310454,
-0.0138994381,
0.0094569512,
0.0041946908,
0.0727893114,
0.0299608801,
-0.0333930179,
-0.0579539798,
-0.0049856957,
0.0515958183,
-0.0298263002,
-0.0551558845,
0.0375285968,
0.0722909942,
-0.0485559553,
-0.0566691682,
-0.0265189521,
0.0106685646,
0.0238560401,
-0.0133260302,
0.0061661238,
-0.0153585915,
0.003530093,
-0.0698165596,
0.0669333264,
-0.0313843824,
0.0141248843,
-0.0513339303,
0.0202041809,
-0.005201756,
-0.0417881161,
0.0048730308,
0.0223523509,
-0.0683331415,
-0.0268272255,
0.0180446077,
0.0421045274,
0.0194843262,
0.0642013326,
-0.0060777315,
-0.0535099283,
0.0303621739,
0.0460447185,
0.043596603,
0.0251125805,
-0.004131496,
-0.0157216694,
-0.001896288,
0.0184714962,
-0.0093784137,
0.0120393783,
0.0227329265,
-0.0015390859,
0.0082447035,
-0.0076707099,
0.0456490815,
0.000790839,
0.0125575652,
-0.0410968475,
0.0262607224,
0.0104310047,
-0.0117412321,
0.0079858955,
-0.0000687631,
-0.0439285822,
0.0044996855,
0.0249656998,
-0.0178696234,
0.0417735092,
-0.021661194,
-0.0019731917,
-0.0098502096,
-0.0616664104,
-0.0426788405,
0.0083486177,
-0.0257511064,
0.0334410071,
-0.0076032742,
-0.0122988243,
0.0098819062,
-0.0130697852,
0.0329371616,
0.0237437505,
0.0830438212,
0.0015631324,
-0.0234747883,
-0.0263929684,
0.0384639166,
-0.0106524974,
-0.0786968991,
-0.0084542697,
0.0468794815,
0.0487840958,
0.0277751107,
0.0237458795,
-0.0443769507,
0.0162228588,
-0.0589252673,
0.0013622125,
0.0566545799,
-0.0552014634,
0.0311050285,
0.0255777966,
-0.0236964002,
-0.035061989,
0.0451213159,
0.0050443429,
0.0120828887,
-0.0097308364,
-0.0042184494,
-0.0429397896,
-0.0356683508,
-0.0126849515,
0.0124951936,
-0.0339413472,
0.0247158911,
-0.0321661942,
-0.040873237,
-0.0147978216,
0.0097303856,
0.0345742851,
0.0068811281,
0.0557672381,
0.0403246656,
-0.0122899571,
-0.0288159158,
0.0152602978,
-0.0203202181,
0.0254961625,
-0.0022881038,
-0.0235805381,
-6.006519728e-33,
0.0021552735,
-0.0219497271,
0.0334242508,
-0.0406967476,
0.0094362944,
0.0025010807,
0.0053509604,
0.0074201208,
0.002100101,
-0.0052164025,
0.0148194237,
0.001573727,
0.0169854127,
0.0364897363,
-0.012885293,
-0.0178754348,
0.0456248485,
-0.0124299889,
-0.0294942632,
0.0035918211,
-0.0117979906,
-0.0006361577,
-0.0011758076,
-0.0669509768,
0.0237391666,
-0.0135522764,
-0.0352052189,
-0.0149149122,
0.0264349282,
0.0551432222,
-0.0767213702,
0.0303930026,
0.0202562753,
0.0051823682,
-0.0585202724,
0.0416200422,
-0.0420747809,
0.0303574204,
-0.000634141,
0.0557724088,
-0.0816504955,
-0.003481277,
0.0195987206,
-0.0149403941,
-0.0179497115,
-0.0240481403,
0.0304398611,
-0.0506594665,
-0.0318808705,
0.031037733,
-0.0144917006,
-0.021142602,
0.0181827284,
-0.0257509723,
0.0669875368,
0.0454393849,
-0.0171499979,
0.0356523544,
-0.0634577572,
0.0396821499,
0.0070188716,
-0.0040041292,
0.1104501858,
0.0309255645,
0.013548349,
0.0567741059,
0.0602589734,
-0.0401092991,
-0.0408275649,
-0.0357255824,
-0.0046108919,
0.0435180999,
0.0254660677,
0.016393913,
0.072496593,
-0.015498762,
-0.0140924985,
-0.0143583314,
-0.0742368773,
0.0338499919,
0.0768781826,
0.0088747814,
-0.0512860455,
-0.0376984291,
-0.0164103191,
0.0086920997,
-0.0491921566,
-0.0198649727,
0.0886167958,
0.0293039028,
-0.0210526809,
-0.0238823164,
0.0593490452,
-0.0391166396,
0.0100264791,
-0.0035336893,
0.0006547286,
0.0492862724,
0.0027422064,
-0.0092536993,
0.0020153637,
-0.0018692362,
0.0669064373,
-0.016452102,
-0.0017779458,
0.0253705252,
0.0144722797,
-0.0078137647,
-0.0509231463,
-0.0059534516,
-0.0436467677,
0.0181587487,
0.040987175,
0.0373856314,
0.0466616638,
-0.0147916228,
-0.0286487211,
0.0083349142,
-0.0259422697,
-0.046211578,
0.013085247,
-0.0015678408,
-0.0411077105,
0.0285450555,
-0.0085894875,
0.0051030242,
-0.0478156507,
-0.0370961353,
0.0330623835,
-0.031571202,
-0.0123948194,
0.0236230977,
2.79e-7,
-0.0509865135,
0.0205849539,
0.0268042032,
0.0026516109,
0.0069057224,
0.0366621055,
0.0374247842,
0.0354854129,
0.0088612419,
-0.0921367258,
-0.0135321282,
0.0449542627,
0.0307098199,
-0.0251635425,
-0.0309577826,
0.0145349903,
0.0052259364,
0.0177470688,
-0.0523890406,
-0.0324857868,
0.0823831484,
0.0364394225,
-0.0341190845,
-0.0119809601,
0.0932586491,
-0.0434779041,
0.0684170052,
-0.0330470055,
0.013592856,
-0.0288865771,
0.0282816645,
0.0152623691,
0.0029247866,
0.0623353273,
-0.0070488392,
-0.064610377,
0.0253737178,
0.0157982409,
-0.0379308462,
-0.0166839082,
-0.0404614359,
0.0305696968,
-0.018424971,
-0.0655817389,
0.062012922,
-0.021891959,
-0.0156474095,
-0.0005574666,
-0.0365652069,
-0.0313457921,
0.0408657454,
-0.0080886222,
-0.0330304503,
-0.0200074445,
0.0477712043,
0.0172810871,
-0.0397314578,
-0.0644027963,
0.0032404189,
-0.0175298844,
-0.0037269481,
-0.0106413458,
-0.0613281764,
0.0240779594,
0.0406253338,
-0.030019071,
-0.0356563181,
2.627154339e-34,
0.0068230196,
0.0091476673,
0.0442294367,
0.0069074477,
-0.0132255107,
-0.0059200237,
0.0384528488,
0.0301141497,
-0.0176575463,
-0.0430112518,
-0.0130470637
] |
http://arxiv.org/abs/2405.15640v1
|
http://arxiv.org/pdf/2405.15640v1
|
GECKO: Generative Language Model for English, Code and Korean
| 2024-05-24T15:30:41 | 2024-05-24T15:30:41 |
Sungwoo Oh, Donggyu Kim
|
cs.CL
|
cs.CL, cs.AI
|
We introduce GECKO, a bilingual large language model (LLM) optimized for
Korean and English, along with programming languages. GECKO is pretrained on
the balanced, high-quality corpus of Korean and English employing LLaMA
architecture. In this report, we share the experiences of several efforts to
build a better data pipeline for the corpus and to train our model. GECKO shows
great efficiency in token generations for both Korean and English, despite its
small size of vocabulary. We measure the performance on the representative
benchmarks in terms of Korean, English and Code, and it exhibits great
performance on KMMLU (Korean MMLU) and modest performance in English and Code,
even with its smaller number of trained tokens compared to English-focused
LLMs. GECKO is available to the open-source community under a permissive
license. We hope our work offers a research baseline and practical insights for
Korean LLM research. The model can be found at:
https://huggingface.co/kifai/GECKO-7B
|
GECKO: Generative Language Model for English, Code and Korean: We introduce GECKO, a bilingual large language model (LLM) optimized for
Korean and English, along with programming languages. GECKO is pretrained on
the balanced, high-quality corpus of Korean and English employing LLaMA
architecture. In this report, we share the experiences of several efforts to
build a better data pipeline for the corpus and to train our model. GECKO shows
great efficiency in token generations for both Korean and English, despite its
small size of vocabulary. We measure the performance on the representative
benchmarks in terms of Korean, English and Code, and it exhibits great
performance on KMMLU (Korean MMLU) and modest performance in English and Code,
even with its smaller number of trained tokens compared to English-focused
LLMs. GECKO is available to the open-source community under a permissive
license. We hope our work offers a research baseline and practical insights for
Korean LLM research. The model can be found at:
https://huggingface.co/kifai/GECKO-7B
|
[
0.0389453657,
0.0434270203,
-0.0319635496,
0.0328947306,
0.0148331067,
0.0139579866,
-0.0051584993,
0.0042917212,
0.0173196662,
-0.0048704855,
-0.0588054471,
-0.016637899,
-0.0099687604,
0.029247012,
0.0493306778,
-0.0787758231,
0.011627079,
-0.0626349449,
0.0149493972,
-0.0250600111,
0.0004761196,
0.0109968251,
0.035220854,
0.0605152473,
-0.0047409865,
-0.0315723531,
0.0273726285,
0.0147165777,
0.0103161531,
-0.0625620335,
0.0240189079,
0.0132008027,
-0.0202885233,
0.0329633057,
0.0000019689,
-0.043716304,
0.0319361128,
-0.0322457366,
0.0086186817,
0.0273256805,
0.0636296794,
-0.0134561369,
-0.0406769291,
0.0327461436,
-0.0199278183,
-0.0639093816,
0.023975905,
0.0100011155,
0.0297246128,
0.0772163495,
0.0003667664,
-0.0750729963,
0.0340509452,
-0.0205377992,
-0.02546525,
-0.0167273022,
-0.0030328515,
0.0134908305,
0.0096952161,
0.0156839956,
-0.0250155274,
-0.0102799153,
0.0353879519,
-0.027101595,
0.0110965269,
0.0371280313,
-0.043379277,
-0.032045804,
0.0059184097,
0.0638968945,
-0.03864263,
-0.0462765358,
0.0194570813,
-0.002967312,
0.0092053516,
-0.0686458796,
-0.024636576,
0.0509839952,
0.0096568642,
0.0288223382,
0.0305199567,
-0.0342928246,
0.0593310073,
-0.016854573,
-0.022556046,
0.0444014445,
0.0495998636,
-0.006296007,
-0.0439529605,
-0.0097714607,
-0.0012273493,
-0.0887198523,
-0.0012345476,
0.035980165,
-0.045511201,
0.0296388194,
-0.0349797197,
-0.0190215893,
0.047396902,
-0.0324734412,
0.0206974111,
0.0401978232,
-0.0436293371,
-0.0035991359,
-0.0554634333,
0.0488481633,
-0.0124229435,
0.022973299,
-0.0722860396,
-0.0095953876,
-0.0218968503,
-0.0054589463,
-0.0641127676,
0.0335508361,
-0.0220807046,
-0.0017866675,
-0.043798659,
0.0628672913,
0.0073501677,
0.044594951,
0.0252134036,
-0.0312585831,
-0.0611179397,
0.0121621154,
0.0054117567,
-0.0001333038,
-0.0430930704,
0.005437694,
0.0083950013,
-0.0623166189,
0.0378414132,
0.0543467887,
-0.0196961071,
-0.0401347503,
-0.0046623875,
0.0414028354,
0.0255786758,
-0.0039692754,
-0.0526687242,
0.009144647,
-0.0229460876,
-0.0340133756,
-0.0061571011,
-0.1181351766,
-0.0069678249,
0.0071483222,
0.0303181931,
-0.0116174836,
0.0165362414,
-0.0030443254,
0.0357658304,
0.0152041251,
-0.0007991822,
0.0056586256,
0.0236833692,
-0.0226766299,
0.021121785,
0.0843178928,
0.0489506163,
0.0160654411,
0.042176988,
-0.0218978468,
-0.0216043591,
0.0774373561,
-0.0001281846,
0.0243417844,
-0.0009127044,
0.0027530049,
0.0534480885,
0.0252659433,
0.041614607,
0.053209506,
-0.0084054526,
0.033870317,
0.0029820679,
0.0315205492,
0.000966356,
0.0414275378,
0.0909102708,
0.0054055382,
0.0164463725,
-0.0272240601,
0.0010972479,
-0.0073135504,
-0.0157009251,
0.0294789225,
-0.0236891992,
0.0608418584,
-0.0370678194,
-0.0549469292,
-0.0383939706,
-0.0470416248,
-0.025077261,
-0.0186626222,
0.053719949,
0.0325686745,
-0.0299733467,
0.0584571548,
0.0113216192,
-0.0430409946,
0.0202982277,
-0.047431469,
-0.0113397203,
0.0858159214,
0.0154443998,
0.0286753122,
-0.0415986776,
-0.008999343,
-0.0146883279,
0.0278745797,
0.0477367677,
-0.0362384319,
-0.0175785311,
-0.0396026447,
0.0161060244,
-0.0130283507,
0.0194174144,
0.052259896,
0.0301092844,
0.0188552719,
0.0325251408,
-0.033224944,
-0.0287199244,
0.0365508422,
-0.0441197529,
0.0071226037,
-0.0029260565,
-0.0165770147,
0.0645382106,
0.0272421725,
-0.0180811957,
0.0010514621,
0.0344823115,
-0.049057059,
-0.0098741138,
0.0618601032,
-0.0427964367,
-0.0687721074,
0.0378255732,
-0.0075707748,
0.0072854119,
-0.0184865352,
-0.0448997281,
-0.029366089,
0.0416875221,
-0.0049862368,
0.0626500323,
-0.0669270456,
0.0539604835,
-0.0281850956,
0.0037464346,
-0.008379478,
0.0594601296,
-0.0035061156,
0.0186679102,
0.0140756182,
0.0628235713,
0.0270718262,
-0.0380969234,
-0.0403104201,
0.0453237966,
-0.0019271931,
0.0403145365,
0.0032285214,
0.0261203721,
-0.0457628891,
-0.058612965,
-0.1170368791,
0.0249459408,
-0.0236145761,
0.0408587307,
0.0135902753,
0.0148172602,
-0.0451651216,
0.0014646346,
-0.0300613008,
-0.0000677273,
0.0280133635,
0.0201142989,
0.0102311438,
-0.0239543933,
-0.0149498926,
-0.0328458697,
-0.0028689692,
0.0139536066,
0.0067802309,
-0.0093035754,
0.0011523111,
-0.0374830067,
0.0533515066,
0.0049626697,
-0.0081691518,
0.0029362291,
-0.0591596812,
-0.0016053329,
-0.0268320553,
0.0909265205,
0.0033260148,
-0.0066655939,
0.0038781757,
0.021782726,
-0.0190567672,
-0.0339143276,
-0.0074561653,
0.0086694527,
-0.0624885187,
-0.0652302429,
-0.0367690995,
0.0075070299,
0.0776945353,
-0.0058561815,
-0.0325763784,
-0.0322830975,
-0.0519935153,
-0.0587665811,
0.0014756874,
0.0559220053,
0.0071640741,
-0.037170995,
0.0280308966,
-0.0145310415,
-0.0316307023,
0.0489615873,
-0.0513247363,
-0.011133112,
-0.0471845903,
0.0250937268,
-0.0765056461,
0.0249797869,
-0.0103405351,
0.0109869428,
-0.0308904126,
-0.0438816547,
-0.0103739882,
0.0182666574,
-0.0097296583,
-0.0319250785,
0.0499305092,
-0.0212501045,
0.0160038397,
-0.002994268,
-0.0350316428,
-0.0322144069,
0.0250961725,
-0.0294879321,
0.0074350019,
0.0327971019,
0.0200817995,
-0.0049277027,
-0.0088085532,
-0.0202027988,
0.0159536116,
-0.0013815043,
0.0142863821,
-0.0160408914,
-0.0046889074,
0.0018859647,
-0.0362937935,
-0.0210758131,
-0.043100968,
0.0255813766,
0.0052727428,
0.0383969657,
0.0174483899,
-0.0523236096,
-0.0112075135,
0.0622867309,
-0.016062066,
-0.042717915,
0.0280937515,
0.0659078807,
-0.0390736088,
0.0040052906,
0.1127833053,
0.0154029755,
-0.0439805984,
0.0088245859,
-0.0400837325,
0.0566874556,
0.0468624644,
0.0257530902,
-0.0794045106,
-0.0024304006,
0.0122884577,
-0.0051032836,
-0.0070603336,
0.0128532927,
0.0134948902,
-0.1381849051,
0.0184169821,
0.0174535699,
-0.017125709,
-0.0095777698,
-0.0361348167,
-0.0226411056,
0.0066302088,
0.0154463174,
-0.0046600811,
-0.0288543236,
0.0143290516,
0.0278591495,
-0.0187941864,
-0.046256084,
0.0022153256,
-0.0573200956,
0.0396335758,
0.0571340248,
0.0518861264,
-0.011081839,
-0.0458848514,
0.03844909,
-0.0595015958,
-0.0211027768,
-0.0129127298,
-0.0102960952,
0.0084972223,
-0.0348032303,
0.0206015408,
0.0085703805,
-0.0447852612,
0.0107695078,
0.0344201401,
0.0546312667,
0.0119754691,
-0.0155114951,
0.0032407164,
-0.056386821,
-0.0305576175,
0.0261837374,
0.0257964302,
-0.0347902551,
-0.0350574367,
-0.0496130362,
0.0049380357,
-0.0542363934,
0.0229309089,
0.0156338029,
-0.041386608,
-0.0069145951,
-0.0698357075,
-0.0248907227,
-0.0158292148,
0.0462126844,
0.0295492318,
0.0041534221,
-0.03057619,
-0.020049205,
-0.0145109734,
0.0068705017,
-0.027501272,
0.0180869661,
-0.0818667412,
-0.0360631272,
0.0129961846,
-0.0279264972,
-0.0127455834,
-0.0076566106,
-0.0840628743,
-0.0604135133,
-0.0120229507,
0.0588045307,
0.0097512705,
0.0259603541,
-0.0294827204,
-0.0107919546,
-0.0136788236,
0.0148072019,
-0.0260950103,
0.0038607416,
0.0334599912,
-0.0079209274,
0.0187199991,
0.0354863368,
-0.0555861779,
0.0192058831,
0.0095363818,
0.0398566499,
-0.0004817029,
-0.0113545451,
-0.0149348257,
0.042434983,
-0.030015694,
-0.0367178582,
0.0021845938,
-0.0570926741,
0.0190781187,
-0.0069730091,
-0.0102248164,
-0.0046604061,
0.0210397393,
0.0520217903,
-0.0226004645,
0.0016523678,
0.0309151579,
-0.0173296332,
0.0391536169,
0.0303703733,
0.0497415736,
0.0257593393,
0.0553991646,
0.0548863858,
0.0132162282,
-0.0180304572,
0.0022666489,
0.0043224907,
0.0172418859,
-0.0314702205,
0.0078505594,
-0.0113436263,
0.0011769687,
0.0088226041,
0.0146603966,
0.01379216,
-0.0367236696,
0.0034206971,
0.0277114548,
0.083995156,
0.0265468433,
0.0758651197,
-0.0093852924,
-0.0038092728,
-0.0252535716,
0.0336109661,
-0.0099363038,
-0.0771112442,
0.0199037455,
-0.0170388259,
-0.0261100493,
-0.0357098691,
0.0879567415,
-0.001973995,
0.0514633954,
-0.0087866839,
0.0196207128,
-0.0224277023,
0.0042891139,
0.0259679258,
0.0064445799,
-0.0200098995,
-0.0362766832,
-0.0297689289,
-0.0767799988,
-0.0265145972,
0.0340475738,
0.0169133581,
0.012627135,
0.0010588528,
0.088600345,
0.0166275762,
-0.0299402438,
-0.0639396906,
0.0283100139,
0.0377556868,
-0.0076510999,
0.0210186429,
-5.933851769e-33,
-0.0320914574,
0.0033424334,
0.0142164584,
0.0128091527,
-0.0310571771,
-0.0113574984,
-0.042258922,
-0.0154419448,
-0.0109877437,
-0.0394342057,
-0.022002656,
0.0085120955,
0.0154480273,
0.0255668536,
0.0487914011,
0.0363885462,
0.0563278683,
0.022273453,
-0.0065101548,
0.053942658,
0.0131872809,
0.0390131176,
0.0270178709,
0.0060382313,
0.0009882831,
-0.0120304404,
0.0228515118,
-0.0005473635,
0.0184571911,
0.0820169151,
-0.0296784304,
0.0521616153,
0.0187188797,
-0.0085387472,
-0.0541704372,
-0.0369532332,
-0.0699736699,
-0.0346152708,
0.0379514173,
0.0481922515,
0.0096091051,
-0.0912268907,
0.0389507711,
-0.0222001337,
-0.0385211818,
0.0298938863,
0.036399737,
-0.0116686914,
-0.0142846918,
-0.0237542056,
-0.1015301049,
0.0338430777,
0.0050119841,
0.044181481,
0.0233908109,
-0.0464899726,
-0.0303204004,
0.034448754,
-0.0100938771,
0.0217043087,
-0.0017861092,
0.0544874072,
0.0866266936,
0.019079404,
0.0111029856,
0.0211403165,
0.0239147414,
0.0287119057,
0.0199459419,
0.0117278593,
0.0355344526,
0.030356599,
-0.0261978321,
0.0350547098,
0.0191926192,
-0.0696671009,
-0.0781221688,
-0.0081398282,
0.0171009488,
0.0378451049,
0.0089318668,
0.0114359921,
0.0088464376,
-0.0256169289,
-0.0451448001,
-0.044871971,
0.0009545522,
-0.0594007969,
0.0227530338,
0.0175230037,
-0.009251466,
-0.0316531435,
0.0168380253,
-0.0322884209,
-0.0430878364,
0.0328301713,
-0.0188429523,
0.0198165681,
-0.0202491321,
0.0630614161,
-0.072467491,
-0.0631385893,
0.0001484163,
-0.0213000141,
0.0263219252,
0.0087523246,
0.0371227823,
-0.0137949865,
-0.0521380752,
0.0281340387,
-0.0588681698,
-0.0181917418,
0.006222832,
0.0187840946,
0.0353880078,
0.0099060321,
0.0100199822,
0.0391549021,
0.0327441432,
0.0065123211,
-0.0685824156,
-0.0064949025,
0.0017696344,
0.0395830572,
-0.0292891115,
0.0190431662,
-0.0488092043,
0.0591968857,
0.0238538571,
-0.0586051568,
-0.0241253823,
0.027681252,
2.674e-7,
0.0340549536,
0.0246013943,
0.0366112962,
0.008556297,
0.0310872141,
0.0059955809,
-0.0295873825,
0.0107011413,
0.0411031321,
0.0436887853,
-0.030472964,
0.0090706339,
0.0176987872,
0.0546847545,
-0.1235435531,
0.0741881952,
-0.020071473,
0.0171144176,
-0.0735205784,
0.007019266,
0.0318463072,
0.0914293528,
-0.0080416724,
-0.0046206638,
-0.0112389475,
-0.038497813,
0.0499810576,
0.0293802787,
-0.0083804801,
0.0453126207,
-0.0020814163,
0.0297462475,
-0.0049882829,
-0.0348865725,
-0.0089391563,
0.0088009983,
-0.001427529,
0.0605766922,
0.0237184931,
0.0208923854,
0.0602831207,
-0.0199635588,
-0.0348251387,
-0.0872543305,
0.064603202,
0.0074831536,
0.0066551855,
-0.0117636155,
-0.0211023632,
-0.0051413858,
-0.0302064829,
-0.0391951054,
-0.0076452815,
0.0570975058,
0.0251919832,
-0.0291257724,
-0.0268878136,
-0.0213746894,
-0.0418847352,
0.0040554898,
-0.0428045131,
-0.0754478276,
-0.0382469222,
0.0456091054,
-0.0038753054,
-0.0126361335,
0.0472743735,
2.280825235e-34,
0.0377298109,
-0.0095838401,
-0.0052139843,
0.0220846869,
0.0081333788,
-0.0330885947,
-0.0156816263,
0.0275380444,
-0.0268836841,
-0.0501218624,
-0.0313031077
] |
http://arxiv.org/abs/2405.15638v1
|
http://arxiv.org/pdf/2405.15638v1
|
M4U: Evaluating Multilingual Understanding and Reasoning for Large Multimodal Models
| 2024-05-24T15:25:28 | 2024-05-24T15:25:28 |
Hongyu Wang, Jiayu Xu, Senwei Xie, Ruiping Wang, Jialin Li, Zhaojie Xie, Bin Zhang, Chuyan Xiong, Xilin Chen
|
cs.CV
|
cs.CV, cs.CL
|
Multilingual multimodal reasoning is a core component in achieving
human-level intelligence. However, most existing benchmarks for multilingual
multimodal reasoning struggle to differentiate between models of varying
performance; even language models without visual capabilities can easily
achieve high scores. This leaves a comprehensive evaluation of leading
multilingual multimodal models largely unexplored. In this work, we introduce
M4U, a novel and challenging benchmark for assessing the capability of
multi-discipline multilingual multimodal understanding and reasoning. M4U
contains 8,931 samples covering 64 disciplines across 16 subfields in Science,
Engineering, and Healthcare in Chinese, English, and German. Using M4U, we
conduct extensive evaluations of 21 leading Large Multimodal Models (LMMs) and
Large Language Models (LLMs) with external tools. The evaluation results show
that the state-of-the-art model, GPT-4o, achieves only 47.6% average accuracy
on M4U. Additionally, we observe that the leading LMMs exhibit significant
language preferences. Our in-depth analysis indicates that leading LMMs,
including GPT-4o, suffer performance degradation when prompted with
cross-lingual multimodal questions, such as images with key textual information
in Chinese while the question is in German. We believe that M4U can serve as a
crucial tool for systematically evaluating LMMs based on their multilingual
multimodal reasoning capabilities and monitoring their development. The
homepage, codes and data are public available.
|
M4U: Evaluating Multilingual Understanding and Reasoning for Large Multimodal Models: Multilingual multimodal reasoning is a core component in achieving
human-level intelligence. However, most existing benchmarks for multilingual
multimodal reasoning struggle to differentiate between models of varying
performance; even language models without visual capabilities can easily
achieve high scores. This leaves a comprehensive evaluation of leading
multilingual multimodal models largely unexplored. In this work, we introduce
M4U, a novel and challenging benchmark for assessing the capability of
multi-discipline multilingual multimodal understanding and reasoning. M4U
contains 8,931 samples covering 64 disciplines across 16 subfields in Science,
Engineering, and Healthcare in Chinese, English, and German. Using M4U, we
conduct extensive evaluations of 21 leading Large Multimodal Models (LMMs) and
Large Language Models (LLMs) with external tools. The evaluation results show
that the state-of-the-art model, GPT-4o, achieves only 47.6% average accuracy
on M4U. Additionally, we observe that the leading LMMs exhibit significant
language preferences. Our in-depth analysis indicates that leading LMMs,
including GPT-4o, suffer performance degradation when prompted with
cross-lingual multimodal questions, such as images with key textual information
in Chinese while the question is in German. We believe that M4U can serve as a
crucial tool for systematically evaluating LMMs based on their multilingual
multimodal reasoning capabilities and monitoring their development. The
homepage, codes and data are public available.
|
[
0.0033506753,
0.0543288253,
-0.0124973655,
0.075848721,
-0.0307567492,
0.0265239589,
0.0784408078,
-0.0081373733,
0.0068426607,
-0.0526354499,
-0.0620641559,
-0.0433923677,
-0.0050876569,
0.0125199603,
0.0499396324,
-0.0579342358,
0.0482460745,
-0.0514080562,
-0.0270680543,
-0.0295121539,
0.0029505822,
0.0033071798,
-0.0314082056,
0.019603245,
0.0113742957,
-0.0656400621,
0.0010758907,
-0.0035092176,
0.0231926721,
-0.0407262743,
-0.0098726852,
0.0508083664,
0.0316143148,
0.0323584899,
0.000002272,
-0.0624155179,
-0.0125033911,
-0.012398582,
0.0309571922,
-0.0015592969,
0.0164433718,
-0.0196465924,
0.0129651502,
-0.031622082,
-0.0142482854,
0.0166918486,
0.0143603822,
0.0488566495,
0.0460319445,
0.0228516273,
-0.0088178469,
-0.0397706367,
0.0064710248,
0.0082649216,
0.0078833941,
-0.0502271391,
-0.0171727072,
0.0593494251,
0.0200134479,
0.0240122974,
-0.0558823794,
0.0513874739,
0.0326285698,
0.012842115,
-0.0483915992,
0.0447524898,
0.0081732478,
-0.035689041,
0.0103458706,
0.0876432434,
0.0042907773,
0.0123692593,
-0.0356826261,
0.0229284205,
-0.0206251796,
0.0368511826,
-0.0881871879,
0.0460586958,
-0.025064107,
-0.0386697613,
-0.0146077434,
-0.0216495767,
0.0234492142,
0.0267337188,
-0.0240024701,
0.0650915429,
0.0170869436,
0.0034382022,
0.0524271205,
-0.0198859423,
-0.0499672182,
-0.0141378017,
-0.0114770224,
0.0165004432,
-0.0374411568,
0.0188379139,
0.0167649519,
-0.020357389,
0.0252346192,
-0.0276505109,
0.0096253157,
0.0503508113,
-0.0487088896,
0.0710855275,
-0.0314760283,
0.0598610565,
-0.0428790003,
0.0476918854,
-0.012878661,
-0.0320013538,
-0.0504131727,
-0.0138992574,
0.0404451862,
0.0030086781,
-0.0202801339,
-0.0343236662,
-0.0272287708,
0.0187136382,
-0.004535344,
0.0032539661,
-0.0200658515,
-0.0059642573,
-0.005585948,
-0.0150458524,
-0.0579141416,
0.0715629831,
-0.0507690795,
0.0119339842,
-0.0043990128,
-0.017598927,
0.0240805838,
0.042008251,
-0.0316946879,
-0.0496551245,
0.0164450556,
0.0987346843,
-0.030903155,
-0.0023812461,
-0.015188803,
-0.0619719662,
0.0119470395,
-0.0938304961,
0.0756511688,
-0.0521526225,
0.0331342444,
0.0355328172,
0.0134969056,
-0.027611522,
-0.0075892704,
0.0178349297,
-0.0095719332,
0.077087447,
0.0109810904,
-0.002890676,
0.032087326,
0.0249464847,
-0.0545535497,
0.07763879,
-0.0095006898,
0.0094369315,
0.0594071113,
-0.0109751578,
-0.0388227887,
0.0286419131,
-0.0192293357,
-0.0082424451,
-0.007849101,
0.0132779861,
0.0265870839,
0.0420311205,
0.0186690968,
0.016018955,
-0.0247140042,
-0.0424572527,
0.0528803356,
0.0594105907,
-0.0063678063,
-0.0247034617,
0.0161094349,
-0.0491150245,
-0.0051677753,
0.0089045241,
-0.0418810323,
0.0809824914,
-0.0356683806,
-0.0115450462,
0.0504933596,
0.0521859266,
-0.0274657123,
-0.0203448813,
-0.0149370348,
0.0172601473,
0.0095136082,
0.0312016699,
0.0561226383,
-0.0353545323,
-0.0067505552,
0.0140151018,
0.0190522969,
-0.0372016057,
-0.0082188295,
-0.0191377662,
0.0248744655,
0.0554608516,
0.0464224853,
-0.0031712514,
-0.0484797284,
-0.0024860688,
0.0422011428,
0.0415051393,
0.0350215845,
-0.0001201629,
0.0162467323,
-0.0544519275,
-0.0338826142,
-0.0373337902,
-0.0321622938,
0.0300054345,
-0.0179183558,
0.0079609388,
-0.0150125995,
-0.0717292652,
0.0174944941,
0.0163788255,
-0.0439931676,
0.0327258855,
-0.0579169206,
0.0126302252,
0.0233865157,
-0.0021593061,
0.0075693196,
0.0163113028,
-0.0073060365,
-0.0423962399,
0.000475437,
0.0473150797,
-0.0347435437,
-0.0254848506,
0.042616941,
-0.0886799097,
-0.0431715623,
0.0446820371,
-0.0163059123,
-0.0145624317,
0.0250641927,
0.0129666114,
-0.0119395079,
-0.056812197,
-0.0212168153,
-0.0125736892,
0.0182723608,
-0.0243702009,
0.0648309812,
-0.0237124227,
0.0531909876,
0.0130392136,
0.0158860851,
0.0759040564,
-0.0211655553,
-0.0427916422,
0.0551844575,
0.0038568643,
0.0043165255,
-0.0147596058,
0.0015743007,
-0.050229989,
0.0177543964,
-0.0913779885,
-0.0013856752,
-0.0107963877,
-0.0215478782,
0.0111302277,
-0.0235941056,
0.0090465853,
-0.0067239394,
-0.027011998,
0.012245317,
0.0173550323,
0.0536416769,
0.0145034669,
-0.019706903,
0.0519177318,
0.0052385642,
-0.0089630084,
-0.055149734,
0.0724974871,
0.0126344403,
0.030729929,
-0.0224309526,
-0.0291181263,
-0.0770573467,
0.0196513925,
-0.0237095933,
-0.0235546865,
0.0558603108,
-0.0053307265,
0.0349952616,
0.0754868463,
-0.0117389504,
0.0028770925,
0.0507550351,
-0.0188457519,
-0.0399042293,
0.0312407631,
0.0327110738,
0.008530084,
-0.0149574522,
-0.0399945602,
-0.0347721092,
0.0493165553,
0.0016312978,
-0.0105109932,
-0.0282158311,
0.0146760857,
-0.0190178566,
-0.0132600265,
-0.0725927576,
0.0223574247,
0.0270245019,
0.0796240643,
-0.0219719578,
0.0049345586,
0.0647955611,
-0.0340398848,
-0.0009234575,
-0.0381511785,
0.0432137735,
-0.037962649,
0.0780629218,
-0.0172107369,
0.0517193265,
-0.0369779095,
0.0153015358,
-0.0567476116,
0.014503249,
0.0356633738,
-0.0035595994,
0.0379159674,
-0.0482213497,
-0.009212302,
-0.0188112762,
-0.0133664329,
-0.0143567957,
0.0203075726,
-0.033019796,
0.0128519386,
-0.0244466681,
-0.0044589452,
-0.0189845543,
0.0273078065,
0.0636457726,
-0.0202043895,
-0.0156889837,
-0.0080662472,
-0.0106223067,
0.013890299,
0.0303304698,
0.0399852991,
-0.0099007441,
-0.0910682082,
0.0432372242,
0.0347445421,
0.0022701414,
-0.0355571955,
-0.0284302961,
0.0434233472,
0.0273965895,
0.0299558081,
-0.0124448258,
-0.0076606008,
0.0615905672,
-0.0430821702,
0.0307846311,
0.0842186213,
0.0111649148,
-0.052345641,
-0.0120995604,
-0.0407070555,
0.0195596591,
0.0522720888,
0.0217083786,
-0.0707964227,
-0.0661633387,
-0.0116346134,
-0.0146114845,
0.037186563,
-0.0248476136,
0.0102551477,
-0.0905750543,
0.0438885018,
-0.0267940313,
-0.0603137501,
-0.0272117592,
-0.031221237,
-0.0008646892,
0.0248756576,
0.0079078032,
0.0190075953,
0.029229315,
-0.0192331988,
0.0038435187,
-0.0002047225,
-0.0503133982,
0.0764352754,
-0.02526916,
-0.0180590302,
0.0187191013,
0.0295548346,
-0.0098981308,
-0.0770299286,
0.0451517664,
0.0222203992,
0.0058559165,
-0.0228958763,
0.0262733996,
0.0079191951,
-0.0161544848,
-0.0278198775,
0.0117241759,
-0.0073407935,
-0.020360766,
0.0037571792,
0.1097777262,
-0.0216544569,
-0.0088446336,
-0.045802895,
0.0282329377,
-0.0620697886,
0.0595746376,
0.0033481156,
-0.0422791988,
-0.0427912548,
-0.0345476009,
-0.021123996,
-0.0628013387,
-0.0062724631,
0.04899114,
0.0017212299,
0.0050073881,
-0.0486728139,
-0.047799509,
0.0196921751,
0.0054306597,
0.0046094242,
0.0018111883,
-0.0238035433,
0.019468056,
-0.0427031741,
0.044580657,
0.0100893145,
0.0339159817,
-0.0305260047,
0.016604092,
0.0174318179,
-0.044899974,
-0.0252391174,
0.0253592543,
-0.0191184115,
-0.0376756378,
-0.0513729639,
0.027738072,
0.0126466667,
0.0054911016,
0.0354656614,
-0.0428141281,
-0.0225133896,
-0.0355026573,
-0.0040278435,
-0.0161593482,
0.0159194637,
0.0129350862,
-0.0050127823,
0.0121387187,
-0.0243703462,
0.0184789374,
-0.0034049784,
0.0470868088,
-0.0119727142,
0.0092566609,
0.0080198655,
0.0241496097,
0.0137237059,
-0.0072343377,
-0.028038457,
-0.0320817754,
0.0061819302,
-0.0123577416,
-0.0130843455,
-0.0241117422,
0.0094739115,
0.023916617,
-0.0373210534,
0.0223486107,
-0.0134310564,
0.0166356545,
0.0089959139,
-0.0080901738,
0.0240578577,
0.0188375115,
-0.0002805915,
-0.0274108853,
0.0157480501,
-0.0063902689,
-0.0062290845,
-0.0158128869,
0.0657802746,
-0.0230638105,
0.1367869526,
0.0376521051,
-0.0118265739,
0.0062316586,
0.001713698,
-0.0268879849,
-0.0424921066,
0.0374573283,
0.0192347821,
0.0567126088,
0.0281455535,
0.020437615,
-0.0152099859,
0.0021267477,
-0.0097928913,
0.0339334272,
0.023898242,
-0.0336924084,
0.0552390143,
0.0310028922,
0.0342947356,
-0.0126025928,
0.0745787174,
-0.0329413749,
0.0102315331,
-0.0106058009,
-0.0055567808,
-0.0428982377,
-0.0262333397,
0.0360712335,
0.0406451263,
-0.0068116277,
0.0277729705,
-0.0529446714,
-0.0830672085,
-0.025178168,
0.0281374045,
-0.0327394083,
0.0176878572,
0.0139971692,
0.0923066363,
-0.0060519315,
-0.032522101,
-0.0309260655,
-0.0100835264,
-0.0413846113,
-0.0010091119,
0.017477233,
-6.822280481e-33,
-0.0105557134,
-0.0298435576,
0.0463678353,
-0.0366355777,
0.015491887,
-0.0152312629,
-0.0541311502,
0.0083048111,
-0.0508552417,
0.0126050804,
0.0190588553,
-0.0323971696,
-0.0169442352,
0.008894424,
0.0323205926,
0.0082232766,
-0.0193459112,
0.0174313262,
-0.0242084581,
-0.0049250135,
0.0594041944,
0.0099377614,
0.0561089069,
-0.0189291593,
0.027683394,
0.0143110026,
0.0047342484,
0.0188169871,
-0.0472564623,
0.056635581,
-0.0561739877,
0.0263598487,
0.0512480587,
-0.0386476964,
-0.0022515145,
-0.0112148132,
-0.0624797121,
-0.0315044038,
0.041507259,
0.0263215937,
-0.0872811154,
-0.0271639694,
0.0275272597,
-0.0128907692,
-0.0027163241,
-0.05171315,
0.0578928217,
-0.0187477358,
0.0219372362,
0.0210734345,
-0.0615083948,
0.0089133596,
-0.0433090143,
0.0088122739,
-0.0025076647,
0.0654461905,
-0.0110609038,
0.0647038519,
-0.0477414429,
0.0195482373,
-0.0233848598,
-0.0258409809,
0.0965513438,
-0.0091877552,
0.0054294858,
0.0694167614,
0.0037649565,
0.0010106413,
0.0161253102,
-0.0029653748,
0.0539523661,
0.0528513826,
0.0447886251,
0.0551899634,
0.004543609,
0.0150846262,
-0.0722723082,
0.0344226472,
-0.0012166131,
0.0619507693,
0.0389003418,
-0.0040228143,
0.0195578448,
0.0005656665,
-0.0433966927,
0.0244342312,
-0.053931199,
-0.0355568193,
0.0606142171,
-0.0564749539,
-0.0170322508,
0.0034048737,
0.0811088681,
-0.0473841652,
-0.0127973398,
0.0148648107,
-0.003110521,
0.0282854345,
0.0145490356,
0.0504797213,
-0.0210995208,
-0.0228890199,
-0.006362054,
-0.0179702435,
0.0434864573,
-0.0002319052,
0.0279396772,
-0.0053186701,
-0.002016553,
-0.0269722585,
-0.0709015653,
-0.0432474203,
-0.0032455653,
0.0189014934,
0.0161416531,
0.019994339,
-0.045274172,
0.0033626815,
0.0104074599,
0.0023356851,
-0.0559490956,
-0.0527881235,
-0.0249646325,
0.0247947145,
-0.0393933728,
0.0077690301,
-0.0449524075,
-0.0362459831,
-0.0043853051,
-0.0649436116,
-0.0038585479,
0.0621509738,
2.842e-7,
0.0570404865,
0.0477357917,
-0.0403023697,
0.0035679967,
0.0173448008,
0.0291094184,
-0.0587948859,
-0.0044328398,
0.0307124741,
-0.02186835,
-0.038078215,
0.0519988537,
0.0340181291,
0.0595631637,
-0.0207642522,
-0.0047949087,
0.0176230166,
-0.0384182669,
-0.0666434914,
0.0268737786,
0.0548154078,
0.0059181722,
-0.0495547168,
-0.0272412933,
0.0261290483,
-0.0349138714,
0.0818501264,
-0.026516892,
-0.0156692713,
-0.0072629144,
0.0216600951,
0.0131531404,
0.0098637072,
0.0376711451,
0.021242829,
-0.0427988544,
0.0546352752,
0.0118611464,
-0.0187072121,
-0.0473203622,
-0.001276963,
0.0030155173,
-0.0248528998,
-0.0953982323,
0.0808464512,
-0.0036105174,
-0.045372244,
0.0028180238,
-0.0281720962,
-0.0219444819,
0.0341094323,
-0.0279333983,
-0.0308867637,
0.0143223442,
0.0041441796,
-0.0241333544,
-0.004084358,
-0.0447612777,
-0.0069657657,
0.0258220695,
-0.0538707264,
-0.0318767056,
-0.0272934847,
-0.0470837578,
0.0497471951,
-0.0479670279,
-0.054886695,
2.097555793e-34,
0.0377276763,
-0.0053805304,
0.0343853906,
-0.0023667514,
0.0301433038,
-0.0301675461,
-0.038221214,
-0.00591572,
-0.0116375377,
-0.0736358389,
-0.0210253708
] |
http://arxiv.org/abs/2406.00028v1
|
http://arxiv.org/pdf/2406.00028v1
|
Persian Homograph Disambiguation: Leveraging ParsBERT for Enhanced Sentence Understanding with a Novel Word Disambiguation Dataset
| 2024-05-24T14:56:36 | 2024-05-24T14:56:36 |
Seyed Moein Ayyoubzadeh
|
cs.CL
|
cs.CL, cs.LG
|
Homograph disambiguation, the task of distinguishing words with identical
spellings but different meanings, poses a substantial challenge in natural
language processing. In this study, we introduce a novel dataset tailored for
Persian homograph disambiguation. Our work encompasses a thorough exploration
of various embeddings, evaluated through the cosine similarity method and their
efficacy in downstream tasks like classification. Our investigation entails
training a diverse array of lightweight machine learning and deep learning
models for phonograph disambiguation. We scrutinize the models' performance in
terms of Accuracy, Recall, and F1 Score, thereby gaining insights into their
respective strengths and limitations. The outcomes of our research underscore
three key contributions. First, we present a newly curated Persian dataset,
providing a solid foundation for future research in homograph disambiguation.
Second, our comparative analysis of embeddings highlights their utility in
different contexts, enriching the understanding of their capabilities. Third,
by training and evaluating a spectrum of models, we extend valuable guidance
for practitioners in selecting suitable strategies for homograph disambiguation
tasks. In summary, our study unveils a new dataset, scrutinizes embeddings
through diverse perspectives, and benchmarks various models for homograph
disambiguation. These findings empower researchers and practitioners to
navigate the intricate landscape of homograph-related challenges effectively.
|
Persian Homograph Disambiguation: Leveraging ParsBERT for Enhanced Sentence Understanding with a Novel Word Disambiguation Dataset: Homograph disambiguation, the task of distinguishing words with identical
spellings but different meanings, poses a substantial challenge in natural
language processing. In this study, we introduce a novel dataset tailored for
Persian homograph disambiguation. Our work encompasses a thorough exploration
of various embeddings, evaluated through the cosine similarity method and their
efficacy in downstream tasks like classification. Our investigation entails
training a diverse array of lightweight machine learning and deep learning
models for phonograph disambiguation. We scrutinize the models' performance in
terms of Accuracy, Recall, and F1 Score, thereby gaining insights into their
respective strengths and limitations. The outcomes of our research underscore
three key contributions. First, we present a newly curated Persian dataset,
providing a solid foundation for future research in homograph disambiguation.
Second, our comparative analysis of embeddings highlights their utility in
different contexts, enriching the understanding of their capabilities. Third,
by training and evaluating a spectrum of models, we extend valuable guidance
for practitioners in selecting suitable strategies for homograph disambiguation
tasks. In summary, our study unveils a new dataset, scrutinizes embeddings
through diverse perspectives, and benchmarks various models for homograph
disambiguation. These findings empower researchers and practitioners to
navigate the intricate landscape of homograph-related challenges effectively.
|
[
-0.0058619366,
0.0032616172,
0.0063154725,
0.039583303,
-0.0521434657,
0.0336656198,
0.0331095047,
-0.0362555981,
0.0578711294,
-0.0170140173,
-0.0182860848,
-0.0446098112,
0.0243895594,
-0.0241973531,
0.0238907672,
-0.0281831324,
0.0489061177,
0.0004759171,
-0.0026298983,
-0.0015776228,
0.006775321,
0.0183327645,
-0.0112801241,
0.0180079322,
-0.0315332897,
-0.0127063738,
-0.0054433849,
0.0419825688,
0.0333958529,
0.0122012617,
-0.003035747,
0.0354207493,
0.0146560445,
0.1142773181,
0.0000015636,
-0.0741505995,
-0.0361083634,
-0.0247855093,
0.0027782684,
-0.0255265739,
0.0481911302,
-0.029868383,
-0.0074715554,
0.0249572247,
0.0084662773,
-0.0587054342,
0.0611774176,
-0.0105240801,
0.0464277714,
0.0359295309,
-0.0258239694,
-0.080053784,
-0.0203809794,
-0.0196528584,
-0.0146842375,
0.0045368099,
0.0002819056,
-0.0303336047,
-0.0698120296,
-0.0409943163,
-0.043300353,
0.0529841222,
-0.0116923368,
0.0277833119,
0.0084476555,
-0.0183709823,
-0.0553000011,
-0.0134070516,
0.0420306399,
0.0679617152,
-0.0541036837,
-0.0677150413,
0.0123745259,
0.0082913954,
-0.0394143723,
0.0321735144,
-0.015921006,
0.0212087501,
-0.01149063,
0.0350505523,
0.0536088608,
0.0017453542,
0.0479137897,
0.0217927154,
-0.0029903157,
0.0044571739,
0.0469717868,
-0.033459872,
0.0550144054,
0.0051272851,
-0.0360753238,
0.0150359021,
0.0345860757,
-0.0057860939,
-0.0125312405,
0.0298478715,
-0.0247428101,
-0.0533308648,
0.0227936059,
-0.0726096556,
-0.0029741065,
0.0417262465,
-0.0133788111,
0.0463282913,
-0.0481711775,
-0.0112908613,
0.0190291349,
0.0037772905,
-0.0216224808,
-0.0145990709,
-0.0614527017,
0.0211227387,
-0.0653980449,
0.0160219427,
0.0050440147,
-0.0020731741,
-0.0555666201,
-0.0145953065,
-0.0327746496,
0.0811022893,
0.0057843872,
-0.00925868,
-0.0457241684,
0.0269058757,
0.034557946,
0.045954477,
-0.0245926082,
-0.0087732095,
-0.0552946515,
-0.0663557127,
-0.0241885316,
0.0356265455,
0.0041828877,
-0.0159670487,
0.0020835039,
0.0050282977,
-0.0003566862,
-0.0404122435,
0.0066546337,
-0.0256602503,
-0.0029374738,
-0.0315719545,
0.0303731486,
-0.1002421901,
0.0588960648,
-0.0073154126,
-0.000801747,
-0.0623316467,
-0.0111868326,
-0.0014495159,
-0.0067690657,
0.1125531569,
0.0078323158,
-0.0201856308,
0.0322155356,
0.0320910402,
0.0089639993,
0.0602554083,
0.0197901111,
0.0050219405,
0.0858717188,
-0.0001469097,
-0.0489174724,
0.0275629964,
-0.051722087,
0.0224904977,
0.0098301936,
-0.058734607,
-0.0311964974,
0.0810349807,
-0.0081590647,
0.0699661747,
-0.0175263006,
-0.0212206021,
0.0160821434,
0.0507990457,
0.0356813446,
-0.0149753597,
0.0202004723,
0.053149689,
0.0256931148,
0.0363409445,
-0.0057036756,
-0.0536502153,
0.0741280019,
0.0534655377,
-0.0009022985,
0.0061807125,
-0.0224118344,
-0.0068742693,
-0.0330715515,
-0.0796922073,
0.0056193378,
-0.0015059583,
0.0896960273,
-0.0022089754,
-0.0459084958,
0.0592703372,
0.0502104126,
-0.0461887717,
0.0382778682,
-0.0150968106,
-0.0073733856,
0.079660669,
0.022341568,
-0.0455484688,
0.0104096299,
-0.006627121,
-0.0160192661,
0.0239048563,
0.0373073146,
-0.0409143455,
0.0475035235,
-0.0765286088,
-0.0205645394,
0.0014133671,
0.040642716,
-0.0217013564,
-0.0106503312,
0.0068644951,
0.0370047241,
-0.0252288859,
0.0039364947,
0.023935087,
-0.0290363412,
0.001101198,
-0.0235837158,
0.0389650762,
-0.0499498211,
0.0521354601,
-0.0303473026,
0.0045440663,
0.0271874759,
-0.0396165103,
0.0218724143,
0.0076250946,
-0.0189991239,
-0.0504881889,
0.0541986823,
-0.1010350883,
-0.027809184,
0.0657426417,
-0.0405153632,
-0.0418738872,
0.0311552957,
-0.0375889912,
0.0349644311,
-0.0347245187,
0.0263048895,
-0.0130220605,
0.0209173877,
-0.0396607406,
0.0327449553,
0.0175998025,
-0.0070281681,
-0.0292480066,
0.0232931245,
0.0435395651,
-0.0036125828,
-0.0582588799,
0.0469673984,
0.0377727523,
-0.0109497076,
-0.0017509888,
-0.0293152779,
-0.0502939895,
-0.0317869708,
-0.0421979353,
-0.0103712948,
-0.0510425344,
0.0176410396,
0.0167458355,
0.0054939897,
0.011504516,
-0.0178576801,
0.0200904831,
-0.0016501403,
0.038421642,
0.0254419111,
0.0355318151,
-0.0630544275,
-0.0429678708,
-0.038407553,
-0.0453380086,
-0.0135921314,
0.0400906652,
-0.0114413826,
-0.0239764564,
0.0045037293,
-0.0230179038,
-0.0453888513,
0.0113328425,
0.0220547765,
-0.016374249,
0.0009159565,
-0.0154971723,
0.0127763674,
0.0269456431,
-0.047194358,
0.0274769031,
0.0573750585,
0.020120021,
-0.0144715682,
-0.0368836559,
-0.0203186553,
0.0181722417,
-0.0341196544,
-0.0175926536,
-0.0490785651,
0.027543515,
-0.0049723042,
-0.0104746073,
0.0018069642,
-0.0039849733,
-0.0245497469,
0.0215997137,
0.0023698683,
0.0007877225,
-0.0032892288,
0.0245526582,
-0.0185510498,
-0.041471906,
0.0477452092,
-0.017335806,
0.0361672528,
-0.0457559004,
-0.0208878778,
-0.0564945452,
0.04033323,
-0.0088147679,
0.0186645947,
-0.0320764631,
-0.0384023599,
-0.0611094199,
0.0208592061,
0.0394082889,
-0.0269851033,
0.062102329,
-0.0303458776,
-0.0022024687,
-0.0201946218,
0.0238949265,
-0.0418713167,
0.0310011432,
-0.0301659517,
0.0384296589,
0.0687119663,
0.0055369805,
0.0156350657,
0.0055878619,
-0.0283802953,
0.034391474,
0.0403127782,
0.0088034784,
-0.0222156662,
0.032555636,
0.06536448,
-0.0009722014,
0.0707443058,
-0.0059912154,
0.0452165231,
0.0112617649,
0.0489545316,
-0.0623901673,
-0.0516403839,
0.0027250808,
0.0293576717,
-0.0806718618,
-0.0260054898,
-0.0449130498,
-0.0172029547,
-0.045041725,
-0.0139039466,
0.1278178692,
0.0286170058,
-0.0605623238,
0.0075892149,
0.0120694749,
0.0060435999,
-0.0138590839,
-0.0501419567,
-0.1057445183,
-0.0391480587,
0.0206581596,
0.0027329966,
0.0602706671,
-0.0181771088,
0.0255191326,
-0.0900400504,
-0.0132127944,
0.010182308,
-0.0004681113,
-0.0591152571,
-0.0126280095,
-0.0187094733,
-0.021268988,
-0.0116390884,
-0.0455431268,
-0.0280376598,
-0.004488749,
0.0191939007,
-0.0425105654,
-0.0587720126,
-0.0174396895,
-0.0759940222,
-0.0063257944,
-0.0114332745,
-0.0013492774,
0.0057149464,
-0.0261509214,
-0.0140205696,
-0.0434076935,
-0.0649814829,
-0.009032717,
0.0140906032,
0.0190477967,
-0.0266741719,
0.0005423027,
-0.0254030377,
-0.0638325438,
0.0066802716,
-0.0231447611,
0.0592223965,
-0.0142360795,
-0.0004153602,
0.0209681205,
0.0326262303,
0.0092257624,
0.0520434044,
-0.007226774,
0.0445337631,
-0.0791584477,
-0.0010381251,
0.0493536256,
-0.0356132798,
0.0372554176,
0.0156965107,
-0.0353489444,
-0.0139882909,
-0.0559835322,
-0.0724929422,
0.0010790004,
0.0129122818,
0.000780112,
0.037308611,
0.0386899039,
0.0253434666,
-0.0687814951,
0.0231294427,
-0.002872101,
0.0067748143,
-0.0128221754,
-0.0410939343,
0.0008799774,
0.0021275212,
-0.0055608423,
0.0366521329,
-0.0884118751,
-0.0280481353,
0.081934087,
-0.0391852595,
0.0196459703,
0.0245049857,
-0.0396775864,
-0.0384974368,
-0.0217282157,
-0.0309813116,
-0.0069654076,
0.0094512878,
0.0416757278,
-0.0076825568,
-0.0309986603,
0.0684390739,
-0.0072986144,
0.0041353973,
0.00917878,
-0.0012105398,
-0.0049476936,
-0.0598167256,
-0.0124096125,
0.0095332116,
0.0322715715,
-0.040421959,
0.0210973937,
-0.0341514759,
0.0092889629,
-0.00109796,
0.0303397831,
0.0105888359,
-0.0206447132,
0.0249918103,
0.0109356176,
-0.0221965853,
0.0425632671,
0.006189506,
0.0229489747,
-0.0089690061,
0.0229926389,
0.0143580483,
-0.0311229918,
-0.0447970591,
-0.0029766785,
-0.0183706433,
0.0094304364,
-0.0091131944,
0.0026778991,
-0.0148558253,
-0.021679027,
-0.0289985333,
0.0059809387,
-0.0222920254,
0.0276158396,
-0.0049522757,
0.0096784448,
0.0183996893,
0.0205182284,
0.1165666059,
-0.0606942587,
0.06871894,
-0.0185992196,
-0.0112961568,
-0.0026706723,
0.0531457476,
-0.0396240689,
0.0001410143,
0.0047251419,
0.0119716395,
-0.0056034946,
0.0383095667,
0.0622626059,
0.0438455716,
0.0171808098,
0.0103406198,
-0.0431055687,
0.007728491,
-0.0602794625,
-0.0230456255,
0.0130604068,
-0.0389352515,
-0.0294987075,
-0.0590964667,
-0.0986035541,
0.0200542882,
-0.0106864786,
0.0196626931,
0.0342260003,
0.0463482365,
0.0323152207,
0.0099266237,
-0.0271141659,
-0.0094846906,
0.0201493204,
0.000068296,
-0.0199850779,
-0.0203493666,
-5.285536145e-33,
-0.001648674,
0.013520563,
-0.02390204,
-0.0335761532,
-0.0474040881,
-0.0146879395,
-0.0257233568,
-0.0116404844,
-0.0460142419,
0.0254883077,
0.0166283529,
0.0290418584,
0.0357107855,
0.0248760991,
0.0163005069,
-0.0599113889,
0.0630816519,
0.0068477308,
0.0003538189,
0.0172178652,
0.0768164769,
0.0310867522,
0.0746565387,
-0.0026453987,
-0.0109379571,
-0.0346132517,
0.0063349768,
0.0280971695,
0.0020215537,
0.0667877793,
0.0076244315,
0.079510726,
0.0167077929,
-0.0168408342,
-0.0080138184,
0.0235133711,
-0.1031372771,
-0.0265078619,
-0.0214762129,
0.0585109293,
-0.0005448359,
-0.0126993284,
0.0401039496,
-0.000195444,
-0.0178858712,
-0.0105222259,
0.0384304859,
-0.0053250166,
-0.0289605707,
-0.0153905582,
-0.0518094264,
0.0464094169,
-0.0174876302,
0.0219590534,
0.0432993844,
0.0146162705,
0.0074833329,
0.021163471,
-0.1130527705,
0.0111627998,
-0.021926336,
0.0224557351,
0.0586490892,
-0.000443972,
-0.0268903654,
0.0306030884,
0.0206920207,
-0.058715716,
0.0288743451,
0.0085573476,
0.0347229503,
-0.0225670487,
0.0175269209,
0.0169329178,
0.025749119,
-0.0428546183,
-0.073171258,
-0.0216473453,
0.0555952936,
0.0686248019,
-0.0147050815,
-0.0320850275,
-0.0416508876,
-0.0345365144,
-0.0277203955,
0.0342544019,
0.0073335022,
-0.0132012097,
0.0068006925,
-0.0142935542,
-0.0122601455,
0.0815755129,
0.0159224849,
-0.0350019597,
0.0453960225,
0.0148845986,
0.016360838,
0.0298752077,
0.0061343573,
0.0182744898,
-0.0001482696,
-0.0349485613,
0.0049551423,
0.0302304495,
0.0218724255,
-0.0300937966,
0.0271144956,
0.0026451291,
-0.0594910234,
0.0488728024,
-0.0367071629,
-0.0038415033,
-0.0184729211,
0.0345930308,
0.0167088229,
0.0208755266,
0.0082181664,
0.0132958731,
0.0622356124,
0.0285196621,
0.0053247842,
0.0347166769,
-0.0315061584,
0.0032725413,
0.0074676708,
0.0395833589,
-0.0360477008,
0.0316356756,
0.0370757505,
-0.0551120862,
0.0057004397,
-0.0174768344,
2.4e-7,
0.0630387217,
0.020251397,
0.0268142018,
0.0006177045,
0.0125044696,
0.0019370546,
-0.0237960499,
0.0214170329,
-0.0435850993,
0.0212105308,
0.0099677742,
0.022304289,
0.0566777438,
0.0276114475,
-0.0527183935,
-0.0133514367,
-0.0702139363,
0.0431051925,
-0.0413672812,
0.0266305972,
0.0321805738,
0.1157470495,
-0.0146457162,
-0.0286155585,
-0.0403283536,
-0.0006124351,
0.0262383781,
0.011962614,
-0.0311342981,
-0.0055709733,
0.0396418199,
-0.0036739933,
0.0066807172,
0.0109660439,
-0.0532102212,
0.0052231392,
0.0767162144,
0.0630179793,
0.0298055187,
-0.0267317407,
-0.0079578869,
-0.027525181,
0.010482464,
-0.0288170408,
0.0729173794,
-0.0157115944,
0.0125181265,
-0.0154454103,
-0.036061354,
0.0419912413,
0.0373520851,
0.005707283,
-0.0396093652,
0.00497662,
0.0287324209,
-0.0517611355,
0.0059683598,
-0.0339695103,
0.0163501687,
0.0461000875,
-0.0005850152,
-0.0027569297,
0.0487353094,
-0.0389731638,
0.0158326477,
0.0152595108,
-0.0055928915,
2.104437945e-34,
0.025613999,
0.005059497,
0.0155457528,
-0.019438006,
0.0271800831,
-0.0582749099,
-0.0135891251,
0.0371639058,
0.0093191927,
-0.0317067057,
0.0152889593
] |
http://arxiv.org/abs/2405.15604v1
|
http://arxiv.org/pdf/2405.15604v1
|
Text Generation: A Systematic Literature Review of Tasks, Evaluation, and Challenges
| 2024-05-24T14:38:11 | 2024-05-24T14:38:11 |
Jonas Becker, Jan Philip Wahle, Bela Gipp, Terry Ruas
|
cs.CL
|
cs.CL, A.1; I.2.7
|
Text generation has become more accessible than ever, and the increasing
interest in these systems, especially those using large language models, has
spurred an increasing number of related publications. We provide a systematic
literature review comprising 244 selected papers between 2017 and 2024. This
review categorizes works in text generation into five main tasks: open-ended
text generation, summarization, translation, paraphrasing, and question
answering. For each task, we review their relevant characteristics, sub-tasks,
and specific challenges (e.g., missing datasets for multi-document
summarization, coherence in story generation, and complex reasoning for
question answering). Additionally, we assess current approaches for evaluating
text generation systems and ascertain problems with current metrics. Our
investigation shows nine prominent challenges common to all tasks and sub-tasks
in recent text generation publications: bias, reasoning, hallucinations,
misuse, privacy, interpretability, transparency, datasets, and computing. We
provide a detailed analysis of these challenges, their potential solutions, and
which gaps still require further engagement from the community. This systematic
literature review targets two main audiences: early career researchers in
natural language processing looking for an overview of the field and promising
research directions, as well as experienced researchers seeking a detailed view
of tasks, evaluation methodologies, open challenges, and recent mitigation
strategies.
|
Text Generation: A Systematic Literature Review of Tasks, Evaluation, and Challenges: Text generation has become more accessible than ever, and the increasing
interest in these systems, especially those using large language models, has
spurred an increasing number of related publications. We provide a systematic
literature review comprising 244 selected papers between 2017 and 2024. This
review categorizes works in text generation into five main tasks: open-ended
text generation, summarization, translation, paraphrasing, and question
answering. For each task, we review their relevant characteristics, sub-tasks,
and specific challenges (e.g., missing datasets for multi-document
summarization, coherence in story generation, and complex reasoning for
question answering). Additionally, we assess current approaches for evaluating
text generation systems and ascertain problems with current metrics. Our
investigation shows nine prominent challenges common to all tasks and sub-tasks
in recent text generation publications: bias, reasoning, hallucinations,
misuse, privacy, interpretability, transparency, datasets, and computing. We
provide a detailed analysis of these challenges, their potential solutions, and
which gaps still require further engagement from the community. This systematic
literature review targets two main audiences: early career researchers in
natural language processing looking for an overview of the field and promising
research directions, as well as experienced researchers seeking a detailed view
of tasks, evaluation methodologies, open challenges, and recent mitigation
strategies.
|
[
0.0663484409,
0.0396212116,
-0.005664398,
-0.0135638677,
-0.0420899689,
-0.0044139987,
0.0321785733,
-0.0328494906,
-0.0066675027,
-0.0445006602,
0.0263601653,
-0.0341924988,
0.0370237306,
0.024393307,
0.021998344,
-0.0754090622,
0.0116919335,
-0.0208922811,
0.0049615791,
-0.0050020195,
-0.0026919451,
0.0559320599,
0.0161075927,
0.0529842675,
0.0037914922,
0.0004689719,
-0.0143733341,
0.0099299708,
0.0268681645,
-0.01933196,
0.0097465236,
0.0401229039,
0.0082068294,
0.0498335399,
0.0000022736,
-0.0632370785,
-0.023836622,
0.007219966,
0.0164233483,
-0.0057528052,
0.0585963465,
-0.0224324595,
0.0165012199,
0.0417781323,
-0.0013811751,
0.0126223182,
0.036881499,
0.0125678945,
0.0859574601,
0.0633040369,
-0.0035382914,
-0.0299439877,
-0.0001353731,
-0.0046793651,
0.0029711158,
0.0288785007,
-0.0101385331,
0.0092173964,
0.004558221,
-0.0213049948,
0.0267515779,
0.0153844962,
0.006673425,
-0.0059508206,
0.0140844826,
0.0472615696,
-0.0437375642,
-0.0487631746,
0.0120585654,
0.0758153126,
0.0199198537,
-0.0274271127,
0.0390680768,
-0.0066552926,
-0.0100972848,
0.0046266457,
-0.0603711158,
0.1104865149,
-0.0259418041,
0.0203479081,
0.0097972797,
0.0159263499,
-0.0129226521,
0.0459920615,
-0.0030687815,
-0.0052388273,
0.0236951727,
-0.0215838049,
-0.0434036702,
0.0287061743,
-0.0146331294,
-0.0941614285,
0.0206544511,
0.0387208275,
-0.037228629,
0.0203279871,
-0.0925419405,
-0.0737182349,
0.05528735,
-0.0387401693,
0.0177270863,
0.0222071633,
0.0220432486,
-0.0010976669,
-0.0028014956,
0.0254078507,
0.0162638985,
0.00456431,
-0.0451838933,
0.0054118107,
-0.0210250542,
-0.0049560438,
-0.0499450415,
0.0403061807,
-0.037531998,
-0.0101539893,
-0.0336885974,
0.0170348901,
0.0318897218,
0.019264685,
-0.0067301509,
0.0379996933,
-0.0385057107,
-0.0112800896,
-0.0021444394,
-0.0206637103,
-0.0266363118,
-0.0056076758,
0.0409337357,
-0.0298267845,
-0.0096002677,
0.0298391934,
0.0184566472,
-0.0005542444,
-0.0107813757,
0.0472390503,
-0.0213979967,
-0.017792929,
-0.0238555428,
-0.111513637,
-0.0315310694,
-0.0700203478,
-0.0235492513,
-0.0818005577,
0.0074195238,
-0.0001113428,
-0.0124557493,
-0.0270341095,
0.0233110581,
0.0815864354,
-0.0504629239,
0.0148827573,
-0.0494986475,
-0.0033237927,
0.0585053898,
0.0076393373,
0.0244579818,
0.0491933636,
0.0772525668,
0.0696447194,
0.0360647179,
0.0253082253,
0.0154716019,
-0.0081832269,
0.0248646792,
-0.0191575699,
-0.0074350536,
-0.0030216256,
-0.050453268,
0.0232679527,
-0.0147612449,
0.096363157,
-0.0226103812,
0.0262973066,
0.0671195164,
0.0426829569,
-0.0218925457,
0.0181542616,
0.0545339249,
0.0130893718,
0.0187583826,
0.0215037372,
0.0008267563,
0.0118437354,
-0.0167543814,
0.046349287,
0.051290784,
0.0448927954,
0.0041771326,
-0.0403391123,
-0.0158295128,
-0.0102558937,
0.002215378,
0.0061936649,
0.0176474135,
0.0417457037,
-0.054737512,
0.0097546242,
-0.0014719564,
-0.1109267473,
0.051859051,
-0.0041182647,
0.0500473529,
0.0931528732,
0.0767481327,
0.0206014607,
0.0195954703,
-0.0263963714,
0.0395326577,
0.0601783618,
0.0306951571,
-0.0050081918,
0.0027289609,
-0.0595336854,
-0.0048234165,
-0.0280232001,
-0.0123753957,
0.0116952937,
-0.0247977655,
0.0150637226,
0.0576147735,
-0.0190568455,
-0.0648858026,
0.0413933098,
-0.0380024686,
-0.0243015401,
0.0037952953,
-0.0176067762,
-0.0072470191,
0.0416551903,
-0.0263747703,
0.0474111289,
0.0022939153,
-0.0130375735,
-0.0258264691,
0.053648077,
-0.0329917222,
-0.0242760759,
-0.0086759534,
-0.0783246085,
0.0147791402,
-0.007836408,
-0.0185571574,
-0.0743475631,
0.0324975327,
-0.0648694932,
0.0098614497,
-0.0462259129,
0.0226761661,
0.0212672371,
0.0194443855,
-0.0413742773,
0.0780770481,
0.0122669786,
-0.0053224904,
0.0464705043,
-0.0193095934,
0.0234559234,
-0.0208810549,
0.0068181548,
0.0074674771,
0.0057841041,
0.0003592185,
0.0532201342,
0.0249106344,
-0.0530940183,
-0.0321256295,
-0.076853931,
0.0294197705,
-0.0168095566,
0.014600073,
0.0633294731,
0.0241197627,
-0.0062494832,
0.0029571615,
-0.0455154888,
-0.0144879129,
0.0499520637,
-0.0052983556,
0.049729161,
0.0098819891,
0.0281720646,
-0.0393172503,
-0.0041558556,
0.0119378865,
0.0429401249,
0.0729014277,
-0.0088754036,
-0.0503593609,
0.0386755541,
-0.0377433971,
-0.0305174757,
0.027234165,
-0.0031842499,
0.0040210239,
0.0219808612,
0.0425370522,
0.0120739974,
0.0632390678,
0.0120263435,
0.0433558412,
-0.0233450197,
-0.012037144,
-0.0120476829,
-0.0431998819,
0.018690668,
0.0468755849,
-0.0782980546,
0.0554634817,
0.0260699373,
0.0501087978,
0.0026030324,
-0.0582128577,
0.0443951339,
-0.0278393552,
-0.0481347367,
-0.0273060594,
-0.0430342108,
0.0025828157,
0.0461493842,
-0.0283131879,
-0.039974954,
0.0575827323,
-0.0523608215,
0.018118266,
-0.0472289212,
0.0032600795,
-0.0593426414,
0.0365864299,
0.0127426516,
0.0574032255,
0.0254275948,
-0.0086582582,
-0.0497641638,
-0.0142487111,
0.0478524826,
-0.0429587401,
0.0358297415,
-0.0696716234,
-0.0087453937,
-0.0172680933,
-0.011015499,
-0.0383026004,
-0.053841535,
0.0061216494,
-0.010352999,
0.0428964831,
-0.009597688,
-0.0373512842,
-0.010193442,
0.0448651463,
0.0187257733,
-0.0152145019,
-0.0353579409,
-0.0187985264,
0.0492273085,
0.0548910461,
-0.0195612554,
0.0233612191,
-0.0419740416,
0.0055962205,
0.0534220673,
0.0326446742,
0.0375891551,
0.0193336047,
-0.0201747231,
-0.0034282566,
-0.0476829261,
-0.00923126,
0.0034096371,
0.060846407,
-0.0542514324,
-0.0075913817,
0.106469132,
0.0378818288,
-0.0065340525,
0.0046821814,
0.0042268601,
0.0418534465,
-0.0174156502,
0.0061337505,
-0.0669142604,
-0.0048179841,
0.0012865316,
0.0737980604,
-0.0508175045,
-0.0552055389,
-0.0550033525,
-0.0961381271,
0.0334656388,
0.0163032878,
-0.0207517203,
-0.0237333197,
-0.0191150215,
-0.0467124097,
0.0092289653,
-0.0020691853,
0.0041491115,
-0.0212527271,
0.0462084822,
0.028498251,
-0.0455419347,
-0.014926509,
-0.0389090218,
-0.0443241224,
0.0235339906,
0.0618839115,
0.0435478128,
0.0088017127,
0.0122882547,
0.0042579724,
-0.0268648174,
-0.0043239756,
-0.0008052711,
0.0813928843,
-0.0057881386,
0.0099927783,
-0.052425988,
0.058274623,
-0.0106290141,
-0.0335830897,
0.0427545607,
0.0091840047,
-0.0128848171,
-0.0310670193,
-0.031470865,
0.0540708527,
0.0104357647,
0.0808295831,
0.0225075558,
-0.0285337772,
-0.0419257283,
-0.0317787714,
0.0346058458,
-0.0994656309,
0.0101061491,
0.0083059799,
0.0069574206,
-0.0236083101,
-0.1058898941,
-0.0543361418,
0.0122689297,
-0.0064209714,
0.0135222552,
0.0091542397,
0.0191678219,
-0.0193509795,
-0.0227156635,
-0.0062982487,
-0.0131574674,
0.0123851933,
-0.0194662623,
-0.0268788468,
-0.0096184164,
-0.0674316362,
-0.002756285,
0.0468698367,
-0.0804466084,
-0.0598957427,
-0.0227766838,
0.0372903906,
0.0217790175,
0.0378920995,
0.0134185236,
-0.0282982495,
0.068094939,
0.0313487537,
-0.0086290659,
0.0198915396,
-0.0210178308,
0.0209480189,
-0.009330892,
0.0062977765,
-0.048593916,
-0.0108260708,
-0.0180111192,
0.0140241766,
0.0110324435,
-0.0318861865,
0.03664352,
0.034885779,
-0.0251488872,
-0.0691898614,
0.0259298719,
-0.0514264964,
-0.0238153599,
-0.0603668988,
-0.0274966136,
-0.0638652071,
0.0237600878,
0.0029839564,
-0.0028932604,
0.0093070101,
0.0059308945,
-0.042754788,
0.0184596684,
-0.0181332659,
0.0174886454,
-0.0239043217,
0.0458566807,
0.0135217188,
-0.0349067077,
-0.0529255532,
0.0075333011,
-0.078189671,
0.0310678706,
0.0019813438,
0.0251264535,
0.0147501109,
-0.0163420495,
-0.005474221,
0.0372607261,
-0.065711543,
-0.0229298957,
-0.0497973152,
0.0220008325,
0.0240347609,
0.0059981593,
0.0844572037,
-0.0284184869,
0.0191223044,
-0.0043718908,
0.0562732294,
0.0283412039,
-0.0378991999,
-0.0555973202,
-0.0087321969,
-0.0254512057,
-0.0234396216,
0.1059583053,
0.0152103752,
0.0094187669,
-0.0005096393,
-0.0351110324,
-0.0319827944,
-0.0355969407,
-0.0366021395,
0.0287410524,
-0.0018370392,
-0.0224799495,
-0.0665920004,
-0.0314431228,
0.0104399538,
0.0211975239,
0.0353397615,
-0.0020190568,
0.0455037728,
0.0204043873,
-0.0062712058,
-0.0002395763,
-0.0668804646,
0.0132414531,
-0.0390645005,
-0.0076137637,
-0.0160429645,
-6.325092656e-33,
-0.0181898512,
-0.04836566,
-0.0042482433,
-0.0355664603,
-0.0295647942,
-0.0131857386,
-0.0568577722,
-0.0128706368,
0.0326124914,
-0.0207551513,
0.0015704625,
-0.0315648429,
0.0119168805,
0.074542515,
0.0054706796,
0.0127113936,
0.049017895,
-0.0065580215,
-0.0601949021,
0.049452316,
0.0524318404,
-0.003681801,
0.0107587539,
-0.03741372,
0.0391611904,
-0.0442517884,
-0.0414679535,
-0.038564425,
0.0102241859,
0.0160484221,
-0.0085665556,
0.0388971716,
0.0147354808,
0.0045737545,
-0.0492793508,
0.0443188511,
-0.041794505,
-0.0284140129,
0.0029415085,
-0.0227461867,
-0.0542588308,
-0.0479991212,
0.0305913799,
-0.0393386371,
0.0367614962,
0.0027347102,
0.0543101951,
-0.026670875,
0.0027064965,
-0.022141993,
-0.0392170399,
0.0252292361,
-0.0408317037,
-0.0181918964,
-0.0118258446,
0.1083011553,
0.0090769259,
0.0106650228,
-0.0490514673,
0.0474855676,
0.0261975937,
0.0042701233,
0.062042471,
0.0084423833,
0.0102256332,
0.0516313426,
0.0076917023,
-0.0038991119,
-0.0661190376,
0.009695366,
-0.0417647287,
0.0120641477,
0.0288469456,
0.0618719347,
0.023479864,
-0.0014569665,
-0.0380761884,
0.067299284,
-0.0302209221,
-0.0105240569,
0.05078502,
-0.0631906912,
-0.0203078315,
-0.0171058103,
-0.0399795882,
-0.0177044757,
-0.0032080282,
-0.0119147412,
0.0129023762,
0.0227238648,
-0.0526406802,
-0.0348705724,
-0.0040908828,
-0.0006551305,
0.0020775597,
0.012935074,
-0.0331535153,
0.0116694104,
-0.0105242645,
-0.0008966678,
-0.0634479597,
-0.0001159325,
0.0574236624,
0.0316685811,
-0.0000478948,
0.0210982263,
0.0245856382,
0.0184776951,
-0.0325041339,
0.0212735757,
-0.0286221672,
0.0752781332,
0.0054448084,
-0.0083825374,
0.0075403443,
-0.013589073,
-0.0186986662,
0.0017142043,
-0.0026807918,
0.0060334466,
-0.0105766281,
-0.0181827545,
0.0069547594,
0.0117883421,
-0.02531044,
0.0232199952,
-0.0181563012,
0.0014557047,
0.079784587,
-0.0287986547,
-0.0091821179,
-0.0134620667,
2.915e-7,
0.0326668434,
0.0203858707,
0.0463902615,
0.0238678157,
-0.0251430441,
0.0205000993,
0.0248191431,
0.052803427,
-0.0092337579,
0.0238099527,
0.016922798,
0.0065217968,
0.0467305034,
0.0371672697,
-0.0886314362,
-0.0374815576,
-0.0632253736,
0.0156835802,
-0.0647393167,
-0.0074266577,
0.0924408436,
0.0351528488,
-0.0119467787,
-0.0188569576,
-0.0057486864,
-0.0233477633,
0.0570181943,
-0.0188205615,
-0.0205463208,
0.0096259844,
0.0406430848,
-0.0059178816,
-0.0063188258,
0.010029885,
-0.0200385824,
-0.0207806956,
0.0259298664,
0.0133496607,
0.0397197604,
-0.0086932229,
0.0214416832,
-0.027237108,
0.0089825494,
-0.0525395125,
0.0599667169,
-0.0324270763,
-0.0250414815,
-0.0063112318,
-0.0001900927,
0.0010777267,
0.0154018793,
-0.0279815104,
-0.035167899,
0.0301806517,
0.0042150849,
-0.000933943,
0.0185248479,
-0.0255525019,
-0.0281200353,
-0.031787321,
-0.0247647949,
-0.027349012,
-0.0286642201,
0.0083911605,
-0.0020560375,
-0.0471115857,
0.0232088529,
2.867043349e-34,
0.0359169804,
-0.0412182398,
-0.0196689777,
-0.0474159867,
0.0300108697,
-0.0115439473,
0.0010968092,
-0.0049010082,
0.0041757328,
-0.0556051433,
-0.0420945957
] |
http://arxiv.org/abs/2405.15590v1
|
http://arxiv.org/pdf/2405.15590v1
|
Profiling checkpointing schedules in adjoint ST-AD
| 2024-05-24T14:20:45 | 2024-05-24T14:20:45 |
Laurent Hascoët, Jean-Luc Bouchot, Shreyas Sunil Gaikwad, Sri Hari Krishna Narayanan, Jan Hückelheim
|
cs.CL
|
cs.CL
|
Checkpointing is a cornerstone of data-flow reversal in adjoint algorithmic
differentiation. Checkpointing is a storage/recomputation trade-off that can be
applied at different levels, one of which being the call tree. We are looking
for good placements of checkpoints onto the call tree of a given application,
to reduce run time and memory footprint of its adjoint. There is no known
optimal solution to this problem other than a combinatorial search on all
placements. We propose a heuristics based on run-time profiling of the adjoint
code. We describe implementation of this profiling tool in an existing
source-transformation AD tool. We demonstrate the interest of this approach on
test cases taken from the MITgcm ocean and atmospheric global circulation
model. We discuss the limitations of our approach and propose directions to
lift them.
|
Profiling checkpointing schedules in adjoint ST-AD: Checkpointing is a cornerstone of data-flow reversal in adjoint algorithmic
differentiation. Checkpointing is a storage/recomputation trade-off that can be
applied at different levels, one of which being the call tree. We are looking
for good placements of checkpoints onto the call tree of a given application,
to reduce run time and memory footprint of its adjoint. There is no known
optimal solution to this problem other than a combinatorial search on all
placements. We propose a heuristics based on run-time profiling of the adjoint
code. We describe implementation of this profiling tool in an existing
source-transformation AD tool. We demonstrate the interest of this approach on
test cases taken from the MITgcm ocean and atmospheric global circulation
model. We discuss the limitations of our approach and propose directions to
lift them.
|
[
-0.0276907738,
0.0373357944,
-0.0310664047,
0.0105232112,
-0.0213802867,
-0.0289528389,
-0.0130750127,
-0.0665964708,
0.0032110543,
-0.0092656147,
0.0552698523,
0.0084608598,
-0.0017800175,
-0.0044344408,
-0.0453150123,
-0.0075814803,
0.067851685,
-0.023951143,
-0.0505356491,
0.0012438135,
-0.0707124919,
-0.0433683097,
-0.0073285033,
-0.0325968564,
0.0427586809,
-0.0444678143,
-0.0032987345,
0.0356681347,
0.0454962812,
0.0216765907,
0.0087968223,
0.011409957,
-0.0179140288,
0.0801013783,
0.000002072,
0.0357142463,
0.0203842465,
0.0491379313,
-0.014535971,
0.1195232198,
0.0243780892,
0.0557721145,
0.0347235724,
-0.0063598356,
-0.0047224811,
0.0283294711,
0.0146476692,
-0.1164733842,
0.0977780148,
-0.0092166187,
0.0299361721,
0.0544026569,
-0.0200096499,
0.0173922051,
-0.0699223951,
-0.0115263565,
0.0082068546,
-0.0186227281,
0.0048740865,
-0.0522490479,
0.036913801,
0.0428657755,
-0.0128955189,
-0.0246295817,
0.0254888367,
-0.014506042,
-0.01539509,
-0.0024958409,
-0.027650306,
-0.0229705963,
0.0137395496,
-0.0250900071,
-0.0064249099,
0.0273662582,
-0.0362920836,
-0.0139815472,
0.0127911214,
-0.0191463493,
0.0006350775,
-0.0585709512,
0.0347300284,
-0.0372046605,
-0.0171704404,
-0.0479177497,
-0.0598850138,
0.0584915318,
0.0150105217,
-0.0067081372,
-0.009009013,
0.032336168,
0.0155265462,
-0.0249576103,
0.0135246227,
-0.0087000541,
-0.0499672294,
-0.0217172727,
0.0014585115,
-0.0284737926,
-0.0424815156,
-0.0206097607,
0.1175046787,
-0.0140821887,
-0.0302131418,
-0.0523172319,
0.0193352159,
0.0208664611,
-0.0411511734,
-0.0200311989,
0.0523670428,
0.0426299609,
0.0382756628,
-0.0418965295,
-0.0021403895,
0.0151249114,
-0.0418954045,
0.0552277341,
0.0195113122,
0.0246378928,
-0.0175934527,
0.0023232154,
0.0153205683,
-0.035883531,
0.0265463982,
0.0313694924,
0.0535230488,
-0.0546988696,
-0.0226965155,
-0.0406214297,
-0.0155666843,
0.0159356911,
0.0042263111,
-0.0001000457,
-0.0447552167,
-0.0454460867,
-0.0311011374,
0.0383921526,
0.0154522313,
0.0153560825,
-0.0644930303,
0.012498307,
0.0271967817,
-0.0460081771,
0.003009157,
0.0218949411,
-0.0302614011,
-0.0249115434,
0.0073662773,
-0.0168190785,
-0.0535685234,
-0.032282576,
0.0222577341,
0.0294723194,
0.0632196516,
0.0192161929,
-0.0020930271,
0.04249667,
0.0367781557,
0.0364365354,
0.0053992765,
-0.0005568781,
0.0018889152,
-0.0434376113,
-0.0314125717,
0.0119051924,
0.0330611542,
0.0227823015,
-0.0281626694,
0.0275150761,
-0.030442806,
-0.0045893164,
0.0132977162,
-0.0288017429,
-0.0232219696,
-0.0245797075,
0.0118460944,
-0.0850050226,
0.0175830517,
-0.0116146738,
0.0388551988,
0.0047993609,
-0.0117067993,
0.0254044663,
-0.0291232783,
0.0371747464,
-0.0235849153,
-0.0082371561,
-0.0159693155,
0.0803194344,
0.0070713717,
0.0062748929,
-0.012495908,
0.0486298203,
-0.0675365105,
-0.0152096702,
0.0163126532,
0.0119166449,
-0.0020064595,
0.0159831923,
-0.0066273515,
-0.0328547247,
-0.0390069522,
-0.0553848296,
-0.0001129871,
-0.0221249219,
0.0031707152,
-0.0670125932,
-0.0428571478,
0.0186601169,
0.0457724668,
-0.0255739577,
0.0446139723,
0.0117371334,
0.0269137118,
0.0249463115,
-0.0058974591,
-0.0432103947,
0.0061306739,
0.0561227202,
0.0380256027,
-0.0211151019,
-0.0159880146,
-0.0070316489,
-0.0229620486,
0.0061572497,
0.0383488685,
-0.0055688084,
0.0627220348,
0.0198156592,
0.0405357778,
0.0253107566,
-0.0517150648,
-0.009395753,
0.0070614591,
-0.0061554848,
-0.0148678403,
0.0160086378,
0.0240788367,
0.0375266895,
-0.0203542635,
-0.0024562352,
0.0174811948,
0.0298990626,
-0.0443756059,
-0.0422325097,
0.0958771929,
0.0132891638,
-0.0394396707,
0.01563951,
0.0776879787,
0.0068234811,
0.0275471043,
0.0063903709,
0.0571613349,
0.0108398255,
-0.0259515122,
0.0481758602,
0.0562426262,
-0.0065431674,
-0.0060319048,
0.0122328065,
-0.004810736,
-0.0028074149,
-0.0249628481,
0.0245472249,
-0.0139914667,
0.0620366037,
-0.0025489635,
-0.0172051694,
0.0598688237,
0.039873451,
0.0536003374,
-0.0243433826,
-0.0015190939,
0.0282320753,
0.0308924783,
0.0463990197,
0.008108886,
-0.0558550619,
0.005115143,
0.001310249,
0.0002770716,
-0.0762287974,
0.0127217807,
-0.0067096432,
0.0478616394,
0.0625272319,
-0.0295155048,
-0.0537671596,
-0.1081522405,
0.005264652,
-0.0015231884,
-0.0018284831,
-0.0333354548,
0.0250973124,
-0.0274062362,
0.0027698907,
0.0152751422,
0.0395133011,
-0.0393008851,
0.0240663625,
-0.0145313079,
-0.0282037724,
-0.0274045635,
0.0070804032,
0.0347532518,
0.0215991475,
0.0665692687,
0.0236106869,
0.0341330208,
0.0373856537,
-0.0271517131,
-0.0340864547,
0.0701833144,
-0.0158537012,
-0.0190170128,
-0.0182138495,
-0.026068626,
-0.0106505891,
-0.0591460168,
-0.0106047513,
0.0025192294,
-0.0860315934,
0.0081837252,
0.0015621122,
0.0054027471,
-0.026438117,
-0.0403621309,
-0.0115813352,
-0.0279203653,
-0.0576268174,
0.0096542798,
-0.0150511703,
-0.0330836475,
0.0157678351,
0.0387041681,
0.0151122687,
-0.0281339269,
-0.0133891841,
-0.014939826,
-0.0194027536,
0.0052217073,
-0.0442795753,
-0.029088892,
0.0499760211,
0.0109148333,
-0.006912922,
0.0086671691,
-0.088048026,
0.0068644742,
0.0266098361,
0.0173231587,
0.0085756732,
-0.0090931039,
-0.0141776642,
0.0217661057,
-0.0036614942,
-0.0370352231,
-0.0513175353,
-0.0203949995,
0.0238703713,
0.0325332843,
-0.0539694093,
0.000460423,
0.0361041017,
-0.070784077,
-0.027450379,
0.0115544489,
-0.0074943332,
-0.069633171,
0.0310601611,
0.022609232,
-0.0003833745,
-0.0434330404,
-0.0477333218,
0.0340854153,
-0.0006334293,
0.023527652,
-0.0185624845,
0.0441968627,
0.0300659407,
0.0400464684,
-0.023348337,
-0.0029573976,
0.0460884646,
0.0480559357,
-0.0506458767,
0.0294225924,
0.0017425899,
-0.0126381973,
0.0602473319,
0.0245054215,
-0.0604069754,
0.0088177193,
-0.1106153801,
-0.0529928729,
-0.0258395504,
-0.0252983645,
0.0010012744,
0.0011909395,
-0.0080678901,
0.0056381589,
0.0028107902,
0.010040204,
-0.0303445775,
0.0394342765,
-0.0311081111,
0.0297263283,
0.0562757924,
-0.0026907418,
0.0163911749,
0.0551702902,
-0.0190140679,
0.0362010784,
-0.0569884256,
-0.0067368867,
0.0560105778,
0.0215931404,
-0.0260288324,
0.0185927916,
-0.018381238,
0.0505447462,
0.0379478186,
0.1154457852,
-0.0503977723,
-0.0370685123,
-0.0384462588,
-0.0247368496,
0.057364732,
-0.005851265,
0.0019298462,
0.0121810455,
-0.0401966944,
-0.0163776465,
0.0300015658,
-0.0693178624,
0.0124361115,
-0.0125287604,
0.0033051493,
0.0353590362,
0.0279913172,
0.0605057478,
0.0032777032,
0.0750294104,
0.0323551819,
-0.0293847769,
-0.1046894342,
-0.023413403,
-0.109854117,
-0.0200788807,
-0.055605229,
0.0813784972,
0.0088836299,
-0.0298546776,
-0.0087738791,
0.0234135929,
-0.0200871248,
0.0587734468,
-0.040345002,
-0.0648794994,
0.0313307829,
0.0566024967,
0.0311884806,
0.0072527295,
0.0004704173,
0.003807191,
-0.0286739226,
0.0729724839,
0.0152981961,
0.0393773653,
0.025464179,
-0.0577777475,
0.0065591098,
-0.057950668,
0.0001585433,
-0.0002095916,
0.0533726774,
0.0006253865,
0.0673997402,
0.0021502292,
-0.0595959015,
0.0712280944,
0.0279224981,
0.0077028079,
-0.0669942871,
-0.0200050697,
0.0328020938,
0.0083194338,
0.0132240998,
-0.0266340896,
0.0673522055,
-0.0529274754,
0.00248454,
-0.0162804388,
0.0804553553,
0.0317700952,
-0.0279642027,
-0.0340308249,
-0.0292159077,
0.0318508707,
0.0733587742,
-0.0038914338,
0.0192437284,
0.0143441716,
0.0310326293,
0.0826090947,
0.0165742468,
0.0282417312,
0.0170211103,
-0.0163254347,
0.0492451973,
-0.0071937051,
-0.0316518247,
-0.0597778149,
0.0001430931,
-0.04547235,
-0.0229852404,
0.0496173315,
0.0601910315,
0.0394278951,
-0.0071796277,
0.0234636422,
0.0078509981,
0.0097765559,
0.0003425731,
0.0323196277,
-0.0333824381,
-0.0051621031,
-0.0330788903,
-0.0117824692,
0.0202424079,
0.0038237432,
0.0520625114,
-0.0538633242,
-0.0135297487,
-0.0397092104,
-0.0391874872,
0.0406005904,
0.0191501584,
0.0131259756,
-0.0710170269,
-0.0625919774,
0.0119395303,
-0.0390477404,
-0.0133561455,
0.0690176338,
-0.055433847,
-0.0093092248,
0.0069271652,
-0.0121730985,
0.03594492,
0.0126542551,
-0.0258184783,
-0.018353736,
-0.0212949086,
0.0306370109,
-6.22157495e-33,
-0.0458362699,
0.0480012447,
-0.021127386,
0.0043271673,
-0.0490389057,
0.0249314792,
0.0021420282,
-0.0314628854,
0.0731970742,
-0.0029111777,
0.0126374261,
0.0015423157,
0.0099785794,
-0.0691453964,
-0.0040321751,
0.0585759543,
0.0323497541,
-0.0055661448,
-0.0607416555,
0.0332266167,
0.0775812119,
0.0289179701,
-0.044238843,
0.0098747406,
-0.045462288,
-0.0021166881,
0.0283057243,
0.0155457184,
-0.0256605484,
-0.0420196131,
0.0341149867,
-0.0117751379,
-0.0677784756,
-0.0037162493,
-0.0224720053,
-0.0700818598,
-0.0665530041,
-0.0250170827,
0.0563469343,
-0.0265654065,
0.0051910221,
-0.0333639383,
-0.0482636355,
0.0720947534,
-0.0234916955,
-0.0197260994,
-0.046458371,
0.0264304094,
-0.0418184921,
-0.041429732,
-0.0407651216,
0.0343587473,
-0.0246405378,
0.0094126659,
0.0349553525,
0.0005392312,
-0.0319537297,
-0.036497131,
0.0044597783,
0.0470592789,
0.042952057,
0.0054600248,
0.0522111505,
0.0624548011,
0.0255653132,
0.0140224174,
0.0439355895,
0.0196514092,
0.0208581407,
0.0246124025,
0.0254927594,
0.0253122877,
-0.0332827419,
-0.013124547,
0.0420824513,
-0.0228055976,
0.0150479944,
0.0198973324,
-0.0154793896,
-0.0599170737,
0.0258137211,
-0.0321511924,
-0.0117097776,
-0.0273323171,
-0.0000726472,
-0.0488520563,
-0.0216942355,
0.0061918776,
0.0388675705,
-0.0043941941,
-0.0643035173,
0.0483525619,
0.0183371026,
0.006692573,
0.082642667,
0.0167719126,
0.0052393726,
-0.0369456373,
0.0085860221,
-0.0023582496,
0.044316873,
-0.0723033249,
-0.0048921616,
0.0036244881,
0.0420109071,
-0.0113049466,
0.0030031351,
0.0155350817,
-0.0054587997,
0.0045490265,
-0.0156913567,
0.0057991892,
-0.0332813933,
0.0240483023,
-0.025199065,
0.0020713108,
0.0084826592,
0.0479772426,
-0.0149228936,
0.015318267,
-0.0177283157,
0.0584330633,
0.0532508977,
0.0378727652,
0.0445217378,
0.0153784556,
-0.0045360494,
-0.0440696031,
-0.0301682707,
-0.0598455071,
-0.0384469517,
0.0235519372,
2.969e-7,
-0.0421931185,
0.0331583172,
0.0343997478,
0.0113133416,
0.0025342037,
-0.0493795834,
-0.0039662765,
-0.0198358893,
0.0063093482,
0.0253247395,
-0.0137758432,
-0.0276745707,
0.045319926,
-0.0219480749,
0.0704345182,
-0.0131485155,
0.0074044676,
0.0194119625,
-0.0514352657,
-0.1095252186,
-0.0267159138,
-0.0269605611,
-0.0014479262,
-0.0160559472,
0.0157581456,
-0.0068191681,
-0.0054885922,
0.0305851959,
0.042645704,
0.0154453395,
-0.0018071906,
-0.053575363,
0.0163115114,
-0.0409654044,
-0.0523588285,
-0.00072052,
0.0368264392,
0.0688582137,
-0.0057366574,
0.0188513473,
0.0150746321,
0.0110511091,
-0.008044878,
-0.0170957893,
0.0010087524,
-0.0180625468,
-0.020077806,
-0.0593246296,
-0.0282376725,
-0.040950451,
-0.0118665528,
-0.0224314947,
0.081363678,
0.0513608903,
-0.0146111045,
0.0330634527,
0.0020784379,
0.0133788213,
-0.0526063144,
0.0061045303,
-0.001907342,
-0.004503184,
0.0049620802,
0.0128077939,
-0.0028999487,
-0.03239825,
-0.0030532849,
3.182609629e-34,
0.0481569245,
0.0062673828,
-0.0025061432,
-0.0030429668,
0.0257989857,
-0.0318020135,
-0.0083793737,
0.0099442555,
-0.0067527234,
0.0322352313,
-0.0084652966
] |
http://arxiv.org/abs/2405.15585v2
|
http://arxiv.org/pdf/2405.15585v2
|
Synergizing In-context Learning with Hints for End-to-end Task-oriented Dialog Systems
| 2024-05-24T14:13:54 | 2024-07-03T05:26:24 |
Vishal Vivek Saley, Rocktim Jyoti Das, Dinesh Raghu, Mausam
|
cs.CL
|
cs.CL
|
End-to-end Task-Oriented Dialog (TOD) systems typically require extensive
training datasets to perform well. In contrast, large language model (LLM)
based TOD systems can excel even with limited data due to their ability to
learn tasks through in-context exemplars. However, these models lack alignment
with the style of responses in training data and often generate comprehensive
responses, making it difficult for users to grasp the information quickly. In
response, we propose SyncTOD that synergizes LLMs with task-specific hints to
improve alignment in low-data settings. SyncTOD employs small auxiliary models
to provide hints and select exemplars for in-context prompts. With ChatGPT,
SyncTOD achieves superior performance compared to LLM-based baselines and SoTA
models in low-data settings, while retaining competitive performance in
full-data settings.
|
Synergizing In-context Learning with Hints for End-to-end Task-oriented Dialog Systems: End-to-end Task-Oriented Dialog (TOD) systems typically require extensive
training datasets to perform well. In contrast, large language model (LLM)
based TOD systems can excel even with limited data due to their ability to
learn tasks through in-context exemplars. However, these models lack alignment
with the style of responses in training data and often generate comprehensive
responses, making it difficult for users to grasp the information quickly. In
response, we propose SyncTOD that synergizes LLMs with task-specific hints to
improve alignment in low-data settings. SyncTOD employs small auxiliary models
to provide hints and select exemplars for in-context prompts. With ChatGPT,
SyncTOD achieves superior performance compared to LLM-based baselines and SoTA
models in low-data settings, while retaining competitive performance in
full-data settings.
|
[
0.0331004895,
0.0146348523,
0.004390229,
0.0311581474,
-0.0141460728,
0.034224961,
0.0026618284,
-0.0146134542,
0.0363128856,
-0.0314534418,
-0.047355365,
-0.048918236,
-0.0460944772,
0.0371753126,
-0.0048321355,
-0.058044523,
0.0393591858,
-0.0389007181,
0.0025264185,
-0.0101268748,
0.0197347403,
0.009940573,
-0.0195951741,
0.0622230619,
0.0153918471,
-0.0521917082,
-0.0512886234,
0.0201031044,
-0.0095363101,
-0.0252837595,
0.0443162248,
0.0447085351,
0.0220303312,
0.0294441488,
0.0000019454,
-0.0494218804,
0.0031895628,
-0.0225411057,
-0.0757000223,
-0.02847353,
0.026619086,
0.0136007192,
-0.0117280586,
0.0242937431,
-0.0364693254,
0.0087952353,
0.0575913861,
0.0633488819,
0.052014146,
0.063506037,
-0.0002371757,
0.0469685346,
0.008328367,
-0.0200010352,
0.0283844527,
-0.0450978652,
-0.0136953155,
-0.0018878611,
0.0192722529,
0.0314345807,
0.0189820603,
0.016756203,
-0.0075519527,
-0.0225809589,
-0.0173075646,
0.0123450747,
0.0007079506,
0.0064951028,
-0.0339806452,
0.082891494,
0.00177804,
0.0155983679,
0.0001556111,
-0.0162749998,
0.0183000229,
-0.0209293403,
-0.0759496987,
-0.0310600754,
-0.0293959882,
-0.0004982565,
0.0009543242,
0.0126729356,
-0.0134716332,
0.0057482771,
-0.0024271968,
0.0240070745,
-0.0217850916,
-0.0523178913,
0.0174799021,
-0.0035429022,
-0.0168512799,
-0.0584034808,
-0.0146625685,
0.0333510488,
-0.0846304223,
0.013837521,
-0.0493128188,
-0.011618427,
0.0235517249,
-0.0392654836,
0.0386872739,
0.0620995909,
-0.0483302549,
0.0596201159,
0.0049606878,
0.0450646356,
0.0033441223,
0.0027675638,
-0.0379198864,
-0.0398900732,
-0.0488686897,
-0.0061298418,
-0.0774847418,
0.0425750837,
-0.0804874301,
-0.0302063283,
-0.0142623913,
0.0876396894,
0.0005403956,
0.0196206905,
0.0124586485,
0.0130591802,
-0.0565241091,
-0.0003948252,
-0.0339535885,
0.0181879923,
-0.0020975966,
0.0262475088,
0.0432357602,
-0.0307219643,
-0.0228565745,
0.0079852417,
0.0087936614,
-0.0247927811,
-0.0130303111,
0.070636265,
-0.0242552627,
-0.0326186977,
-0.0155725628,
0.0047964971,
-0.0295092557,
-0.0543766841,
-0.0292086899,
-0.0355077833,
0.0081122462,
0.031289231,
-0.0083018243,
0.0016639783,
-0.0022162977,
0.0482956357,
0.0082117934,
0.0420908481,
-0.0210865438,
-0.0068455827,
-0.0037503557,
-0.0000819494,
-0.0692580342,
0.0833243877,
0.0041111181,
0.0310931019,
0.0141340131,
0.0022717221,
-0.0162834693,
0.0279270913,
0.0030145943,
0.0091406563,
0.0346729532,
0.0393085033,
0.0375759751,
-0.0003168134,
0.0257385708,
0.0428210944,
-0.022043068,
0.0142483627,
0.0900535285,
0.0641646609,
0.0330268517,
0.0059674904,
-0.0078642257,
0.0381414182,
0.009613621,
-0.0074912994,
-0.0361191668,
0.0399246365,
-0.0008352257,
0.0294959024,
0.0261969436,
0.0348300301,
-0.0458657742,
-0.0228627659,
-0.010355412,
0.00900352,
-0.0134441806,
0.0362953357,
0.0463696122,
0.0056568803,
-0.07474906,
0.0222408697,
0.0350487269,
-0.031213101,
-0.0508341677,
-0.029413389,
0.0196802411,
0.0616949908,
0.0124535011,
-0.0369822681,
-0.0575676002,
-0.0514972061,
0.0564872511,
0.0450218543,
0.0270463396,
-0.0203322861,
0.0268417392,
-0.0461938754,
-0.0294051431,
-0.0015878157,
-0.0042232848,
0.0102062793,
-0.0197834596,
-0.0137071805,
0.007493,
-0.0101864012,
-0.0130243367,
-0.0050583906,
0.0130539443,
0.0008104913,
0.0159717221,
0.0199789014,
0.0475167967,
0.0079907523,
0.0232103914,
-0.0284175519,
0.0284190029,
-0.0151819652,
0.0336046442,
0.0027086607,
-0.077784203,
-0.0714709386,
0.0131655652,
0.0435097516,
-0.0030566202,
0.0085175559,
-0.0356948338,
-0.0581412427,
-0.0155381197,
-0.0409464091,
0.042943459,
-0.0586611405,
0.0130533567,
0.0278141461,
-0.0143939666,
-0.0340149887,
0.0497278385,
0.0123822568,
0.0101835495,
0.0636515841,
0.0086734677,
0.017684333,
-0.0746375695,
0.0050984169,
0.0040858933,
-0.0105931805,
-0.0026250386,
-0.0028532553,
0.0167127084,
-0.0338416919,
-0.045152355,
-0.1346284896,
0.0030697703,
-0.0286907107,
0.0215079188,
0.0078335665,
0.0251915641,
-0.0082554528,
-0.0412336811,
-0.0330581032,
0.0256958697,
0.053343717,
0.0006632802,
0.0370926932,
0.0250414163,
-0.0172616709,
-0.0419189036,
0.0321419202,
0.0148876775,
0.0160005204,
-0.0145989479,
-0.0305535104,
-0.0696625113,
0.0175076108,
-0.0057194931,
0.0519830175,
-0.0385071374,
-0.0010536887,
0.0212569237,
0.0135499071,
0.0600188263,
0.0569294505,
0.0144338394,
0.0071793715,
-0.0067411079,
-0.0223306566,
-0.0932175294,
0.0574754477,
-0.0102309519,
-0.0040667984,
-0.0333761424,
0.0104053887,
0.0163796023,
0.050629396,
-0.0039664349,
0.0231264941,
-0.0408713818,
0.0235164817,
0.0128541235,
0.0115160141,
-0.0350353904,
-0.0113935778,
-0.0427002199,
0.0071351277,
-0.0368997417,
-0.059567865,
0.0738720968,
-0.0125239156,
-0.0423776507,
-0.0475920103,
-0.0398603454,
-0.0383875184,
0.0465950593,
-0.0369122885,
0.0201658476,
0.0520451479,
-0.0684193447,
-0.0627419576,
0.0188131463,
0.0017507878,
-0.0164818801,
0.0638478249,
-0.010212929,
0.0187378563,
-0.0336439982,
-0.0785111487,
-0.0662677884,
-0.0484297723,
-0.0233807508,
0.0142100779,
0.001825518,
-0.0213717353,
0.0095788566,
0.0150856366,
0.0013790902,
0.0500267632,
-0.0037621083,
-0.0083855717,
-0.0340322778,
0.0036051895,
0.0594914854,
0.0084507223,
-0.0286507364,
-0.038615454,
0.0635246187,
0.0363639221,
0.0754096657,
-0.0669911802,
-0.0088403048,
-0.0307222698,
0.0501624905,
-0.0134018688,
-0.0109008579,
0.0156065188,
0.0660558045,
-0.042185571,
-0.0127789965,
0.1068707258,
-0.0050382554,
-0.0370734185,
-0.0058788708,
-0.0397919826,
0.0377468951,
0.021566676,
0.0246502999,
-0.0886698589,
-0.06187075,
0.0114686573,
0.015713213,
0.011090395,
-0.0133608328,
-0.0170179326,
-0.0762282237,
0.0279789902,
-0.0170145761,
-0.0273371469,
0.0077025369,
-0.0085144229,
0.0264285505,
0.0075050485,
0.0227916483,
0.0178151429,
-0.0263418891,
-0.0271096118,
0.0320577621,
0.0068511707,
-0.0113785202,
0.0183612704,
-0.0404772162,
0.0565869436,
0.0729284659,
0.0226184521,
0.0466710143,
0.0063609383,
0.0661312044,
0.0075483001,
0.0235574991,
-0.0133761875,
0.0230392385,
-0.0124217849,
-0.0102098016,
-0.0313915461,
0.0448280014,
-0.0100290515,
-0.0051546656,
0.0353060961,
0.0726232007,
-0.0463869199,
0.0511253104,
0.003832757,
-0.0098742144,
0.0618214011,
0.0774905235,
0.0009028597,
-0.0620075315,
-0.0255272165,
-0.020593388,
-0.0131812748,
-0.0020420607,
0.0519192591,
-0.0128127523,
-0.0039668404,
0.0148380417,
-0.0481641926,
-0.0607400239,
0.0243833084,
0.0364994816,
0.0125678172,
0.0267221332,
-0.0361550599,
0.0091946051,
-0.0023255195,
0.0351279043,
-0.0387919657,
0.01998711,
-0.0615569167,
-0.0052760462,
0.0089323195,
-0.0594429262,
0.0029893909,
0.0101357484,
-0.0999566987,
-0.0491638631,
-0.0419116654,
0.0556158535,
0.0225598104,
0.0515466072,
0.0161959957,
-0.019926833,
0.0069922442,
0.0394839421,
-0.033081755,
-0.026831042,
0.0510612428,
0.0390023068,
-0.0123227872,
0.0403175242,
-0.0702563301,
-0.0146549316,
0.0147439502,
0.0324128754,
-0.0352304094,
-0.0492583402,
0.0249311533,
-0.0227151737,
0.0525655672,
0.0309606213,
-0.050773859,
-0.0295813065,
0.0715130493,
0.0062462282,
-0.0096355295,
-0.0775335059,
0.0122880163,
-0.0230313614,
-0.0334984697,
0.0304661505,
0.0245402642,
-0.0285315663,
0.0497397035,
-0.0644226149,
0.0096410699,
-0.0038056369,
0.0061822967,
0.03471848,
0.0055826907,
-0.0288327783,
0.0138986465,
0.0392165743,
0.023976231,
0.0002655436,
0.1045431271,
0.0354772881,
-0.0247919671,
0.0199442394,
0.0280370079,
-0.0600905567,
-0.0340843312,
-0.0380534455,
0.0120488601,
0.0261297803,
0.0179038867,
0.0225984138,
-0.0434923321,
0.0193777401,
-0.0259045213,
0.0498807244,
-0.0040315604,
-0.035811916,
0.0315419249,
0.0169870984,
-0.007967459,
-0.0054583699,
0.0720037073,
-0.0215042364,
0.0196786243,
-0.0008437742,
0.0030944939,
-0.0548480861,
-0.0171671938,
0.0426896252,
0.0432101488,
-0.0204093829,
0.0044928407,
-0.0215472709,
-0.0165018141,
-0.001390635,
-0.0079956185,
-0.0390435793,
-0.0390838273,
0.0456839614,
0.0359276012,
-0.0032917119,
-0.0436623394,
-0.0092366952,
-0.0111077121,
0.0016517156,
-0.0171023216,
-0.0332932696,
-5.892601101e-33,
-0.030495951,
-0.014850717,
0.0381722711,
0.023318572,
-0.0142072877,
-0.0146812974,
-0.0234519485,
0.0045004254,
-0.0324953273,
-0.0032656575,
-0.0315478444,
0.0070682648,
0.0145157818,
0.0424799286,
-0.0134747969,
-0.0151701346,
0.0491571352,
0.0077219922,
-0.0207886137,
0.0304339416,
0.0517569929,
0.027110137,
0.0420663171,
-0.0460247658,
0.0606622845,
-0.0200597811,
0.0201204661,
0.0369150713,
0.0276164152,
0.045427151,
-0.0038446621,
0.0300604589,
0.0084176334,
-0.0317443088,
-0.0344968662,
0.0456692725,
-0.0502168946,
-0.0158512183,
0.0568292812,
-0.00829814,
-0.0013505763,
-0.0141582405,
-0.0194985121,
-0.0248368271,
-0.0083056157,
0.0135154733,
0.0036813226,
-0.0311226305,
-0.0329742543,
0.0086098397,
-0.0468459688,
-0.0013814662,
-0.0055069998,
-0.0762735382,
-0.0187743325,
0.0903259143,
0.011974236,
-0.0443229079,
-0.0021082447,
-0.004839249,
0.0184699837,
0.0037694785,
0.05755594,
0.0621058196,
-0.004169838,
0.0593764335,
0.0101030897,
0.0203888882,
-0.0057412423,
-0.0305284262,
-0.0312989913,
0.0372220688,
0.0127328951,
0.0280372836,
0.014116806,
0.0334961079,
-0.0505105332,
0.029913798,
0.0570881404,
0.0318438448,
0.023378022,
0.0319381692,
-0.0546931177,
-0.0182865299,
-0.0160777234,
0.0125488471,
-0.0269860327,
-0.0527215265,
0.0648525581,
0.0371686742,
0.0300224107,
-0.0158057548,
0.0471745841,
-0.0127691291,
-0.0055547138,
0.016693186,
-0.0092428029,
-0.0009706737,
-0.0141608147,
-0.009975601,
-0.0755269378,
-0.0333548449,
0.0211062487,
0.036338333,
-0.0045204926,
0.0197597668,
0.034642946,
-0.0079087103,
-0.0762267783,
0.0126051856,
-0.0486912951,
0.056797225,
-0.0417323709,
0.0007604157,
0.0578986965,
-0.020048473,
-0.0235058349,
0.1321010441,
-0.0355651937,
-0.0609036833,
-0.0478564054,
-0.0491162874,
0.0123004559,
-0.0154188937,
-0.0037668846,
-0.002381742,
0.00652631,
-0.0022414715,
-0.006728794,
-0.0027652243,
0.0119790956,
0.030124126,
2.594e-7,
-0.0180117935,
0.0113406898,
0.0138789592,
0.0313841067,
-0.0033418669,
0.0514168851,
-0.017522946,
0.007541758,
0.0450754501,
0.0071953037,
0.0031474445,
-0.0062184259,
0.0638838261,
0.0056468039,
-0.1321215481,
-0.0223987475,
-0.0658738464,
-0.0200918559,
-0.0372066908,
0.0024194974,
0.0472825468,
0.0768624023,
-0.0149375433,
-0.0002395899,
0.0157028288,
-0.0494758934,
0.0562357195,
-0.048445683,
-0.0092388205,
0.018532237,
0.0442352183,
-0.0299552344,
-0.0146620693,
0.0266566016,
-0.0022137023,
0.0266452692,
-0.0169689637,
0.0452251174,
0.0129288314,
-0.0421711616,
-0.0193353202,
0.0065638004,
0.0062186858,
-0.0391192362,
0.0472158343,
-0.0721150115,
-0.0131134903,
-0.0304878131,
0.0118070422,
-0.0426030755,
-0.0324312523,
-0.0020681724,
0.0088455798,
-0.0088223219,
0.010080914,
-0.0212481748,
-0.0114529235,
-0.0434042029,
-0.0148945544,
0.0252087172,
-0.032619942,
-0.0765314251,
-0.057056535,
0.0049365968,
0.0145236682,
-0.051235728,
0.0001032117,
2.249011123e-34,
0.0140640559,
-0.0328013711,
0.0317819603,
0.0848478153,
-0.0160405487,
0.0023317023,
0.043498829,
0.0021172364,
0.0840335637,
-0.0705225617,
-0.0293338597
] |
http://arxiv.org/abs/2405.15556v1
|
http://arxiv.org/pdf/2405.15556v1
|
Certifiably Robust RAG against Retrieval Corruption
| 2024-05-24T13:44:25 | 2024-05-24T13:44:25 |
Chong Xiang, Tong Wu, Zexuan Zhong, David Wagner, Danqi Chen, Prateek Mittal
|
cs.LG
|
cs.LG, cs.CL, cs.CR
|
Retrieval-augmented generation (RAG) has been shown vulnerable to retrieval
corruption attacks: an attacker can inject malicious passages into retrieval
results to induce inaccurate responses. In this paper, we propose RobustRAG as
the first defense framework against retrieval corruption attacks. The key
insight of RobustRAG is an isolate-then-aggregate strategy: we get LLM
responses from each passage in isolation and then securely aggregate these
isolated responses. To instantiate RobustRAG, we design keyword-based and
decoding-based algorithms for securely aggregating unstructured text responses.
Notably, RobustRAG can achieve certifiable robustness: we can formally prove
and certify that, for certain queries, RobustRAG can always return accurate
responses, even when the attacker has full knowledge of our defense and can
arbitrarily inject a small number of malicious passages. We evaluate RobustRAG
on open-domain QA and long-form text generation datasets and demonstrate its
effectiveness and generalizability across various tasks and datasets.
|
Certifiably Robust RAG against Retrieval Corruption: Retrieval-augmented generation (RAG) has been shown vulnerable to retrieval
corruption attacks: an attacker can inject malicious passages into retrieval
results to induce inaccurate responses. In this paper, we propose RobustRAG as
the first defense framework against retrieval corruption attacks. The key
insight of RobustRAG is an isolate-then-aggregate strategy: we get LLM
responses from each passage in isolation and then securely aggregate these
isolated responses. To instantiate RobustRAG, we design keyword-based and
decoding-based algorithms for securely aggregating unstructured text responses.
Notably, RobustRAG can achieve certifiable robustness: we can formally prove
and certify that, for certain queries, RobustRAG can always return accurate
responses, even when the attacker has full knowledge of our defense and can
arbitrarily inject a small number of malicious passages. We evaluate RobustRAG
on open-domain QA and long-form text generation datasets and demonstrate its
effectiveness and generalizability across various tasks and datasets.
|
[
0.0396326296,
0.0743543655,
0.0163490139,
0.0193277933,
-0.0141919283,
-0.0256236698,
-0.0186206233,
0.010939477,
-0.0059184469,
-0.0283143371,
0.0140942801,
-0.0449210219,
-0.0040932186,
0.0064175148,
0.0498750024,
0.0266255438,
0.0459781066,
0.0007020577,
-0.0032194338,
-0.0055645206,
0.005399981,
0.0005510636,
0.0199604202,
0.0102422135,
0.0129583757,
0.0051152487,
-0.0382523015,
0.011692211,
0.0017237571,
-0.027930215,
0.0158545226,
0.0354930945,
-0.0211650971,
0.0408945382,
0.0000018223,
-0.0507622436,
-0.0587380454,
0.0457854569,
-0.0277785398,
0.0126168458,
0.0231954865,
0.0265957415,
-0.0216432326,
0.0673820078,
-0.036380358,
0.0153214633,
0.0259639174,
0.0633604452,
0.0919576734,
0.0667386726,
-0.0015176304,
-0.0273964871,
0.0272131022,
-0.0341730118,
0.0368382037,
-0.0482938103,
0.0042095981,
-0.0490970314,
0.0025495414,
0.0357786417,
0.009390167,
-0.0005555534,
-0.0070616072,
-0.008569329,
-0.0039131795,
0.0156691223,
-0.0704323873,
-0.008169625,
-0.0018338548,
0.0307670627,
0.0549680442,
-0.0056994045,
0.0349951312,
0.010925591,
0.0167666823,
0.026243126,
-0.0161961485,
0.0312849171,
-0.0267302431,
0.026049979,
-0.0238278545,
-0.0125341928,
0.0326361433,
0.0064573451,
-0.0054329648,
0.0288275331,
0.0307324901,
-0.0410260186,
-0.0257854741,
-0.0154132508,
0.0096399663,
-0.0296232067,
-0.0279147811,
0.0256036203,
0.0017327735,
0.0362708569,
-0.0695785135,
-0.0644450039,
0.0759388581,
0.0026548188,
0.0062912405,
0.0016498428,
0.0013084521,
-0.007200174,
-0.0088044386,
0.0478499979,
-0.030696949,
0.0201949421,
0.0116715916,
-0.0182098504,
-0.0303028617,
-0.0311799627,
0.0129932594,
0.0808728486,
-0.0629152507,
0.0275529698,
-0.0563094243,
0.0549156889,
-0.0195176695,
0.0300001893,
-0.0390279368,
0.0507461503,
-0.0448186137,
0.0082399519,
-0.0187910013,
-0.0328139514,
0.0192055609,
0.0076625613,
0.0215306804,
-0.0752007514,
-0.0288487282,
0.022373857,
0.0025689921,
-0.0217897147,
-0.0442148894,
0.057713449,
-0.012617684,
-0.0130593963,
-0.0501742959,
-0.1389018595,
-0.0283644553,
-0.008530885,
0.003567222,
-0.052460216,
0.0082040438,
-0.0287998021,
-0.0438455082,
-0.0538989119,
0.0076198592,
0.0692723766,
-0.0405515842,
0.0212779697,
-0.0409758091,
-0.0349388085,
-0.0021133511,
0.0338717885,
0.0177026466,
0.0205587782,
-0.0043287179,
0.046845533,
0.0487476811,
0.0339733586,
0.0523247235,
-0.0118115721,
0.0658643395,
0.0268831067,
0.011123823,
0.0228102449,
0.0401762426,
-0.0505534038,
-0.0369460285,
0.0696969181,
-0.0088364827,
-0.0221101083,
0.060820315,
0.0217747949,
-0.0053787013,
-0.0001889194,
-0.0436635502,
0.0263210032,
0.0004707616,
0.0255560074,
-0.0592443049,
-0.0021552173,
-0.0766160488,
-0.0062637203,
0.0815177113,
-0.0024685632,
0.0029968016,
-0.0353215374,
-0.0248154812,
0.0043575559,
-0.0176419225,
-0.0108210053,
-0.0103679365,
0.037854448,
-0.0319930166,
-0.0232938845,
-0.0229640622,
-0.0867777094,
0.0240951385,
-0.0525628179,
0.0230524875,
0.0225114804,
0.1046713367,
-0.0208847392,
0.0109885558,
-0.0049605696,
-0.0058153486,
0.0387282372,
-0.0267769061,
0.0102855768,
0.0120190289,
-0.0480431393,
-0.068765752,
0.0177107062,
0.0134859178,
-0.0015597668,
0.0308492873,
0.0055674994,
0.0335781202,
-0.0163925439,
-0.0029201719,
0.0227986872,
0.022600038,
-0.0635050312,
0.0603223257,
-0.0047248816,
0.0560690723,
0.0393662341,
-0.0217478964,
0.0619379766,
0.0443282314,
0.0367613845,
-0.0300802998,
-0.0261840932,
0.0257856306,
0.0282767694,
-0.019105766,
0.0069682226,
0.0126286261,
0.0358560681,
-0.0173958484,
-0.0337715708,
0.0060950234,
-0.0343128331,
-0.0175322574,
-0.0765139833,
0.0007527975,
0.0217752308,
0.0220142882,
-0.0609936491,
0.0455204137,
-0.0003628798,
0.0071689915,
-0.0152611267,
-0.0027944455,
0.0190861709,
-0.0366976336,
-0.0322496034,
0.0200917106,
-0.0410649888,
0.0332872085,
0.0346502736,
-0.0084019843,
-0.0262872949,
-0.0293679219,
0.0039238269,
0.0263892226,
-0.0059158769,
0.0660899729,
-0.0137890968,
0.0142982258,
-0.0123712141,
-0.0329117812,
-0.0546625964,
-0.0309600662,
0.0410689227,
0.025302304,
0.0388130061,
0.0227705278,
0.0298544634,
-0.0008356139,
-0.0107788024,
0.0060394164,
0.0416825116,
-0.0044437372,
0.0055128653,
-0.0351005606,
0.040787261,
-0.0452958979,
0.0126672322,
0.0131318448,
-0.0284496788,
0.0336874574,
-0.02373695,
-0.0095156087,
-0.0036293184,
0.0530013628,
-0.0261789393,
0.021755496,
-0.0062101744,
-0.026761448,
-0.0084270537,
-0.0367408395,
-0.0302563179,
0.0009914468,
0.0004841701,
0.069945924,
-0.0198393017,
0.0499886759,
-0.0263374895,
-0.0185026228,
0.0049066199,
0.0469265506,
-0.0397473276,
-0.0263836533,
-0.0315193497,
-0.0118056815,
0.0065102284,
-0.0268080421,
-0.0561670624,
0.0502201878,
-0.0425886102,
0.0191191174,
-0.0278259702,
-0.0053945696,
0.0025251932,
0.0204356406,
-0.0406906307,
0.0044652848,
0.0222937688,
-0.0211171191,
-0.0502211303,
-0.0477435179,
0.0618151575,
-0.0712948814,
0.025002446,
-0.0338366404,
-0.0074857827,
-0.0238599256,
-0.0237659961,
-0.0869313926,
-0.0252954327,
0.0581056066,
-0.0123757273,
-0.014707082,
-0.009493513,
-0.0236061905,
-0.0397367701,
0.039638415,
-0.0127199413,
-0.0310941264,
-0.0235777777,
-0.0089681614,
0.0425833799,
0.023406053,
-0.0227875132,
0.0218137447,
-0.0165521074,
-0.0237436518,
0.1134718135,
0.052325692,
-0.0245850664,
0.0410048403,
-0.0033629558,
0.0123497695,
-0.0208357107,
0.017559249,
0.0561345518,
0.0017076862,
-0.0160305407,
-0.0138172964,
0.0371177197,
0.0393399633,
-0.0069166101,
-0.0346594639,
0.0254743733,
0.0405881628,
0.0016258222,
0.0240043122,
-0.0701976493,
0.0195939206,
0.0022750271,
0.1139391735,
-0.0093183545,
-0.0316556916,
-0.0288641974,
-0.1351444125,
0.05598896,
0.0078260209,
0.0026604468,
-0.0179043245,
-0.0388096534,
-0.0010895437,
0.0432127714,
0.0181979593,
0.0220110323,
0.0214068536,
0.0108410306,
0.0201627873,
0.025417991,
-0.0016896734,
-0.0565077774,
-0.0256295092,
0.0057724933,
0.0780322105,
-0.0068798568,
0.019688962,
-0.0238853414,
0.015302198,
-0.0387069993,
-0.0280162729,
0.003852413,
0.039105434,
0.0177266784,
-0.0204756223,
-0.012579185,
0.0175709687,
-0.0185340289,
-0.0215389188,
-0.0062926896,
0.0194014031,
-0.0218627695,
0.0326613374,
-0.0539256856,
-0.022712497,
0.0365231112,
0.0321663059,
-0.0239723269,
0.012307602,
-0.0371727459,
-0.0189204756,
0.0301059484,
-0.060781341,
0.0089798933,
0.0442170314,
-0.024979068,
-0.0095055457,
-0.0841905177,
-0.026915377,
-0.0483981669,
0.0428105593,
-0.0153601933,
0.0151923792,
-0.0105809402,
-0.0163795091,
-0.0756053329,
-0.0365280844,
-0.0609878041,
0.011992448,
-0.027272258,
0.0190780833,
-0.0007509002,
-0.052664727,
-0.0168909002,
0.0010544916,
-0.0324291438,
-0.0721732453,
0.0193692259,
0.0297878925,
0.0244763214,
0.0033855105,
-0.029335821,
-0.0467099547,
0.0958914906,
0.0820816457,
-0.0356920436,
0.0139825474,
0.0169483721,
-0.030409148,
-0.0189600959,
-0.0030804002,
-0.0981106162,
-0.0153020816,
0.0122736841,
-0.0096013211,
0.0431551784,
-0.0166441351,
0.0040827664,
0.0597066656,
0.0182649866,
0.0050328667,
0.0478870608,
0.0070782327,
0.049276419,
-0.0170521531,
-0.0118536511,
0.0013984166,
0.0510454699,
-0.0200009122,
-0.0346767493,
0.0479822047,
0.0364643671,
-0.020423213,
0.0892593786,
-0.0395666547,
-0.0259563681,
-0.0072566043,
0.0162499882,
0.031566795,
0.0120312767,
-0.0151165444,
0.0306269303,
0.0151215587,
0.0477771945,
-0.0114815151,
-0.0040531596,
-0.0019490995,
0.0030447228,
0.0171569791,
0.0290137082,
-0.1028714851,
0.0015098574,
-0.0064136675,
-0.0190613735,
-0.0083895009,
0.0549296029,
0.0017313439,
0.0042060576,
0.0219100527,
-0.0381533913,
-0.0182282683,
0.0020483793,
-0.0096953055,
0.0092018899,
0.0087193148,
0.0181083698,
0.0319310613,
0.0600242242,
-0.013101114,
0.0094355699,
-0.039792452,
-0.0255884379,
-0.0378816724,
-0.0729140639,
0.005049956,
0.0340442955,
-0.0366954431,
-0.0587863773,
-0.0110183014,
-0.052091226,
0.04197786,
0.0188032892,
0.0312839895,
0.0249330923,
-0.0023946608,
0.0501340814,
0.0129793826,
0.0253223218,
-0.0722858608,
0.0254982933,
-0.0405487381,
-0.0280512311,
0.039609421,
-5.894157162e-33,
0.0101082837,
-0.0643544495,
-0.0092631113,
-0.053012643,
0.0146617396,
0.0060019433,
-0.0158448163,
-0.0304443501,
0.0123520596,
-0.0342661738,
-0.0235746279,
0.0206810683,
0.0270540044,
0.0511256531,
-0.041899588,
-0.0359075367,
0.0160647146,
0.0050046472,
-0.0512524657,
0.0436937697,
0.0915882736,
-0.0320919454,
-0.0097799338,
-0.0668107197,
0.0598765761,
-0.045849923,
-0.0446737781,
0.0068049054,
-0.0282911621,
0.0849371925,
0.0014879151,
0.0322854891,
0.0077192518,
0.0136008319,
-0.0070925439,
0.0336067863,
-0.0478972606,
-0.0339669511,
-0.0188885555,
-0.0404050611,
-0.0175244529,
-0.0344758853,
0.0659056902,
-0.0257832725,
0.0135949384,
0.02222706,
-0.0040023173,
0.0061065648,
-0.0328461863,
-0.085395582,
0.009799283,
0.0227935128,
-0.0048686741,
0.0315311961,
-0.0119155338,
0.0209000949,
0.0540560931,
0.0155767566,
-0.0279503316,
0.0298537612,
0.0077885087,
0.0265090838,
0.0194208398,
0.0104484595,
-0.0221237503,
0.0488886461,
0.0017033683,
-0.0045486693,
-0.0227882974,
-0.0107938526,
-0.0386142954,
0.0159410704,
0.0444996282,
0.1184826344,
0.0468801036,
-0.0109461118,
-0.0480139554,
0.0868582502,
-0.033011429,
0.0266299229,
0.0284403581,
-0.023974428,
-0.0490834601,
0.0072174333,
-0.0326091088,
0.0122424001,
-0.0525964685,
-0.0185974538,
0.0341477692,
0.0533808768,
-0.0693793222,
0.0004526659,
0.0192723181,
-0.029556267,
0.0079303887,
0.0460091308,
-0.0064902734,
-0.039628204,
-0.0387756899,
-0.0064018727,
0.0435032807,
0.0153820058,
0.0585543774,
-0.0288299005,
-0.0208312031,
-0.0034100031,
-0.0020273891,
0.0010536616,
-0.0850157738,
0.0073483675,
-0.0015616198,
0.0773398653,
0.0108300271,
-0.0121036721,
0.0170749202,
-0.016482627,
0.0215774607,
0.0045339414,
-0.0150653785,
0.0293590948,
-0.0328574665,
-0.003873562,
-0.0022134921,
0.0058171721,
0.0353451259,
0.0348257162,
0.0284686256,
0.0355169326,
0.0236083791,
-0.0070526083,
0.0149888145,
0.0049778246,
2.694e-7,
0.0140109416,
-0.0126715163,
0.0671885237,
0.0226930957,
0.0076000788,
0.031522803,
0.0101756481,
0.0659564883,
-0.0379442498,
-0.0126985535,
-0.0390872955,
-0.0263814889,
0.0333619267,
0.0635966659,
-0.0852660015,
-0.0285604894,
-0.0480012186,
-0.0065871347,
-0.040717002,
-0.0208103992,
0.1191585809,
0.0416090786,
-0.0251533967,
-0.0344092697,
-0.0457926914,
-0.0657292604,
0.0277837086,
0.06583976,
-0.0286223404,
-0.004886731,
0.0326501057,
0.0044037928,
0.0456338003,
-0.0159486812,
-0.0624804758,
0.0127546154,
-0.029375555,
0.0372396745,
0.0197219457,
-0.0840026736,
-0.0204681531,
-0.0289797373,
0.0383397341,
0.0247640423,
0.0543277115,
0.0528440177,
0.0153035223,
0.0173780117,
-0.0541022122,
-0.0421866253,
0.0100820027,
-0.0074457796,
-0.0085115787,
0.0088170022,
-0.0109884972,
0.0382117853,
0.0288964212,
-0.0676142797,
-0.0351509228,
-0.0396437645,
-0.0311456118,
-0.0266089048,
0.0076676193,
0.0458748676,
0.0129200909,
-0.035115201,
0.0187139977,
2.705352035e-34,
0.0099709649,
-0.024549162,
-0.0114710974,
0.0208526459,
0.0313726328,
-0.0053333659,
-0.0008665402,
0.0291052889,
0.0286629349,
-0.0602162071,
-0.0385711119
] |
http://arxiv.org/abs/2406.00027v1
|
http://arxiv.org/pdf/2406.00027v1
|
Adapting PromptORE for Modern History: Information Extraction from Hispanic Monarchy Documents of the XVIth Century
| 2024-05-24T13:39:47 | 2024-05-24T13:39:47 |
Hèctor Loopez Hidalgo, Michel Boeglin, David Kahn, Josiane Mothe, Diego Ortiz, David Panzoli
|
cs.CL
|
cs.CL, cs.IR, cs.LG, H.3; H.3.2; H.4; I.7; I.2; I.2.7
|
Semantic relations among entities are a widely accepted method for relation
extraction. PromptORE (Prompt-based Open Relation Extraction) was designed to
improve relation extraction with Large Language Models on generalistic
documents. However, it is less effective when applied to historical documents,
in languages other than English. In this study, we introduce an adaptation of
PromptORE to extract relations from specialized documents, namely digital
transcripts of trials from the Spanish Inquisition. Our approach involves
fine-tuning transformer models with their pretraining objective on the data
they will perform inference. We refer to this process as "biasing". Our Biased
PromptORE addresses complex entity placements and genderism that occur in
Spanish texts. We solve these issues by prompt engineering. We evaluate our
method using Encoder-like models, corroborating our findings with experts'
assessments. Additionally, we evaluate the performance using a binomial
classification benchmark. Our results show a substantial improvement in
accuracy -up to a 50% improvement with our Biased PromptORE models in
comparison to the baseline models using standard PromptORE.
|
Adapting PromptORE for Modern History: Information Extraction from Hispanic Monarchy Documents of the XVIth Century: Semantic relations among entities are a widely accepted method for relation
extraction. PromptORE (Prompt-based Open Relation Extraction) was designed to
improve relation extraction with Large Language Models on generalistic
documents. However, it is less effective when applied to historical documents,
in languages other than English. In this study, we introduce an adaptation of
PromptORE to extract relations from specialized documents, namely digital
transcripts of trials from the Spanish Inquisition. Our approach involves
fine-tuning transformer models with their pretraining objective on the data
they will perform inference. We refer to this process as "biasing". Our Biased
PromptORE addresses complex entity placements and genderism that occur in
Spanish texts. We solve these issues by prompt engineering. We evaluate our
method using Encoder-like models, corroborating our findings with experts'
assessments. Additionally, we evaluate the performance using a binomial
classification benchmark. Our results show a substantial improvement in
accuracy -up to a 50% improvement with our Biased PromptORE models in
comparison to the baseline models using standard PromptORE.
|
[
0.0352298804,
0.1124112234,
0.0009460411,
0.0828505084,
-0.0033949008,
0.0220659021,
0.011738833,
-0.0045666383,
-0.0378544331,
-0.0337111801,
-0.0217122138,
-0.0038803439,
0.0183402002,
-0.0482802391,
-0.0286013149,
-0.0616031438,
0.0322440267,
0.0129882777,
-0.007841601,
-0.0067880303,
-0.0151309548,
0.0273005217,
-0.0219268892,
-0.0097629055,
0.0064336169,
0.0117263468,
-0.0037553099,
0.0024348083,
0.0275190547,
-0.0126308855,
0.0055900691,
0.0632664785,
-0.0243128967,
0.0219225977,
0.0000020486,
-0.0778325498,
-0.0498425514,
-0.0051261326,
-0.0054410077,
-0.0188338179,
0.0114797419,
-0.0464596078,
-0.0080590704,
-0.0106282895,
-0.0516254827,
-0.0545914732,
0.0416045561,
0.0242693219,
0.016355373,
0.0714196339,
-0.0308336373,
-0.0651665181,
0.0379312374,
-0.0478927381,
-0.0047520711,
-0.0069063082,
0.0173258446,
-0.0379685611,
-0.0090521723,
0.0307716709,
-0.0101213623,
0.0310629643,
0.021392392,
-0.0228682831,
0.0240226835,
0.0023793131,
0.0076538348,
-0.038307108,
-0.0008534786,
0.0242855102,
0.0411856771,
0.0374422297,
0.0293421689,
0.0210477505,
-0.0099762408,
0.0459369235,
-0.0020949086,
0.0235164035,
-0.0474401116,
-0.0167237073,
0.0500771888,
0.0259537268,
0.0014625917,
0.0074799405,
-0.0754874051,
0.0582134426,
0.0178588424,
-0.0414233282,
0.0539668202,
0.0224081762,
0.0149525357,
-0.0344867781,
0.0201970711,
0.023510538,
0.0088449558,
0.0212893821,
0.0063124094,
-0.0145484405,
-0.0099756382,
-0.0494872592,
-0.0544464961,
0.0211929493,
-0.0339271836,
0.037907701,
-0.0081846584,
0.0305241551,
0.004781934,
0.0114857266,
-0.0036905974,
-0.0281074792,
-0.0191862788,
0.0145816794,
-0.0488445945,
0.0037450541,
-0.076295577,
0.038296707,
-0.0121971425,
0.0061839866,
0.0026845175,
0.0144723449,
0.0199254248,
-0.0004364675,
-0.0256449711,
-0.0137987481,
-0.0455255695,
0.0101239039,
-0.0439402349,
-0.0061172047,
-0.0466578268,
-0.080144085,
0.0023969128,
-0.0056295027,
0.0268012192,
-0.0693259239,
-0.0001856168,
0.012455089,
-0.0144239571,
-0.0216344967,
0.0217605177,
-0.0397689305,
-0.0091281524,
-0.0481637344,
-0.0044109565,
-0.0485158302,
0.0661077276,
0.021769492,
0.0167518109,
-0.0359838866,
0.0247616358,
-0.0019833497,
-0.0562101305,
0.0566698387,
-0.0004741714,
-0.0141730839,
0.0276234578,
0.0157781746,
-0.0092106387,
0.0461524539,
0.0691921785,
-0.0058500352,
0.0577100404,
0.005794147,
-0.0068316795,
0.053445816,
-0.0063490272,
-0.0524229445,
-0.0062567014,
0.0414190777,
0.0459555052,
0.0420696177,
0.0498194136,
0.0241967086,
0.0034814125,
-0.0076819463,
0.017256096,
0.0027906483,
0.0296596531,
0.0368433706,
0.0573556945,
-0.0357773043,
0.0403689481,
-0.0069327722,
-0.0601846874,
-0.0312801562,
0.0154231451,
0.0180399884,
-0.05581351,
0.0604163408,
0.0045836493,
-0.0128104044,
0.0289407242,
-0.0092629241,
-0.0435624309,
-0.0148875471,
0.0740031078,
0.0365626886,
-0.0442225374,
0.0530279204,
0.0239361934,
-0.0695555285,
0.0146497218,
0.011178378,
0.0396102928,
0.0399488322,
0.0526433997,
-0.0475058779,
0.0102132084,
-0.0351147987,
-0.0138340499,
0.0267268959,
0.0338092297,
-0.0411514454,
0.0517979339,
-0.0439166054,
-0.035495691,
0.0423615538,
0.0143025806,
0.0128588444,
0.0246613789,
0.043593403,
0.0306980964,
-0.006699305,
0.0287619587,
0.0612898208,
-0.0462263972,
-0.0683497936,
0.0020317778,
0.0385493673,
0.0029350871,
0.0423220284,
-0.023007961,
-0.0523503609,
0.0518518835,
-0.0298261326,
0.0258206334,
0.008357021,
-0.0709078908,
-0.0448565111,
-0.0050626891,
-0.0430138633,
-0.0385673344,
0.0019201179,
-0.0317673907,
-0.0077156182,
-0.0348958932,
0.0090720104,
0.0466874577,
-0.0287139174,
0.0405894071,
-0.0012863305,
0.0107420152,
-0.0419158228,
0.0583273284,
0.0079913866,
0.0093631214,
0.0379757397,
-0.0658361688,
-0.0036711202,
-0.02521546,
-0.0404117219,
0.0059881872,
0.0123488978,
-0.0232129898,
0.0399526432,
-0.010544599,
-0.0073858616,
-0.0650013983,
-0.0846785381,
0.0459577888,
-0.0182030369,
0.0321472399,
0.0276873186,
0.0380624942,
-0.0144477161,
-0.0034803841,
0.0077447784,
-0.0310212355,
-0.0275109299,
-0.0037639032,
0.055200953,
0.0214112177,
-0.0291738305,
-0.0227283798,
-0.0037284077,
-0.0251949579,
0.0599692687,
0.0549446605,
-0.0060196109,
-0.0112435976,
-0.0060891504,
-0.0205714628,
0.0256558973,
-0.0158671159,
-0.0112024201,
-0.0127075473,
0.0149854543,
0.0383091234,
0.0605006255,
0.0282998513,
0.0786896795,
-0.0024003456,
-0.005714844,
-0.0012180356,
-0.002211283,
-0.0289603975,
-0.0281020217,
-0.0226016324,
-0.043292474,
0.0079700639,
0.0135223167,
0.0101226065,
0.0213853996,
-0.0176251419,
-0.0138428425,
0.0159658063,
0.0011739922,
-0.0182145014,
-0.0106698321,
-0.0041351621,
0.0403854586,
-0.0345769972,
0.0074200192,
0.0215923749,
-0.0077433679,
0.0141953891,
-0.024527343,
0.0076876758,
-0.0626106039,
0.0584538691,
0.0316621736,
0.0198657587,
0.0334031284,
-0.0401832275,
-0.0458241217,
-0.0229496621,
0.0326738991,
-0.0372217856,
0.0374542624,
-0.0351396762,
-0.020814199,
-0.0175104514,
-0.0380859114,
-0.0792963281,
-0.046652317,
-0.0074295322,
0.0071723736,
0.0029103239,
0.0007432913,
-0.016636664,
0.0264630411,
-0.0243529342,
0.0377682,
-0.0076385587,
-0.0183267314,
-0.0288058091,
0.0779108927,
0.0157109536,
0.0080247847,
-0.0041344897,
-0.0453990363,
0.0117057972,
0.0058646468,
-0.0594223551,
-0.016661251,
-0.0292663444,
-0.0036266602,
0.0060427226,
-0.0724452808,
-0.0124858683,
-0.0131338779,
-0.004302098,
-0.056593962,
0.0142937517,
0.1421599686,
0.0290505812,
-0.0462290756,
-0.0085222637,
0.0187264439,
-0.0133572109,
0.0123767694,
0.0071632373,
-0.099188596,
0.0196840465,
0.0072183367,
-0.0065642283,
0.0048980801,
-0.0470917709,
-0.0086036827,
-0.1381342113,
-0.0069084573,
-0.0201890022,
-0.0392084345,
-0.0502064079,
-0.0562619232,
-0.0035920087,
-0.0753804669,
-0.0268627964,
0.0402576923,
-0.0214389842,
0.0171246696,
-0.017832581,
-0.0223557912,
-0.0197041109,
0.0010559221,
-0.0729823485,
-0.0063432273,
0.0307560042,
0.0458035171,
-0.0365648381,
-0.0247817524,
0.0068308474,
-0.0405714735,
-0.0112715773,
-0.004503625,
0.0956142619,
-0.0103827221,
0.0023446479,
-0.0114258537,
0.0079131378,
0.004187189,
0.0029724538,
0.050546065,
0.0364909098,
-0.0032492627,
0.0138715049,
-0.0038218745,
-0.025735423,
0.023311561,
0.0552192144,
-0.0156184426,
-0.0107365502,
-0.0083305174,
0.0015880413,
0.0388828963,
-0.0036488087,
-0.0109857172,
-0.0116215367,
-0.0351664014,
-0.0677104369,
-0.0556360893,
-0.0480181202,
0.0057484605,
0.0423730351,
0.04582344,
-0.0482869968,
0.0405040495,
-0.0005292592,
-0.0642031506,
0.0545848869,
-0.0148929972,
-0.0304459054,
-0.0375641733,
-0.0090224445,
-0.0036333362,
0.0033579825,
-0.0227219183,
-0.0418875366,
-0.0589335673,
-0.0029568614,
0.0197949372,
0.0099051334,
0.0259128232,
0.0814507306,
-0.0345819555,
-0.0117883459,
0.0317769609,
0.0617211349,
-0.0709536895,
-0.0044951178,
0.0134371966,
-0.0086812004,
-0.0201412067,
0.0133072613,
-0.0188203584,
0.031623777,
-0.0126277115,
-0.0132098775,
-0.0333645716,
-0.0879567116,
-0.0048308717,
-0.0260648001,
0.0883497298,
-0.0431653894,
-0.0372619256,
-0.0432530791,
0.0220994093,
-0.053493388,
0.0177009217,
0.0179126356,
0.0163093098,
-0.0029751696,
0.0369322225,
-0.002300248,
-0.0238223951,
-0.010916708,
0.0113624539,
-0.0637107641,
0.0211890303,
0.0324522071,
0.0224077608,
-0.0002622696,
0.0229202304,
0.0000291017,
0.0525174178,
0.0303692743,
0.0720128715,
-0.0367231034,
0.0305441376,
-0.0111938817,
0.03698726,
0.0211939923,
0.013916784,
-0.0430981703,
0.0323216654,
-0.039234221,
0.0285102408,
0.0328012332,
-0.0392025821,
0.0169811863,
0.0262623653,
-0.0048374319,
0.0292490441,
0.0318773724,
0.0142710414,
0.0287927613,
0.0300017036,
0.0093771201,
-0.0085064964,
0.0304656643,
0.0077252332,
-0.0469227768,
0.0398373082,
0.019116614,
-0.0206756312,
-0.0476641469,
-0.0035997478,
0.0146842683,
-0.0780590028,
-0.0382054038,
-0.0300288629,
-0.0761701986,
-0.0044702808,
0.0340569392,
-0.0365700312,
-0.0338538513,
0.0054429914,
0.0390826911,
0.0379134007,
-0.016946692,
-0.038455613,
-0.0377540775,
0.0505496301,
-0.014851287,
-0.0543071181,
-0.008021282,
-6.507076079e-33,
-0.0461191796,
-0.0165674929,
0.0220439397,
0.0147362296,
-0.0390133597,
0.0032636062,
-0.0195413958,
-0.041621644,
-0.0207831822,
0.0318593308,
-0.0248480178,
-0.0211394653,
0.0256375503,
-0.0007442117,
-0.0119024822,
-0.0006109627,
0.0474093072,
0.0174050033,
-0.0131277973,
-0.0136963706,
0.016979428,
0.0413822606,
0.0557355285,
-0.124195978,
0.0527982488,
0.0154631659,
-0.0126861567,
0.0219627917,
0.0644316897,
0.0355506875,
-0.0156837087,
-0.0178524945,
-0.0048560132,
-0.0769612491,
0.0346515961,
0.0742229596,
-0.1392078251,
-0.0122546973,
-0.0104425866,
0.0608869307,
0.0651876852,
-0.0499621332,
0.0084861824,
0.0214494988,
0.001389531,
-0.0368981026,
0.0787759125,
-0.0226474758,
0.0117625855,
-0.0370256267,
-0.0608719811,
0.0346057042,
0.0143234972,
-0.000234551,
0.0485116504,
0.0475333445,
0.0185543559,
0.0029373043,
-0.0277441796,
-0.0399752893,
0.0283342525,
0.0199356861,
0.00058881,
-0.0118287494,
-0.0279149935,
0.0252090544,
-0.0048438897,
-0.0274437387,
-0.0322755463,
0.0567385815,
0.0027900168,
0.0464603044,
0.0103024654,
0.0701640844,
0.0455352552,
-0.0070557678,
-0.1000156477,
0.0204162598,
0.0315312035,
0.0016530901,
0.0033431773,
-0.0346851386,
-0.0102953408,
0.0094843023,
-0.0126420064,
-0.0443072133,
0.0069996445,
-0.0056969537,
-0.0063557499,
-0.0139905512,
-0.0081724804,
0.0080774017,
0.0176749155,
-0.044250872,
-0.0192972068,
0.0270280931,
-0.0308854412,
0.0236052405,
0.003350351,
0.0143328458,
-0.0321798734,
-0.0148921851,
0.0603925437,
0.0688794404,
0.0304619502,
0.016462272,
0.0180533454,
-0.0121939005,
-0.0847010836,
0.0013809744,
-0.0129765058,
0.0221718028,
-0.0103440247,
0.0450628884,
0.0040636468,
0.0417886637,
0.0050751986,
0.0300875176,
0.0359568186,
0.0147347543,
-0.0051854933,
0.0015038288,
0.0426152535,
0.0155476499,
0.0153648434,
-0.0158755369,
0.0310245939,
0.0402899683,
0.0040476532,
-0.0291423667,
-0.0174788609,
-0.0401754975,
2.845e-7,
0.0616355054,
0.0131530613,
0.0511240438,
0.0205529649,
0.0027190824,
0.0106037837,
0.0553940833,
0.0352323987,
0.0086149443,
-0.0322429091,
0.023507379,
-0.0077563045,
0.0385039672,
0.0666575432,
-0.0894866213,
-0.092476368,
-0.067015551,
-0.0342280082,
-0.054525774,
0.0259668697,
0.0727856383,
0.0948230624,
0.0275342241,
-0.0104453731,
-0.0486941971,
0.0070287674,
0.0218779184,
0.0076663932,
-0.0257069729,
0.048382584,
0.0338125266,
-0.0076482343,
-0.005366846,
0.056700848,
0.0130564263,
0.0448173694,
-0.011354696,
0.0539814681,
-0.0035085229,
0.0286722761,
0.033737842,
-0.0501910187,
0.0005435708,
-0.053341005,
0.071908541,
2.627e-7,
0.0202351995,
0.0036088745,
-0.0590691306,
0.0310495272,
0.0265086591,
0.0053213383,
-0.0037998066,
0.0395077504,
0.04917337,
0.0127929077,
-0.0321873911,
-0.0800176337,
0.0057957345,
0.0415680371,
-0.0060920813,
-0.0542489402,
0.0436216071,
-0.0079011759,
-0.0198448617,
-0.0059089921,
0.0337154903,
2.625069903e-34,
-0.0057724458,
0.00609947,
-0.0423037633,
0.0157441329,
0.0504217595,
-0.0406989343,
-0.0038525269,
-0.003595138,
0.0092967125,
-0.0631051287,
0.0171249509
] |
http://arxiv.org/abs/2406.03199v1
|
http://arxiv.org/pdf/2406.03199v1
|
Bayesian WeakS-to-Strong from Text Classification to Generation
| 2024-05-24T13:33:11 | 2024-05-24T13:33:11 |
Ziyun Cui, Ziyang Zhang, Wen Wu, Guangzhi Sun, Chao Zhang
|
cs.CL
|
cs.CL, cs.AI, cs.LG
|
Advances in large language models raise the question of how alignment
techniques will adapt as models become increasingly complex and humans will
only be able to supervise them weakly. Weak-to-Strong mimics such a scenario
where weak model supervision attempts to harness the full capabilities of a
much stronger model. This work extends Weak-to-Strong to WeakS-to-Strong by
exploring an ensemble of weak models which simulate the variability in human
opinions. Confidence scores are estimated using a Bayesian approach to guide
the WeakS-to-Strong generalization. Furthermore, we extend the application of
WeakS-to-Strong from text classification tasks to text generation tasks where
more advanced strategies are investigated for supervision. Moreover, direct
preference optimization is applied to advance the student model's preference
learning, beyond the basic learning framework of teacher forcing. Results
demonstrate the effectiveness of the proposed approach for the reliability of a
strong student model, showing potential for superalignment.
|
Bayesian WeakS-to-Strong from Text Classification to Generation: Advances in large language models raise the question of how alignment
techniques will adapt as models become increasingly complex and humans will
only be able to supervise them weakly. Weak-to-Strong mimics such a scenario
where weak model supervision attempts to harness the full capabilities of a
much stronger model. This work extends Weak-to-Strong to WeakS-to-Strong by
exploring an ensemble of weak models which simulate the variability in human
opinions. Confidence scores are estimated using a Bayesian approach to guide
the WeakS-to-Strong generalization. Furthermore, we extend the application of
WeakS-to-Strong from text classification tasks to text generation tasks where
more advanced strategies are investigated for supervision. Moreover, direct
preference optimization is applied to advance the student model's preference
learning, beyond the basic learning framework of teacher forcing. Results
demonstrate the effectiveness of the proposed approach for the reliability of a
strong student model, showing potential for superalignment.
|
[
0.0281750932,
0.0637043789,
0.0158031117,
-0.0316087715,
-0.031968005,
0.023175234,
0.0187639073,
-0.0502926409,
-0.0073349662,
-0.0385432169,
-0.0283749867,
-0.1035315245,
0.0106169181,
0.0030734036,
0.0453716852,
-0.0517290123,
0.0137651516,
-0.0259358417,
0.0153675638,
0.0017780459,
0.0080414489,
0.0052367221,
-0.0192444958,
0.0249524973,
0.0201172046,
0.0096343728,
-0.0177150499,
0.0112033281,
0.0481274202,
0.0192194115,
0.0271705519,
0.0689728186,
0.014155658,
0.0411254503,
0.0000016214,
-0.0204874128,
-0.0272147153,
-0.0178938787,
0.0178876612,
-0.0082007898,
0.0326813236,
-0.0044337302,
-0.0220870245,
0.0150951073,
-0.0271438546,
0.0093730753,
0.0657134131,
0.0405636467,
0.0502344631,
0.0663478747,
0.0163687151,
-0.0089954408,
0.0011610984,
-0.0422998853,
0.0338673331,
0.0008207684,
0.020695176,
-0.0384358168,
-0.0033067176,
0.0156966392,
0.0079977205,
0.0142124882,
-0.0081714038,
-0.0373228788,
0.0385725759,
0.011941066,
0.000543437,
-0.0140638081,
-0.0015638121,
0.0325697623,
0.0264640227,
-0.0233131237,
0.0098648416,
-0.0141537702,
0.0157482252,
0.0513238832,
-0.0395373031,
0.0246521309,
-0.0148736788,
0.0187729485,
0.0290273987,
-0.0253655314,
0.0071161152,
0.0244751871,
0.02662885,
0.0220684707,
0.0229270924,
-0.0403109752,
0.005889297,
0.0690200105,
-0.0563273355,
-0.0644978359,
0.0046724589,
0.004074641,
-0.0738167241,
0.0233830567,
-0.0668419302,
-0.019930901,
0.0142140742,
-0.0422553159,
0.0325944647,
0.0210361462,
0.0066617923,
0.003021054,
0.0233641826,
0.039540045,
-0.0123142395,
-0.0053284978,
-0.0299262349,
0.0648540109,
-0.0366412103,
-0.029477207,
0.0199280307,
0.009888459,
-0.0058714156,
-0.0116118742,
-0.0658285245,
0.0054020989,
-0.0237333942,
0.0165180843,
0.0240775738,
0.0467596427,
-0.0665575117,
0.020217428,
-0.0365082659,
0.0569139682,
-0.0184545796,
0.015873149,
0.0670221299,
0.0299071874,
0.0166071653,
0.0203669313,
-0.0036551759,
-0.0369891375,
-0.0169750955,
0.0224117227,
-0.0422291532,
0.0072323759,
-0.0344916172,
-0.0232794844,
-0.0741165206,
-0.0624811761,
-0.0029278684,
-0.0701124445,
0.0114953155,
0.0051389765,
-0.031785015,
-0.0537900552,
0.0022544328,
0.0061100265,
-0.0206988659,
0.0069726072,
0.0312908739,
-0.0086988704,
0.0285823643,
-0.0090881055,
-0.0227798503,
0.045269005,
0.0047332547,
0.0535646081,
-0.0084777512,
0.013570521,
-0.0012913935,
0.0128978072,
-0.0058176797,
0.0286736898,
-0.0359411687,
-0.0221350323,
-0.0046983887,
0.0640872121,
0.0157260206,
0.0683794096,
-0.0196290445,
0.0084945178,
0.094469741,
-0.0240305848,
-0.0085258894,
-0.0058580986,
0.0499732941,
-0.0021448885,
-0.0095268954,
0.0757827088,
-0.0380930975,
0.0317177139,
-0.0284182224,
0.0344016366,
0.0011933098,
0.0490077958,
-0.0005678892,
-0.0488883965,
0.009631562,
-0.0151077909,
-0.0454890616,
-0.011533861,
0.0652013123,
0.0078665325,
-0.0353422686,
-0.0238542594,
0.0299879406,
-0.0802506208,
0.0399239548,
-0.0315419883,
0.0629148483,
0.0698399171,
0.0395018607,
0.0097949719,
-0.0393213816,
-0.0182650127,
-0.0042774654,
0.0512403473,
0.041927658,
-0.0377378352,
-0.0379633494,
-0.0475682802,
-0.0394924283,
0.009259047,
0.031389948,
0.0301826708,
-0.0160789844,
-0.0224161856,
0.0060561253,
0.0030227457,
-0.0522094108,
0.0374287181,
-0.029086465,
-0.0084019415,
-0.0336892605,
0.0219898596,
0.019508535,
0.0099728163,
-0.0303511433,
0.0257508513,
0.0156345461,
0.0035180049,
-0.0235827435,
-0.006643102,
-0.0543202013,
-0.0251658056,
0.0324685909,
-0.0560359433,
-0.0183977094,
-0.0028012735,
-0.0194797944,
-0.0561094321,
0.0206425097,
-0.0217531305,
-0.0268380027,
-0.0123713072,
0.0394845381,
0.0014579672,
-0.0432084873,
-0.074582614,
0.072149016,
0.0313677676,
0.0120568527,
0.0418739617,
-0.0087514604,
-0.009378247,
-0.0349003635,
-0.0197757743,
-0.0169056896,
0.0087955343,
0.015331286,
0.0830062479,
0.0150369573,
-0.0468081906,
-0.0725183785,
-0.0478428788,
0.0184969902,
-0.0172242783,
0.0120126363,
0.048269026,
-0.0204366315,
0.0114537235,
0.0361725315,
-0.0714888573,
-0.0235748626,
-0.003024746,
-0.0086568603,
0.0182471965,
0.0028387248,
0.0171396006,
-0.0639779344,
-0.0092909681,
-0.0052714292,
0.0140641592,
0.0641473606,
0.0223172661,
-0.0469285436,
0.0752657428,
-0.0292393249,
-0.0112572443,
0.0285101775,
-0.0273586474,
0.031241484,
-0.0036503279,
0.0925950184,
0.0829958692,
0.024953803,
0.0306644961,
0.0481575616,
0.0197586603,
-0.0328333713,
-0.01197071,
-0.0705937967,
-0.0114608584,
-0.0229861718,
-0.0338463783,
-0.0146334935,
0.0513970442,
0.0146465478,
-0.0331667475,
-0.0690863505,
0.0344696678,
-0.0087039927,
-0.0126611227,
0.0067198705,
-0.0074608489,
-0.0171190165,
-0.0004480595,
-0.0438295491,
-0.063590534,
0.0843194127,
-0.025888864,
0.0435943566,
-0.0304990467,
0.0005389426,
-0.0017718655,
0.0063753584,
0.0206411947,
0.030785609,
0.0066612931,
0.0238776561,
-0.1323851496,
0.0235875584,
0.0406841859,
-0.0231854431,
0.0194612313,
-0.0554786772,
0.0049660932,
-0.0094123147,
-0.0292416252,
-0.0361790583,
-0.0164482184,
0.0057992912,
0.0139478855,
0.0742981136,
-0.052859813,
-0.015928369,
-0.0018144477,
0.0343010686,
0.0050598378,
-0.0448142998,
-0.0188700445,
-0.0302286483,
0.0377748758,
0.0696049333,
-0.0446842797,
-0.0221108329,
-0.0796551853,
-0.0271523595,
0.039662756,
0.0471805111,
0.0122352866,
-0.0283877943,
-0.0193859152,
0.0228872634,
0.0108932331,
-0.0078780251,
0.0032164843,
0.0636471733,
-0.0125375772,
-0.0342003666,
0.1063278243,
0.013647791,
-0.0325153321,
-0.018380804,
0.0303000882,
0.0952789932,
0.0155126899,
0.0122710867,
0.0078934869,
-0.0081027308,
-0.0163381863,
0.0681937188,
-0.020756714,
-0.0326558799,
0.0115708355,
-0.0980260223,
0.0024654628,
0.0450973734,
-0.0407756194,
-0.0005655352,
-0.0284761786,
0.0120652802,
-0.0163926166,
0.0183951091,
0.0010466033,
0.0126415631,
0.0501449555,
0.0644966885,
0.0286740474,
-0.0659888387,
-0.0459177122,
-0.059294939,
0.0780506805,
0.0884312168,
0.0311881676,
0.0118228877,
-0.0333281159,
-0.000558893,
-0.0094508911,
-0.0028381117,
0.0079033133,
0.0373130962,
-0.0105337715,
0.0092538251,
-0.0269255284,
0.0658084825,
-0.0182438251,
0.0018793307,
-0.0137261851,
0.0705395713,
-0.0070387283,
-0.0260054953,
-0.0154959317,
0.0152810561,
0.0546180233,
0.0638402998,
0.0207867492,
-0.0920736492,
-0.0112253539,
-0.0238609854,
0.0017529292,
-0.0607005991,
-0.0005453441,
0.0271797292,
-0.0231803283,
-0.0084006591,
-0.1055570617,
-0.0037595597,
0.0279023796,
0.0598429777,
0.0227757934,
0.0063250684,
0.0146845244,
-0.0548296943,
0.0142970895,
0.0132427225,
0.0097896019,
0.0107439291,
-0.0581310391,
-0.0650323406,
0.0080537507,
-0.0372212417,
-0.0065288874,
0.0100406148,
-0.0648835078,
-0.1197830439,
0.0010896439,
0.0647860616,
0.0275524557,
0.0344921462,
0.0291143507,
-0.0497532412,
0.0211185478,
0.0538959093,
0.0217882246,
0.0187850427,
-0.0140015436,
-0.0010675134,
-0.0388114639,
0.0120730484,
-0.02361436,
-0.0048431833,
0.0450005122,
0.0341353975,
0.0051081921,
-0.013899724,
0.0348912291,
-0.0239056367,
0.0066686966,
-0.0895780995,
-0.0141267367,
-0.0646845475,
-0.0391630828,
-0.0094640534,
-0.0222938899,
-0.0283412989,
-0.0227892511,
0.0599221438,
0.0046812911,
0.0197439771,
0.0071279025,
0.0136079146,
0.0735512525,
-0.0271945,
-0.0032053816,
0.0196286011,
0.0443118513,
0.0584432036,
0.0082111843,
-0.0090889931,
0.0076135374,
-0.0224240515,
0.0338446535,
0.0210375246,
0.0585346781,
0.0413634628,
0.0030926398,
0.0435790084,
-0.0056022978,
-0.0759381875,
-0.012927508,
-0.0679763034,
-0.0136642959,
0.005950111,
-0.0061385632,
0.0774058551,
-0.0316015147,
-0.0030221578,
-0.0225692131,
0.0482326746,
-0.0002010288,
0.0010377404,
0.0116508901,
0.0173947588,
-0.0123082539,
0.0156159708,
0.0724821165,
0.0075215297,
0.0138727538,
-0.0326550528,
-0.0162927434,
-0.0188043211,
-0.0269955713,
0.0215286724,
-0.0045276331,
0.0084675327,
-0.0241545662,
-0.0253966954,
-0.0326676294,
0.0030635193,
0.0325807594,
0.0217434093,
-0.0194477979,
0.0502112769,
0.0553896613,
-0.0788199902,
0.0482714809,
-0.0366050676,
0.0010773045,
-0.0045463359,
0.0151139759,
-0.0355523601,
-6.051868831e-33,
-0.0421142131,
-0.0501957648,
0.001649576,
-0.044896964,
-0.0129248472,
-0.0430655256,
-0.012149197,
0.023570912,
0.0533011034,
0.0176068563,
0.0271176193,
-0.0150316814,
0.0037738539,
0.0784972608,
0.0287068393,
0.0143551202,
0.0529346243,
-0.0149197578,
-0.0481823795,
0.0362428203,
0.0696347654,
0.0029804402,
0.0396990292,
-0.0563446954,
0.0440691635,
-0.0085322354,
-0.0382852443,
0.0049647032,
-0.0229675677,
0.0380663052,
0.0049940716,
0.0000406238,
-0.0013379349,
-0.0194002278,
-0.0137235057,
0.039593868,
-0.0138091538,
-0.0092142606,
-0.0037532877,
-0.0267175902,
-0.0373672917,
-0.0314695463,
0.0061911792,
-0.0016907635,
-0.0164805148,
-0.0088878814,
0.0361766592,
-0.0233502928,
-0.0124149788,
-0.0353296027,
-0.0656881779,
0.0120281521,
-0.0329175554,
-0.0096160732,
-0.0438286662,
0.0416712835,
-0.0012018677,
-0.0264938045,
0.0027990215,
0.0390518159,
0.0438353568,
0.0420886427,
0.0841628388,
0.0544990189,
-0.0258365888,
0.0637333691,
0.0091861635,
-0.0178859849,
0.0032534425,
-0.0258685239,
-0.0071132714,
0.0473093539,
0.0429615714,
0.0791459084,
0.0016968611,
0.0279382393,
-0.0540780872,
0.0487398766,
0.0187610853,
-0.0400837362,
0.0421887785,
0.0156949721,
-0.0294017512,
-0.0322433673,
-0.0323925316,
-0.084256798,
-0.0045089824,
-0.037319418,
0.0257211048,
0.0188713949,
-0.0405897312,
-0.029272465,
0.0122164115,
0.0064823846,
-0.0189914498,
-0.0015209905,
-0.0681082532,
0.0302673597,
-0.031929031,
0.0044783382,
-0.0824152008,
-0.0322597921,
0.0283926576,
0.0524402373,
-0.0181511417,
0.049490951,
0.0717177391,
0.0252691582,
-0.0433908142,
0.0195644535,
-0.0334079377,
0.0198063087,
0.0139846122,
-0.0089404387,
0.0353479981,
-0.0413937345,
-0.014258719,
0.0436257794,
0.0191276576,
-0.0157147944,
-0.0070851897,
-0.0342300422,
-0.0130060511,
-0.0064188126,
-0.0034138174,
0.0071422956,
0.0030550507,
-0.0157201141,
0.0215193555,
-0.0524594598,
0.0017545533,
-0.0049277456,
2.448e-7,
0.0220062099,
0.0192500353,
0.0247851219,
-0.0302479248,
-0.0001178995,
0.0376908071,
-0.0157563239,
0.0022246137,
0.0111550689,
-0.0238404013,
0.0147438338,
0.0006655261,
0.0432944559,
0.0161471721,
-0.0734101459,
-0.0049206344,
-0.0747692287,
0.0292846449,
-0.0368731096,
-0.0165415164,
0.0816641301,
0.0629749075,
-0.0013171699,
-0.0263630655,
0.0022785033,
-0.0296983067,
0.0717734545,
-0.0175877213,
-0.0139483251,
0.0213493984,
0.0396683961,
-0.0710540041,
-0.027578555,
0.0180037338,
-0.0280363876,
-0.019852262,
0.0134359095,
0.0271772593,
0.0153891882,
-0.0096729621,
0.0196830798,
0.0369411297,
0.0315577202,
-0.0244497694,
0.0423955172,
-0.0662994534,
-0.0209699515,
-0.0253724009,
-0.0139406584,
-0.0190591291,
-0.0188738648,
-0.0216235816,
0.0177103989,
0.0094919745,
-0.0094916923,
0.0440633595,
-0.0120660514,
-0.0419907458,
-0.0230502598,
0.0165208951,
-0.0393498018,
-0.0727092624,
-0.0175260231,
0.0407041498,
-0.0043641473,
-0.0794390514,
-0.0078557432,
2.343454518e-34,
0.0477069393,
-0.0025613024,
0.010819165,
-0.019574251,
0.0243777093,
-0.0246549156,
-0.0211993102,
0.0233311523,
0.0427955016,
-0.0461304933,
-0.0233033746
] |
http://arxiv.org/abs/2405.15525v2
|
http://arxiv.org/pdf/2405.15525v2
|
Sparse Matrix in Large Language Model Fine-tuning
| 2024-05-24T13:12:14 | 2024-05-30T00:08:51 |
Haoze He, Juncheng Billy Li, Xuan Jiang, Heather Miller
|
cs.CL
|
cs.CL
|
LoRA and its variants have become popular parameter-efficient fine-tuning
(PEFT) methods due to their ability to avoid excessive computational costs.
However, an accuracy gap often exists between PEFT methods and full fine-tuning
(FT), and this gap has yet to be systematically studied. In this work, we
introduce a method for selecting sparse sub-matrices that aim to minimize the
performance gap between PEFT vs. full fine-tuning (FT) while also reducing both
fine-tuning computational cost and memory cost. Our Sparse Matrix Tuning (SMT)
method begins by identifying the most significant sub-matrices in the gradient
update, updating only these blocks during the fine-tuning process. In our
experiments, we demonstrate that SMT consistently surpasses other PEFT baseline
(e.g. LoRA and DoRA) in fine-tuning popular large language models such as LLaMA
across a broad spectrum of tasks, while reducing the GPU memory footprint by
67% compared to FT. We also examine how the performance of LoRA and DoRA tends
to plateau and decline as the number of trainable parameters increases, in
contrast, our SMT method does not suffer from such issue.
|
Sparse Matrix in Large Language Model Fine-tuning: LoRA and its variants have become popular parameter-efficient fine-tuning
(PEFT) methods due to their ability to avoid excessive computational costs.
However, an accuracy gap often exists between PEFT methods and full fine-tuning
(FT), and this gap has yet to be systematically studied. In this work, we
introduce a method for selecting sparse sub-matrices that aim to minimize the
performance gap between PEFT vs. full fine-tuning (FT) while also reducing both
fine-tuning computational cost and memory cost. Our Sparse Matrix Tuning (SMT)
method begins by identifying the most significant sub-matrices in the gradient
update, updating only these blocks during the fine-tuning process. In our
experiments, we demonstrate that SMT consistently surpasses other PEFT baseline
(e.g. LoRA and DoRA) in fine-tuning popular large language models such as LLaMA
across a broad spectrum of tasks, while reducing the GPU memory footprint by
67% compared to FT. We also examine how the performance of LoRA and DoRA tends
to plateau and decline as the number of trainable parameters increases, in
contrast, our SMT method does not suffer from such issue.
|
[
0.0513499118,
0.0554200299,
-0.0064881737,
0.0623288266,
0.005846581,
0.0276029725,
-0.0120167742,
0.0380359218,
-0.043309778,
-0.0497144088,
-0.0657955185,
-0.0352076404,
0.0591934547,
-0.0208040569,
0.0402533822,
-0.046061188,
0.0347570814,
-0.0645118952,
-0.0478445105,
-0.0016698125,
0.0038724192,
0.0247861706,
-0.0027896829,
0.0021244022,
0.0201924928,
-0.0042733639,
0.0086008226,
0.0174173955,
0.0240542237,
0.0318388417,
0.0246352553,
0.0609454177,
0.0152126784,
0.0172009096,
0.0000021722,
-0.0158729367,
-0.0023203988,
-0.0312062502,
0.0116366372,
0.0693725497,
0.072679624,
0.0067331935,
-0.0056894883,
-0.000087341,
-0.0203057658,
-0.0455019772,
0.0603271052,
0.0019518023,
0.025758246,
0.0640258864,
-0.0099177025,
-0.028905347,
0.0138785401,
-0.044510521,
-0.0342083797,
-0.0820004493,
0.020075297,
-0.0216897689,
-0.016468389,
-0.0181421693,
-0.0372468382,
0.0100488421,
-0.0274287723,
-0.0126374103,
0.0069501265,
0.0104980562,
-0.0467531942,
-0.0357530601,
0.004549636,
0.0740401894,
-0.0615816563,
-0.0077531119,
0.0260247402,
-0.0119244112,
-0.0176042728,
-0.0133664785,
-0.0066220001,
0.0002808275,
0.0168979242,
0.0130415121,
0.0086467611,
-0.0502053648,
0.0390985347,
0.0332459733,
-0.0103376778,
0.0366127566,
0.0474121831,
0.0037254023,
-0.0136744911,
-0.0128916735,
-0.0262782574,
-0.0534816012,
-0.0086245602,
0.0114480043,
-0.0451392457,
-0.0179779343,
-0.0853165537,
-0.0357174948,
0.0053742193,
0.0296601932,
0.1259508431,
0.0519829839,
0.0043447227,
0.0516014732,
-0.035862349,
0.0236878321,
-0.0063140383,
0.0137942266,
-0.0306293331,
0.0115639772,
0.0153352004,
0.0320614129,
-0.0255272891,
0.012749156,
-0.050080739,
-0.0067385146,
-0.0315670334,
0.0039550308,
0.0019869853,
-0.0057369131,
-0.0226417538,
-0.0095430017,
-0.0244000666,
-0.0142673273,
0.0195741784,
0.0302849561,
-0.0324287191,
0.0342454053,
-0.0294977482,
-0.0498591699,
-0.0145776179,
0.0243082717,
0.0441889912,
-0.0267224964,
0.0343818702,
0.0281687845,
-0.0142791066,
-0.0632646605,
-0.0534556769,
0.0286334343,
-0.0022167305,
-0.104469128,
-0.0183721781,
-0.0393399708,
-0.0043199016,
-0.0078031872,
0.0322983414,
-0.0393642336,
-0.020987248,
0.0151933106,
-0.0509745069,
0.056754414,
-0.0276356079,
0.0004383674,
0.0278376471,
-0.0066034165,
0.0402037576,
0.0535369106,
0.0482205786,
0.0548324063,
0.0215834714,
-0.0119706634,
-0.0262147728,
0.0293389782,
-0.0068116789,
0.0486740433,
-0.0235115066,
-0.0012662048,
0.0293794014,
0.0122166425,
0.0377529673,
0.0338429958,
-0.040174745,
-0.0031468936,
0.0519837439,
0.0515071228,
0.020449819,
0.0491402969,
0.0337478071,
-0.0376427397,
0.0233864691,
0.0099704135,
-0.0288770478,
0.028137017,
-0.0374820642,
0.0025054046,
0.043662563,
0.070914261,
-0.0270966068,
-0.0236342214,
-0.0147617776,
-0.0058172122,
-0.009448112,
0.0312507376,
0.0264178105,
-0.0148419635,
-0.0792715997,
0.0188481137,
0.0588658564,
-0.0371824093,
-0.0325237662,
-0.0685179979,
0.0210010242,
0.053025011,
0.0135475425,
-0.0199665613,
-0.0556599386,
-0.0109691331,
0.0604760237,
0.0085628778,
0.0095416289,
0.012923005,
0.0130499657,
-0.0949504673,
0.0038434153,
-0.0165625121,
0.0337770768,
0.0031321966,
0.0149512803,
0.0349217579,
0.0236550849,
-0.0081576435,
-0.0303759798,
0.0387560762,
0.0148085859,
0.0104623986,
0.0274313074,
0.0089236395,
-0.0004700558,
-0.015121771,
0.0053982721,
-0.0576027595,
0.0208327118,
-0.0195670091,
0.0616251566,
-0.0752550587,
-0.0225421172,
0.0177209266,
0.0361939669,
0.0460690297,
0.0098020174,
0.0693097413,
-0.0483456478,
-0.029070016,
0.0174640846,
-0.0324376784,
-0.0049102525,
-0.034923438,
0.0246751774,
-0.0548354723,
0.0427576751,
-0.0263529755,
0.0108854249,
-0.0142134838,
-0.0217385013,
0.0271363892,
0.022837529,
0.023850454,
-0.0437434763,
-0.0553957671,
0.0360854752,
0.0181206055,
0.0312263034,
0.0069500967,
0.0372841433,
-0.0098087508,
-0.0412349366,
-0.0647585019,
0.0378961265,
-0.0603059754,
0.0141309006,
0.0379232652,
0.0166480169,
-0.00833584,
-0.0118483556,
0.0255365726,
0.0500138775,
-0.0328232758,
0.017201975,
0.0060915765,
0.0122418646,
-0.0307440441,
-0.0731996149,
-0.0282307528,
0.0436117314,
0.0523326695,
-0.0114410734,
-0.028175164,
-0.0612014271,
0.0313887671,
-0.0252191704,
0.0153595963,
-0.0117002558,
-0.0467834547,
-0.0199844688,
0.0009136961,
0.0352919921,
0.0667120516,
-0.0107123284,
0.0445482582,
0.060602773,
0.0187447779,
0.0071211825,
0.0041545047,
-0.0002447498,
-0.0673932433,
-0.0359041728,
0.0043352274,
-0.0145096229,
0.0534020625,
-0.036324434,
-0.0086288629,
-0.0403075442,
-0.0006270114,
-0.0247414224,
0.0048409156,
-0.0117718577,
0.0238827895,
-0.0089692334,
0.0210540816,
-0.0199762583,
-0.0621994473,
0.0337207019,
-0.0460324176,
0.0006936371,
-0.0203561876,
-0.0320917033,
-0.0631706193,
-0.0134734027,
-0.0078471741,
0.022372853,
-0.0415903814,
-0.0049314434,
-0.0160149653,
-0.0363771133,
0.0187680498,
-0.0064156274,
0.0637530312,
-0.0268752761,
0.036267627,
-0.0353414379,
-0.0344592258,
0.017718792,
-0.0157294348,
-0.0282946154,
0.0190908574,
0.026192423,
-0.0590114072,
-0.0022404008,
0.0059864121,
-0.0172946192,
0.0649510399,
-0.0278245732,
0.0030594112,
-0.0392300859,
0.005731334,
0.0728883967,
0.0492522866,
0.020260131,
-0.0243676472,
-0.0107673528,
0.0018845639,
0.0183954407,
-0.0155016743,
-0.0563522838,
0.0398023427,
0.0473107658,
0.0082343277,
-0.0215795916,
0.0240883641,
0.0389514267,
-0.042890165,
0.0063888198,
0.0822407901,
0.0048622061,
-0.0683217943,
0.0003991226,
0.0211557727,
0.0699317306,
0.032662645,
0.0447595119,
-0.0584934503,
-0.0459597893,
0.0056525441,
0.0532460138,
0.0265538804,
-0.0172831677,
-0.002929088,
-0.0589869283,
0.0461501591,
0.0189379565,
-0.0003029596,
-0.0367999114,
-0.0289304368,
0.0321633369,
-0.0072523654,
0.0203647241,
0.0085404031,
-0.0492183417,
-0.0102161737,
0.0148180816,
0.0087928064,
-0.085131444,
-0.0142350327,
-0.0054454668,
0.023349531,
0.0182865392,
0.0207056198,
0.0191304926,
-0.0261943042,
0.0454195328,
-0.023090668,
-0.012903004,
-0.0654390529,
0.0186988264,
0.0239896178,
-0.0353578217,
0.0127058877,
0.0331186987,
-0.0161315631,
-0.0071138465,
-0.0351399481,
0.0503508598,
-0.0041893125,
-0.0097319959,
0.060073927,
-0.0332354903,
-0.007590828,
0.0287982207,
0.0366471559,
-0.0241150353,
-0.0100009739,
0.0056791883,
-0.0378307626,
-0.0269035716,
-0.0110041872,
-0.0414577797,
-0.0052620005,
0.0201893877,
-0.0205332544,
0.0104107177,
0.0677256584,
0.0206904225,
-0.0259949453,
0.0228347536,
-0.0006944548,
0.0071693202,
-0.0318285152,
0.0344444849,
-0.0552995428,
0.0086758612,
-0.0557933822,
-0.0494880565,
0.040498551,
0.0041582696,
0.0071652289,
0.0109452521,
-0.1347192973,
-0.0627119765,
0.0045551895,
0.0486543998,
0.0125692328,
0.0235809088,
0.0132589648,
-0.0894267559,
0.0002526208,
0.0211396329,
-0.0110509181,
0.0949515551,
-0.0021758168,
-0.0710842684,
-0.0219901297,
0.0494873039,
-0.024579851,
-0.0083181458,
0.0635649934,
0.0308699291,
0.0092106396,
-0.0258884765,
-0.0156242782,
-0.0023568659,
0.0167676639,
-0.030747937,
-0.0301624164,
-0.0593053885,
0.0463407338,
0.0297610033,
0.0186051745,
-0.0209209379,
-0.0410885289,
0.0076742084,
0.0050489707,
-0.0382902473,
0.019021662,
0.0098088877,
0.0313198529,
-0.0471792407,
0.0287357029,
-0.0077242833,
-0.0070139202,
0.0518721156,
0.0219234154,
-0.0213167351,
0.059742555,
0.0147261843,
0.0267763063,
-0.0315861478,
0.003283707,
0.0014031891,
-0.0045280042,
-0.0951568186,
0.0682870448,
0.0068697063,
-0.0921307579,
-0.029182177,
0.0385594293,
0.0835230574,
-0.0308269933,
0.0519094206,
0.0138108563,
-0.0288625509,
-0.014233022,
0.0040940228,
-0.0030101561,
0.0079104966,
-0.0073438426,
0.0171498302,
-0.0262809414,
0.0681403279,
0.0534328818,
0.0373845957,
0.0575245284,
0.0132784201,
0.0041314727,
-0.0186279248,
-0.0258562341,
-0.0235106722,
0.0285661165,
-0.0222953241,
-0.0200964175,
0.005638178,
-0.066665642,
0.0073373038,
-0.0266693179,
-0.0128436219,
-0.0014635489,
0.0248629712,
0.0259114504,
0.0107211582,
0.0413763039,
-0.0347850062,
-0.0472746231,
0.0218617823,
-0.0314237587,
-0.0103601301,
-6.718479922e-33,
-0.019321274,
0.0074683242,
0.0009921951,
0.0175608359,
-0.0043568574,
-0.0144828418,
-0.0378916636,
0.0037966406,
-0.0352874734,
0.0116930632,
-0.0257860012,
-0.0125566311,
0.0118627027,
-0.0030621428,
-0.0318696313,
-0.0026256472,
0.0644463673,
0.0035260739,
-0.01943589,
0.0282598361,
-0.0050643841,
0.051306963,
0.0486610942,
-0.0256060679,
-0.0599695742,
-0.0000442794,
-0.0286162179,
-0.0026858561,
-0.0157682896,
0.0441042148,
-0.0365653448,
0.0410842486,
0.0092408611,
-0.0037055947,
-0.0200005434,
-0.0215010606,
-0.0721576884,
-0.0009918243,
0.043421004,
0.0060609351,
-0.0111421552,
-0.0547953546,
0.0249902923,
0.0302746873,
-0.0462720059,
-0.0101294853,
-0.0033799252,
-0.0225344654,
-0.0497704223,
-0.0057929158,
-0.0762553215,
0.0217869114,
-0.0088808443,
-0.0113281952,
0.0423481576,
-0.0461118557,
-0.0168768708,
0.0126252696,
-0.0428236462,
-0.0207082164,
0.0555717796,
0.0942768827,
0.0461829863,
0.0457789414,
-0.0215534158,
0.0333668962,
-0.0048546828,
0.0382029116,
0.0368465632,
0.016554242,
0.0497376174,
0.0236661583,
0.0113246338,
0.0221677236,
0.0338478312,
-0.0405326225,
-0.0991806015,
-0.0377628841,
-0.0097227963,
0.05504651,
0.02446693,
0.0279032346,
0.0074963118,
-0.0211854186,
-0.0480657481,
0.0050994609,
-0.0574909002,
0.0104709044,
0.0208576135,
0.0331902131,
-0.054620754,
0.017344309,
0.0145825073,
-0.0271263588,
0.0159859695,
0.0010571402,
-0.0346061885,
0.0380036905,
0.0137212435,
-0.0037458171,
-0.0262192972,
-0.0783401355,
0.0192261338,
-0.0303542074,
0.0265375469,
-0.025832437,
0.0405180231,
0.0498656444,
-0.0681529567,
0.0057071154,
-0.0098345717,
0.0286131296,
0.0295146015,
0.077825658,
-0.0064700134,
-0.0065492997,
-0.0048109922,
0.0983699858,
0.0057861661,
0.0176285747,
-0.0067310743,
-0.0380588248,
0.0219699517,
0.019533433,
-0.0210754555,
0.0177631509,
-0.0296465792,
0.0148842633,
0.0178581197,
-0.0404565781,
-0.0237232465,
0.0113561759,
2.778e-7,
-0.0065913936,
0.0748405904,
-0.0437804312,
-0.0506551303,
-0.0034370874,
0.0040573711,
0.0067949914,
-0.023540251,
0.0227450002,
-0.010427841,
-0.0039980006,
0.0351481363,
0.0446888059,
-0.0038715384,
-0.0563394316,
0.0511335582,
-0.0611889698,
0.0479182042,
-0.0563436635,
-0.0561265573,
-0.0430404991,
0.1126208305,
0.0248273741,
-0.0185662638,
0.0137453964,
-0.0112980343,
0.0486890376,
-0.0198538806,
-0.0092594195,
0.0340622887,
0.0553615242,
0.0226080772,
-0.049022276,
-0.0317589119,
-0.0114370398,
0.0356638618,
-0.0246124249,
0.0680078864,
-0.0441411734,
0.0308812447,
0.0304624159,
0.0302111059,
-0.0103701307,
-0.0445993394,
0.0671107322,
0.0013781582,
0.0033697069,
-0.0683954284,
0.0350472555,
-0.0168276951,
0.0269224644,
-0.0623273551,
0.0141168023,
0.0404912122,
0.0271209106,
-0.023045646,
-0.037275333,
-0.0776772052,
0.0375604406,
0.0385419987,
-0.0268998016,
-0.0361582302,
-0.0723901764,
-0.0339136273,
0.0537588112,
-0.0250204895,
0.0347828194,
2.491227244e-34,
0.0352418944,
-0.0139359394,
0.0307419281,
0.0146045154,
0.0256473403,
0.0035390824,
-0.0237709414,
0.0279293582,
-0.0125809992,
-0.1205100492,
-0.0095647145
] |
http://arxiv.org/abs/2405.15523v1
|
http://arxiv.org/pdf/2405.15523v1
|
Mosaic Memory: Fuzzy Duplication in Copyright Traps for Large Language Models
| 2024-05-24T13:05:05 | 2024-05-24T13:05:05 |
Igor Shilov, Matthieu Meeus, Yves-Alexandre de Montjoye
|
cs.CL
|
cs.CL, cs.LG
|
The immense datasets used to develop Large Language Models (LLMs) often
include copyright-protected content, typically without the content creator's
consent. Copyright traps have been proposed to be injected into the original
content, improving content detectability in newly released LLMs. Traps,
however, rely on the exact duplication of a unique text sequence, leaving them
vulnerable to commonly deployed data deduplication techniques. We here propose
the generation of fuzzy copyright traps, featuring slight modifications across
duplication. When injected in the fine-tuning data of a 1.3B LLM, we show fuzzy
trap sequences to be memorized nearly as well as exact duplicates.
Specifically, the Membership Inference Attack (MIA) ROC AUC only drops from
0.90 to 0.87 when 4 tokens are replaced across the fuzzy duplicates. We also
find that selecting replacement positions to minimize the exact overlap between
fuzzy duplicates leads to similar memorization, while making fuzzy duplicates
highly unlikely to be removed by any deduplication process. Lastly, we argue
that the fact that LLMs memorize across fuzzy duplicates challenges the study
of LLM memorization relying on naturally occurring duplicates. Indeed, we find
that the commonly used training dataset, The Pile, contains significant amounts
of fuzzy duplicates. This introduces a previously unexplored confounding factor
in post-hoc studies of LLM memorization, and questions the effectiveness of
(exact) data deduplication as a privacy protection technique.
|
Mosaic Memory: Fuzzy Duplication in Copyright Traps for Large Language Models: The immense datasets used to develop Large Language Models (LLMs) often
include copyright-protected content, typically without the content creator's
consent. Copyright traps have been proposed to be injected into the original
content, improving content detectability in newly released LLMs. Traps,
however, rely on the exact duplication of a unique text sequence, leaving them
vulnerable to commonly deployed data deduplication techniques. We here propose
the generation of fuzzy copyright traps, featuring slight modifications across
duplication. When injected in the fine-tuning data of a 1.3B LLM, we show fuzzy
trap sequences to be memorized nearly as well as exact duplicates.
Specifically, the Membership Inference Attack (MIA) ROC AUC only drops from
0.90 to 0.87 when 4 tokens are replaced across the fuzzy duplicates. We also
find that selecting replacement positions to minimize the exact overlap between
fuzzy duplicates leads to similar memorization, while making fuzzy duplicates
highly unlikely to be removed by any deduplication process. Lastly, we argue
that the fact that LLMs memorize across fuzzy duplicates challenges the study
of LLM memorization relying on naturally occurring duplicates. Indeed, we find
that the commonly used training dataset, The Pile, contains significant amounts
of fuzzy duplicates. This introduces a previously unexplored confounding factor
in post-hoc studies of LLM memorization, and questions the effectiveness of
(exact) data deduplication as a privacy protection technique.
|
[
0.0415169783,
0.0770185664,
-0.0498348027,
0.0613114201,
-0.0604199246,
0.0104579246,
0.0433632806,
-0.0071886261,
0.0070127244,
-0.0495250933,
0.0061377008,
-0.0162029825,
0.0025875263,
-0.0121847894,
0.0013795013,
0.0497473739,
0.0559143946,
-0.0272708554,
0.0277804863,
-0.035958834,
-0.0199110806,
-0.0017282593,
0.0242133923,
0.0274931416,
-0.0066906731,
0.0071828146,
-0.0019192498,
-0.0129384315,
0.0025390792,
-0.0210702494,
-0.0103503242,
0.0842018649,
0.0205200743,
0.1053562686,
0.0000019564,
-0.0656693131,
-0.0421902277,
0.0208453666,
0.0013609666,
0.1012015268,
0.0129864793,
0.0534560718,
-0.0064123692,
-0.0010379544,
-0.018763613,
0.0522379354,
0.0064807246,
0.0094909444,
0.0578742176,
0.0624502636,
-0.0123703433,
-0.0211889893,
0.0129763409,
0.0006634662,
0.0427391827,
-0.0297444724,
0.0190876517,
0.0055481018,
-0.0035313938,
-0.0144436629,
-0.0104403887,
0.0564581044,
-0.0033760599,
0.0368942432,
-0.0499540046,
0.0263572801,
-0.0433616154,
0.006558157,
0.0302547906,
0.0094993375,
-0.0468320474,
-0.0529810078,
0.029673256,
0.0120690791,
0.0551234744,
0.0294162929,
-0.0139479954,
0.0577353649,
-0.0027811064,
0.0342283323,
0.0042556943,
-0.0383111127,
0.0352948867,
-0.0211202241,
-0.0238039698,
-0.0160004683,
0.02196135,
-0.0374783799,
-0.0095173903,
-0.0304513704,
0.0380853973,
-0.0343628041,
0.0207979791,
-0.0200519208,
0.0453928225,
0.008168255,
-0.0992089435,
-0.0267930664,
0.0487815216,
-0.0649386495,
0.0431870073,
0.0266711563,
0.011423259,
0.0382703245,
-0.0560681671,
0.0453933813,
-0.035440024,
0.0605571531,
0.0182891488,
0.0088366522,
-0.0336768664,
-0.0211054366,
0.0020355924,
0.0783208981,
0.0712454543,
-0.0134541709,
-0.0383078679,
0.0124505386,
0.0005255233,
0.0608418062,
-0.0792070031,
-0.0127480561,
0.0192046296,
0.0202065725,
0.0181256626,
-0.0572392382,
0.0259744953,
0.0024016581,
0.0048248079,
-0.0506196544,
-0.0319780521,
-0.0108929742,
0.0064394656,
0.0004038183,
0.008083703,
0.039326746,
0.0039508631,
-0.0509053171,
-0.0114443442,
-0.062092755,
0.0149899563,
-0.0183782708,
-0.0146906786,
-0.0540012196,
0.0053667161,
-0.0055286246,
-0.0001835371,
-0.0573985986,
-0.0005797522,
0.0303559024,
-0.0692056045,
0.0305974577,
0.0063973162,
-0.010571206,
0.0009238542,
0.0303294659,
0.0025110578,
0.0350952819,
0.004449415,
0.0445410088,
0.07199651,
-0.0102837477,
0.0597708561,
0.0442944914,
0.0357021317,
0.0057975245,
0.002170065,
-0.0229663905,
0.030540226,
-0.0446705036,
0.0138227129,
0.0467937663,
-0.0090942085,
-0.0683418289,
0.0863717273,
0.0683107525,
-0.0109539498,
0.0442935601,
-0.0465304404,
0.0203288477,
0.0180162694,
0.0352414064,
-0.0385202803,
0.0000935108,
-0.0432936773,
-0.0009307203,
0.011071478,
0.0199791286,
-0.0025899764,
0.0240785405,
-0.0104861781,
-0.0413974784,
0.0020402404,
-0.0010816096,
0.0051384917,
-0.000824491,
0.0140632745,
0.0916787684,
0.0026971565,
-0.0551264472,
-0.0064465543,
-0.0496597029,
0.0014403394,
0.0246555079,
0.0662504807,
-0.0149199599,
0.0160903297,
-0.0195335373,
-0.0719794333,
-0.0063877618,
0.0206847247,
0.008205628,
0.0132125877,
-0.063840732,
-0.0281062797,
-0.0209032651,
0.0263002422,
0.0191995539,
0.0238209087,
0.0327886604,
0.026635889,
0.0357755423,
-0.0365719236,
0.0118247792,
-0.012781526,
-0.0338593647,
-0.0172464158,
-0.0128879463,
0.0447492935,
0.0095266718,
-0.0498997271,
0.0188252609,
0.0454097316,
0.0029631592,
-0.016267499,
-0.0336901061,
-0.0122930389,
-0.0301845856,
0.0321699642,
-0.0040664114,
0.0080046551,
0.0229888018,
-0.0024310143,
0.0035164652,
0.0089210207,
-0.0528780371,
0.049061913,
-0.0356324874,
0.044407852,
-0.0526963174,
0.0643850267,
-0.0234089252,
0.0505210273,
-0.0269425288,
-0.0005206914,
0.020656582,
-0.0564874671,
0.0635594726,
-0.0569766834,
-0.0387268662,
-0.0119078783,
-0.0482919663,
0.0363048166,
0.0052384208,
-0.0190205816,
-0.008581223,
-0.0364212058,
0.0125846844,
-0.0177824851,
0.0035598411,
0.0033352494,
-0.0010674384,
-0.0248323847,
-0.034053456,
-0.0075617214,
-0.0214843228,
-0.0708515123,
-0.0329099298,
0.020592548,
-0.0205012709,
-0.0221416298,
-0.065843083,
0.0310520139,
-0.0085451705,
0.0056195259,
0.0513475388,
-0.022521643,
-0.0386478268,
-0.0311773736,
0.0655309483,
-0.0359103195,
0.0045491159,
-0.0048217992,
-0.0086775208,
-0.0019351632,
0.0072940732,
-0.0376729518,
0.0482851714,
0.0093773501,
0.0031909703,
0.0596919842,
-0.0177619699,
-0.0456466563,
-0.0232511349,
-0.0015488392,
-0.0300561991,
-0.0342841893,
-0.0370768309,
0.0389746055,
-0.0056293951,
0.0240428336,
-0.0181383472,
-0.0052785384,
-0.0098991953,
-0.0366483964,
-0.0180137828,
-0.007601202,
-0.0531470813,
-0.0489185974,
0.044040069,
0.0194860008,
-0.0586314164,
0.0315181538,
-0.0648447275,
0.0188867487,
-0.0285827462,
-0.0025044973,
-0.0086031137,
0.0438070707,
0.0013862014,
0.0375713967,
0.0189948119,
-0.0394684188,
-0.0312301852,
-0.0106147993,
0.0851371735,
-0.0710150823,
-0.016560534,
-0.0011121753,
-0.009384891,
-0.014524213,
-0.0208478048,
-0.0378543735,
-0.0480829626,
0.0395057909,
-0.0058387504,
-0.0041642422,
-0.0002325538,
-0.0394815728,
-0.0141711263,
0.0553186983,
0.0138865085,
0.0009015956,
0.0260663088,
-0.0162167326,
0.0554629192,
0.0282412432,
-0.0163832027,
0.0148055935,
-0.0395744741,
-0.0388222523,
0.061418023,
0.0422876216,
0.0385471731,
0.0434964597,
0.0092617925,
0.0322855897,
-0.0127195055,
0.0087966267,
0.0515987203,
0.0098172501,
-0.0209066123,
0.0217025038,
0.0718014538,
0.0356380045,
-0.0535509251,
-0.0086256424,
0.0107540917,
0.002088028,
0.0156809781,
-0.0317673013,
-0.0946651101,
0.0025157172,
0.0403940156,
0.0093964254,
-0.0383635573,
-0.0222669858,
-0.0088484306,
-0.1100638956,
0.0026759135,
0.0220734272,
-0.0485228524,
-0.0405506268,
0.0282224417,
-0.0242673196,
0.0156443212,
0.0057080793,
-0.0256018117,
0.0072362656,
0.0025693621,
0.001266427,
-0.0089139044,
-0.0568855107,
-0.0071290266,
-0.0502348952,
0.0511445664,
0.0146820741,
-0.0062397928,
-0.0200627483,
-0.0124470461,
0.029282093,
-0.0436361693,
-0.0036196788,
0.0197745562,
0.0630758479,
0.0027658732,
-0.0085338578,
-0.0151868882,
0.0293292161,
0.0136323329,
-0.0469962135,
0.0130328685,
0.0487749353,
-0.0015052011,
-0.0371286497,
0.001059887,
-0.0069562281,
-0.0170799848,
0.05755141,
0.0490289927,
-0.0076458757,
0.0139231263,
-0.0257234089,
0.0286315288,
-0.0082276119,
-0.0293717422,
0.0340820178,
-0.0524019152,
-0.0390255414,
-0.0463659316,
0.0109974761,
-0.0333423503,
0.0224669464,
0.0402060337,
0.0375893749,
0.0317613594,
-0.0198864602,
-0.0520737581,
-0.0320517495,
-0.0473190211,
-0.0055036126,
-0.0510106087,
-0.0117024416,
-0.0376131497,
0.0353460908,
-0.0242402665,
0.0127182975,
-0.07346908,
-0.0653426051,
0.0543585867,
0.0144545129,
0.0255821422,
0.0131333107,
0.0095321899,
-0.0774756223,
0.0984140113,
0.0528331697,
-0.0566584952,
0.0313875191,
0.0395365283,
-0.0544264577,
-0.006077827,
0.0227418467,
-0.0357144512,
-0.0441635549,
0.0672637969,
-0.0187990256,
-0.0233562216,
-0.0188761689,
0.0163257439,
0.0384260975,
0.011040031,
-0.0026090785,
0.0815321058,
-0.0045716972,
0.0452709049,
-0.0015906175,
-0.0142228073,
0.0066042221,
-0.0208301451,
0.0335945003,
-0.0136802634,
0.0115267774,
0.026331,
0.021621827,
0.0396376364,
0.0151690608,
-0.0228863955,
-0.0169981979,
0.02783278,
0.0186277069,
-0.0042134784,
-0.0370205268,
0.0556434691,
-0.0026771822,
0.0304717254,
-0.0190833453,
0.0369130857,
0.0164569337,
0.025339989,
-0.0182614084,
0.0383024327,
-0.0453292057,
-0.0426206551,
0.0080284933,
0.0246209558,
0.0567126237,
-0.0248500742,
0.0310365278,
-0.0397116207,
-0.0091772377,
-0.0113879452,
-0.0057688835,
-0.0356872119,
-0.0358632579,
-0.0231345575,
-0.0095042065,
-0.0368024036,
0.0460758582,
0.1009724438,
-0.0119130611,
0.005013728,
-0.0420260541,
-0.0411301404,
-0.0053108344,
-0.019232383,
0.0660027489,
0.0082882894,
-0.0229821093,
-0.0271972734,
-0.0873816535,
-0.085353829,
-0.0335500464,
-0.0114672519,
0.0385553353,
-0.0009915703,
-0.0205452666,
0.0479599088,
-0.0045391526,
-0.0138358856,
-0.0294980742,
0.0181363467,
0.0097885951,
-0.0341886319,
-0.0048649656,
-5.834779636e-33,
0.0176620781,
0.0091075879,
-0.04683882,
0.0323429331,
0.0112462109,
0.0066456064,
0.0006577258,
-0.0214567482,
-0.0045897849,
-0.0421848632,
-0.0309159607,
-0.016207017,
0.0051544327,
0.0147013441,
0.0036906991,
-0.0074406317,
0.0001542555,
0.0210389588,
-0.0494395234,
0.0401491933,
0.0595959947,
0.0190570019,
0.0232867897,
-0.026395591,
-0.0046589435,
-0.0164476149,
0.0015423682,
-0.0281791743,
-0.0718236566,
0.0591898821,
-0.0091589512,
-0.001033345,
0.0267193671,
-0.002528203,
-0.0227069892,
-0.0418761857,
-0.0189674031,
-0.0317291282,
0.0119101098,
-0.0023201988,
0.0990090743,
-0.0569817573,
0.0376097113,
-0.0282126758,
-0.0449178778,
-0.008767819,
0.0199223924,
-0.0094861267,
-0.082908988,
-0.0422313474,
-0.0135856308,
0.01334194,
-0.0372348987,
0.0418058224,
0.0293184947,
-0.0281586573,
0.0111147165,
0.036006663,
-0.0703670308,
0.0462854989,
0.0089846766,
0.0387735292,
0.0451148339,
-0.0167139228,
0.0342331715,
0.0758385658,
0.0291646309,
-0.0149602257,
0.0286703445,
0.0021293913,
0.0308533143,
0.0109479986,
0.0412024073,
0.0642356277,
0.0945730433,
-0.0354052149,
-0.0048385067,
0.0246735942,
-0.0258241612,
0.016779026,
0.0889491588,
0.0118053509,
-0.0516236238,
-0.0209970362,
-0.0396772623,
0.0285154693,
-0.0355265141,
-0.0041513159,
0.0285979006,
-0.0067545371,
-0.0272615142,
-0.0063356427,
0.0360843576,
-0.0423388965,
-0.0493883714,
-0.0259722862,
-0.0165927038,
-0.0050046179,
-0.0062642246,
0.0122690629,
-0.0343698896,
-0.0212313794,
0.0404160731,
-0.0453080013,
0.0160880126,
-0.0011046061,
0.0642352849,
0.0475478433,
-0.0834085494,
0.0125126848,
-0.0089431908,
0.0024446603,
0.0386309624,
-0.0050166976,
0.0069550443,
0.0024184885,
0.0142672705,
-0.0333036035,
0.046501372,
0.0196685418,
-0.0411286838,
0.0651951209,
-0.016264217,
0.0064062173,
0.0161040071,
0.0076882639,
-0.0274368003,
0.0415908061,
0.0644146875,
-0.0468322076,
-0.0023040092,
-0.0107937101,
2.732e-7,
0.0113967108,
0.0029405658,
-0.0062757097,
-0.0212890562,
0.0288495142,
0.0794710889,
-0.0693189278,
0.0691518635,
0.0009911492,
0.0569370724,
-0.0133559536,
-0.0061469544,
0.0312702432,
0.024388833,
-0.0589674935,
0.0089417305,
-0.0634715334,
-0.0181501862,
-0.0829365924,
0.0076900297,
0.0452989414,
0.0832954049,
0.0396273211,
-0.0354758836,
0.0128143858,
-0.1091701537,
0.0336157121,
-0.0176728535,
-0.016210543,
-0.0044355737,
0.0232094135,
0.020966202,
-0.0124487188,
0.0013926706,
-0.0171526615,
0.0313822925,
-0.0298713632,
0.0092961183,
0.069937773,
-0.0211265702,
0.0108442884,
-0.0629418492,
-0.0208774712,
-0.0875648186,
0.0506078564,
0.0196847357,
-0.0506088212,
-0.0258246623,
0.0121465735,
-0.0541772731,
0.0389715992,
-0.0004000227,
-0.0016735832,
0.0123844007,
-0.0167104322,
0.0315238014,
-0.0000994939,
-0.0953469127,
0.002564135,
-0.0787793323,
-0.0292531718,
-0.0198638104,
-0.0062180292,
-0.0006346488,
0.0210980419,
-0.0411155485,
0.0661810338,
3.001521739e-34,
0.0069419122,
-0.0417404659,
0.0074965796,
0.02873604,
0.0360391922,
-0.0147000104,
0.0110446149,
0.0501246303,
-0.0279963799,
-0.0318943001,
-0.0028988924
] |
http://arxiv.org/abs/2405.15485v1
|
http://arxiv.org/pdf/2405.15485v1
|
Learning Beyond Pattern Matching? Assaying Mathematical Understanding in LLMs
| 2024-05-24T12:04:54 | 2024-05-24T12:04:54 |
Siyuan Guo, Aniket Didolkar, Nan Rosemary Ke, Anirudh Goyal, Ferenc Huszár, Bernhard Schölkopf
|
cs.AI
|
cs.AI, cs.CL, cs.LG
|
We are beginning to see progress in language model assisted scientific
discovery. Motivated by the use of LLMs as a general scientific assistant, this
paper assesses the domain knowledge of LLMs through its understanding of
different mathematical skills required to solve problems. In particular, we
look at not just what the pre-trained model already knows, but how it learned
to learn from information during in-context learning or instruction-tuning
through exploiting the complex knowledge structure within mathematics.
Motivated by the Neural Tangent Kernel (NTK), we propose \textit{NTKEval} to
assess changes in LLM's probability distribution via training on different
kinds of math data. Our systematic analysis finds evidence of domain
understanding during in-context learning. By contrast, certain
instruction-tuning leads to similar performance changes irrespective of
training on different data, suggesting a lack of domain understanding across
different skills.
|
Learning Beyond Pattern Matching? Assaying Mathematical Understanding in LLMs: We are beginning to see progress in language model assisted scientific
discovery. Motivated by the use of LLMs as a general scientific assistant, this
paper assesses the domain knowledge of LLMs through its understanding of
different mathematical skills required to solve problems. In particular, we
look at not just what the pre-trained model already knows, but how it learned
to learn from information during in-context learning or instruction-tuning
through exploiting the complex knowledge structure within mathematics.
Motivated by the Neural Tangent Kernel (NTK), we propose \textit{NTKEval} to
assess changes in LLM's probability distribution via training on different
kinds of math data. Our systematic analysis finds evidence of domain
understanding during in-context learning. By contrast, certain
instruction-tuning leads to similar performance changes irrespective of
training on different data, suggesting a lack of domain understanding across
different skills.
|
[
-0.0423237197,
0.0221592356,
-0.018241832,
0.0638427958,
-0.0298374183,
0.011460159,
-0.0100025227,
0.014329127,
-0.0508592837,
-0.0254603326,
-0.0582819469,
-0.0011649085,
-0.0118833743,
-0.0026362916,
0.0550477095,
0.0080605671,
0.018656604,
-0.0417553261,
0.0119270412,
-0.0144035369,
-0.0543796346,
-0.0311234575,
-0.052419696,
-0.0012307863,
-0.0322669372,
-0.034981247,
0.0277305897,
-0.0622028634,
0.010782348,
-0.0426934287,
-0.0059717158,
-0.0018289845,
0.0454751961,
0.03406284,
0.0000019671,
-0.07428509,
0.0060624667,
0.0229207147,
-0.0335024558,
-0.0087308232,
-0.0207758695,
-0.0091374051,
-0.0019493229,
-0.0259319209,
-0.0718856081,
0.0295602996,
0.0199737251,
-0.0193337444,
-0.0119051319,
0.0515805483,
0.009684084,
-0.0296890028,
0.0522539653,
0.0063107428,
0.0891619623,
-0.0365070365,
0.0289607979,
0.0128926411,
0.0312051512,
0.0126667749,
0.0019717247,
0.0603739545,
0.0333203524,
-0.0094878869,
0.017356826,
0.0587832481,
0.0194272585,
0.0441937521,
0.0161176007,
0.0478270687,
-0.017459074,
-0.0205334909,
-0.0046098437,
0.0089082606,
-0.0025452096,
0.063808091,
-0.0838283151,
0.0144702671,
-0.0177427176,
-0.0034581809,
0.0229742248,
-0.0182722639,
0.0504090562,
0.0152110737,
0.0032180883,
0.1226619184,
0.0123326974,
-0.0527006425,
-0.0048088389,
0.0108786337,
-0.0351822898,
-0.0249997564,
0.0180268046,
0.0513649695,
0.0273300335,
0.0280523952,
-0.0025453891,
-0.0398970693,
-0.0156503096,
-0.0489471592,
0.058364898,
0.055504743,
-0.0175146125,
0.0046818908,
-0.060308855,
0.0183160771,
-0.0156571008,
0.0413464941,
-0.0105149755,
0.0304272473,
-0.0563693121,
-0.0240752157,
0.041945409,
0.0357607491,
0.0369090885,
0.0265632328,
-0.0209289584,
0.0662989616,
0.0065985029,
-0.0038171243,
-0.0135467118,
-0.0452928767,
-0.0079529649,
-0.0031826876,
-0.0881265998,
-0.0130014224,
-0.0037689791,
0.0465096831,
-0.0167754125,
-0.0149920881,
-0.0357728414,
0.0058434503,
-0.022347942,
-0.0127807204,
0.0250564199,
0.0837173909,
0.0138059724,
-0.0133862644,
-0.0522395484,
-0.01938577,
0.029153334,
-0.0457004905,
0.0423661023,
-0.0658857301,
0.0132581294,
0.0372436605,
-0.031398844,
-0.057379812,
0.0056431307,
-0.0085374545,
0.0045068874,
0.0681487918,
0.0113948379,
-0.0043820986,
0.0851428062,
-0.0070134057,
-0.0330202505,
0.0421084613,
0.0241533034,
-0.0485707559,
-0.0078409417,
-0.0263767689,
-0.0016150016,
0.0148428855,
0.0109941904,
-0.0105135161,
-0.0219609737,
0.0139857167,
-0.0006168731,
0.05196869,
0.0327263512,
-0.0040015141,
-0.0401660539,
-0.0171109457,
0.0202157013,
0.1058311909,
0.0201057494,
0.0033982638,
0.0363249145,
-0.0444659293,
0.0350688696,
0.029712921,
-0.0656649545,
0.0338093415,
-0.0450521819,
0.0136170695,
-0.0014575548,
-0.0229596607,
-0.0319761932,
-0.0812248588,
-0.0099668307,
0.0021163239,
-0.050399527,
-0.0276956148,
0.0837966353,
-0.0169956908,
0.0194987766,
0.0381396152,
0.001410648,
-0.0064060893,
0.0136459004,
-0.0463864766,
0.0069479384,
-0.0168107357,
0.0139169348,
-0.084199056,
-0.0502965897,
-0.0011345735,
0.015624078,
0.0006499537,
0.0269832648,
-0.0082250405,
-0.026132768,
-0.0132870758,
-0.0149490964,
-0.0501293391,
0.0359328128,
0.0043458245,
-0.0201614723,
0.0141141117,
0.0225656666,
-0.0734075382,
-0.0204713345,
0.0346024558,
-0.0719817653,
-0.0623603947,
0.0022869124,
-0.0174959674,
0.0133626685,
-0.0270932186,
0.0081593897,
0.0463848226,
0.037044622,
-0.0349118672,
0.0207917988,
0.0012729848,
-0.0236966945,
-0.0305789914,
0.0583908558,
-0.0578738898,
-0.0132397162,
-0.0232657865,
-0.0235392861,
-0.019952137,
0.0820192695,
-0.0204041712,
0.0092234593,
-0.0040479237,
0.0072486922,
-0.0104837287,
-0.0030130914,
-0.0242931265,
0.0647250861,
0.017740421,
0.0228404403,
0.0061310567,
0.0118930945,
0.01661608,
-0.0103003941,
0.0254525039,
-0.071507439,
0.0087985191,
0.0265039708,
-0.0246806052,
-0.0381706282,
-0.0088554006,
-0.0120630106,
0.0232349578,
0.0014600632,
-0.0014065669,
0.089073047,
0.0032076677,
-0.0288388915,
-0.0109624509,
0.0051605385,
0.0156079782,
-0.007899615,
-0.0295631103,
-0.0259095635,
0.0531761274,
0.0145419706,
-0.0156122921,
-0.0594150014,
-0.014853294,
-0.0637188256,
0.0274277534,
0.0222710222,
0.0410887003,
-0.0537417941,
0.0480241627,
-0.0039054195,
0.0126057928,
-0.0287289135,
0.0327889957,
0.0494239181,
-0.041455958,
0.0176282935,
0.0871884674,
-0.0239707697,
0.0226518847,
0.0429239944,
0.0081454376,
-0.0481522717,
0.0131502319,
-0.0312182624,
-0.0134210419,
-0.033005625,
-0.0141960382,
0.007724823,
0.0603995249,
-0.0298439804,
-0.0254404396,
0.0025321716,
0.0095135076,
0.0029876707,
0.0124168294,
-0.027970342,
0.0441125035,
-0.014725971,
0.0280215647,
0.025056798,
-0.074159652,
0.0697985739,
0.0013876024,
0.0147368638,
-0.0774876848,
-0.0331803225,
-0.0353288129,
0.0126562789,
-0.0009416941,
-0.0615368932,
0.0012316025,
0.0157393161,
-0.0415042527,
0.0292213783,
0.0163813401,
-0.0020934809,
0.053659454,
-0.0773236454,
0.0221360009,
0.0053468137,
0.0040743663,
-0.0279793572,
0.003946566,
-0.0063567497,
0.0373071395,
-0.0246441755,
-0.0386630408,
-0.0283953715,
-0.0215279553,
0.048992157,
0.0147798816,
-0.0150957974,
-0.0281778146,
-0.0401545651,
-0.0276876297,
0.0103895003,
0.0096768197,
0.0345494114,
-0.080272764,
-0.0142263211,
-0.0332949758,
0.0045639444,
0.0002514179,
-0.0372270048,
0.0194163192,
-0.0200089514,
-0.0085192798,
-0.0210609734,
0.0325544663,
0.0014413857,
-0.0178053752,
0.0067250072,
0.0725049227,
-0.0009143937,
-0.0466333665,
-0.0088799717,
-0.0568484552,
-0.0376206301,
0.0333803929,
0.0418068878,
-0.0523858927,
0.0169444568,
-0.0324408934,
0.0075957282,
-0.0019572587,
-0.0523149893,
-0.0101691904,
-0.0572614484,
-0.0002265446,
0.0182711612,
-0.0558976047,
-0.0101503162,
-0.0566832758,
0.0045785261,
0.0101056304,
-0.0052161058,
-0.0164907631,
-0.0233751759,
0.0226409826,
0.0566790104,
0.0015746402,
-0.0999743938,
-0.0113915028,
-0.043471694,
-0.0105060861,
0.0241192132,
0.0006915555,
-0.0572629198,
-0.0907114893,
0.0183633491,
0.0448476374,
0.0299117081,
-0.0046707806,
0.0084891003,
0.0192632601,
-0.0072003752,
0.0183029454,
0.0462882929,
-0.0762652755,
0.0031320949,
0.0108719971,
0.1320360005,
-0.0451890491,
0.018667832,
-0.0417109504,
-0.0269932002,
-0.0028659231,
0.0391874388,
0.0424530879,
0.0062825484,
-0.0194457471,
-0.0559854023,
0.0145651754,
0.0212533753,
-0.0212791748,
0.0094947554,
-0.0054643741,
0.0161806941,
-0.0242472794,
0.0352320969,
0.0203075055,
0.0248095747,
0.0111552924,
-0.0030241907,
0.0080207977,
-0.0329529084,
-0.0333244391,
0.0635321364,
0.0066123283,
0.0200527031,
-0.0583304986,
0.0007254417,
0.0338789374,
-0.0310731158,
-0.0317285061,
-0.0238127708,
-0.0807686076,
-0.0664537326,
-0.0425182804,
0.0476433262,
0.0205291081,
-0.0002759487,
-0.0045291674,
-0.0466478802,
0.0202829428,
0.0442178175,
-0.012258267,
0.0360147431,
0.011718682,
-0.0024987364,
0.0025365588,
-0.0204066057,
-0.0180395488,
-0.0072690663,
0.0073179938,
0.0026648066,
-0.007008811,
-0.0283481423,
-0.003035048,
0.030149471,
0.080776982,
0.0317483954,
-0.0014248351,
-0.0209487323,
0.0111154532,
0.0178039037,
0.0242858473,
0.0027892927,
-0.0171509683,
0.0212812535,
0.0361056663,
-0.0133983372,
-0.0230903309,
-0.0028091848,
0.0487697199,
-0.0379262529,
0.036026597,
0.0126917539,
0.0166843161,
0.0425523482,
0.0037943101,
0.0108470256,
-0.0126821343,
-0.0183070507,
0.0447100699,
0.013990107,
0.0076932465,
0.015870275,
0.0508131683,
-0.0486742258,
-0.0254027471,
-0.0687510744,
-0.0908523872,
-0.01016133,
0.0290459376,
0.0362705104,
-0.0251373071,
-0.0244937651,
-0.0812056884,
0.0170512646,
-0.0183456503,
0.0018239529,
0.0018162488,
0.0634857491,
0.1010593921,
0.0329587683,
0.0281790216,
-0.0376771167,
0.075872995,
-0.0071197832,
0.0657942295,
-0.0467276722,
0.0309291724,
-0.008496468,
0.0149602946,
0.0326603055,
0.0196196474,
0.0073920316,
-0.0007351346,
-0.0368627086,
-0.0506773926,
0.0332351699,
0.0067491545,
-0.0225083381,
-0.0149162794,
0.0410959572,
0.0769682825,
-0.0465464741,
-0.018499691,
0.0261576325,
0.0296402574,
0.0218957216,
-0.0541313067,
-0.0266924407,
-5.608813622e-33,
0.0039751241,
-0.0315401033,
0.0345684141,
0.1162170693,
0.0367748253,
0.0056187925,
-0.0307135601,
0.0094912695,
-0.0268494934,
0.0204910226,
0.0058370526,
0.0317614265,
0.0102432547,
-0.0252479445,
0.0398864821,
-0.0072779018,
0.0015933785,
-0.0361165553,
0.0170224234,
-0.0491621345,
0.0154999048,
0.0200371966,
0.0869151205,
-0.0989050046,
0.0246680602,
-0.0009727741,
0.0408968925,
0.0274980441,
-0.0266457722,
0.0475989543,
-0.0329673812,
0.0292605832,
-0.0120625291,
-0.0236803796,
-0.0534475222,
-0.0426637568,
-0.0138766337,
-0.0669185072,
0.0122785084,
0.0009243427,
-0.0269360542,
-0.0572807677,
0.0865225866,
0.0316215232,
-0.067415446,
-0.0077424399,
0.0531526208,
-0.039268598,
-0.013860167,
0.008074333,
-0.0412571616,
0.0223457441,
-0.0164583903,
-0.0177381858,
0.0580144487,
-0.0309388004,
-0.0118015176,
-0.0164334606,
-0.0235582571,
0.0454632789,
0.0048299157,
0.0049281288,
0.0487690233,
0.0140140494,
0.0235974472,
0.0373820364,
0.0385104492,
0.0127882985,
0.0236577671,
-0.0459259935,
0.0344031826,
0.0557792149,
0.0117357252,
0.0057668528,
0.0366630666,
0.0435897037,
-0.0806643143,
0.0593747869,
-0.0150426272,
0.0131904082,
-0.0080969706,
0.01446913,
-0.0147230383,
-0.0178523194,
-0.0208323039,
0.0334630907,
-0.0157320127,
-0.0065150163,
-0.0188669153,
-0.0535881296,
0.0395748056,
0.0064364122,
0.0138362721,
-0.0197516214,
-0.0155767631,
-0.0416998602,
0.0208659302,
0.0233895388,
0.0306019951,
-0.0150910821,
-0.09397313,
-0.0245972369,
-0.0328183658,
0.015722679,
0.0829685107,
0.0272242632,
-0.0019496294,
0.0179228261,
-0.0080288528,
-0.0283647906,
-0.0228582807,
-0.0652001947,
0.094212316,
0.0452096537,
0.020470215,
-0.0060769762,
0.0035403268,
0.0488807037,
0.0000669413,
-0.0432003923,
0.0044615758,
0.0404601768,
-0.033434432,
0.0440963507,
-0.0452999771,
-0.030764088,
0.0032126103,
0.014511711,
0.0239000693,
-0.0830569267,
-0.0013616752,
0.0014390226,
2.812e-7,
0.0690374523,
0.0517777428,
-0.0122418888,
0.0066080471,
0.0409298614,
0.0125482529,
-0.0647653863,
-0.023165565,
0.0190489236,
-0.0381338485,
0.0100340005,
0.0183341056,
0.0447689928,
0.0354327038,
-0.0398487262,
-0.0234540217,
0.0007787345,
0.0082115326,
-0.0236789975,
0.0044941995,
0.0056398232,
0.0283522848,
0.0286371466,
-0.004035804,
-0.0013397835,
-0.0584714226,
0.0761092976,
-0.026271509,
0.0403814726,
0.0195156652,
0.0138525153,
-0.0189257674,
0.035338819,
0.0607556887,
-0.0230194516,
0.0390914604,
0.0001657728,
0.0332214981,
0.017213963,
0.0367058665,
0.0154930996,
-0.0239075515,
0.0108110523,
-0.0505970418,
0.0046236054,
-0.0516245328,
-0.024026487,
-0.0129581764,
0.0350360274,
-0.0074357241,
0.040560279,
-0.0196026843,
0.0564483963,
-0.0156401377,
-0.0046339128,
0.001669421,
-0.015662726,
-0.0208738148,
-0.0222536586,
0.0612100065,
-0.0647060499,
-0.0584920831,
-0.0455415323,
-0.0261875,
0.053399533,
0.0109098461,
0.0047973008,
2.340430375e-34,
0.0092261778,
0.0061501083,
0.0200441368,
0.0682437569,
0.0098655829,
-0.0119982036,
-0.0305850357,
0.0423680991,
0.0332633331,
-0.0660290793,
-0.0117224306
] |
http://arxiv.org/abs/2405.15471v1
|
http://arxiv.org/pdf/2405.15471v1
|
Emergence of a High-Dimensional Abstraction Phase in Language Transformers
| 2024-05-24T11:49:07 | 2024-05-24T11:49:07 |
Emily Cheng, Diego Doimo, Corentin Kervadec, Iuri Macocco, Jade Yu, Alessandro Laio, Marco Baroni
|
cs.CL
|
cs.CL
|
A language model (LM) is a mapping from a linguistic context to an output
token. However, much remains to be known about this mapping, including how its
geometric properties relate to its function. We take a high-level geometric
approach to its analysis, observing, across five pre-trained transformer-based
LMs and three input datasets, a distinct phase characterized by high intrinsic
dimensionality. During this phase, representations (1) correspond to the first
full linguistic abstraction of the input; (2) are the first to viably transfer
to downstream tasks; (3) predict each other across different LMs. Moreover, we
find that an earlier onset of the phase strongly predicts better language
modelling performance. In short, our results suggest that a central
high-dimensionality phase underlies core linguistic processing in many common
LM architectures.
|
Emergence of a High-Dimensional Abstraction Phase in Language Transformers: A language model (LM) is a mapping from a linguistic context to an output
token. However, much remains to be known about this mapping, including how its
geometric properties relate to its function. We take a high-level geometric
approach to its analysis, observing, across five pre-trained transformer-based
LMs and three input datasets, a distinct phase characterized by high intrinsic
dimensionality. During this phase, representations (1) correspond to the first
full linguistic abstraction of the input; (2) are the first to viably transfer
to downstream tasks; (3) predict each other across different LMs. Moreover, we
find that an earlier onset of the phase strongly predicts better language
modelling performance. In short, our results suggest that a central
high-dimensionality phase underlies core linguistic processing in many common
LM architectures.
|
[
0.0056708334,
0.0181324519,
-0.0403753594,
0.0684458464,
-0.0085682813,
0.0337881558,
0.0199389886,
-0.0252394732,
-0.0524870269,
-0.0344469585,
-0.0502935089,
-0.0471634716,
0.0098273391,
0.0513242632,
0.0484832823,
-0.0647366196,
0.0091775535,
-0.0487230979,
-0.0403160118,
-0.0046354365,
-0.0499587618,
0.001037398,
0.0100268414,
0.0523625091,
-0.01223931,
-0.0021514466,
-0.0057011461,
-0.0421308614,
-0.0284846481,
-0.0022896137,
-0.0280526206,
0.0023199236,
0.0365608148,
0.0177695099,
0.0000019363,
-0.0469896272,
0.0220571794,
-0.010884596,
-0.0066522504,
-0.0193184689,
0.0375802778,
-0.0592027307,
-0.0219058525,
-0.0262129772,
-0.0228844304,
-0.0513916835,
0.0403049365,
0.0717236847,
0.0356196016,
0.0513075851,
-0.0316374674,
-0.0733791739,
0.051480338,
-0.0153597137,
-0.0417077392,
0.0086603342,
0.0106615406,
-0.0319389328,
-0.0382880829,
0.033714965,
-0.0572249815,
0.0115145585,
0.0406798273,
0.03502075,
-0.0133817699,
0.0635208264,
-0.0039288239,
-0.0185987875,
0.0026036156,
0.0404425785,
-0.0488944538,
-0.0035603505,
-0.0027328706,
0.0021066628,
0.0043002041,
-0.0214945953,
-0.0314589292,
0.0298919119,
0.0342157297,
0.0425673574,
0.0043288218,
-0.0126381256,
0.0229520481,
-0.0198287107,
-0.0687297583,
0.0643985495,
0.0250031482,
-0.0161080752,
-0.0123469112,
-0.0024995077,
0.0145279719,
-0.03553278,
-0.0037297097,
0.0421949849,
0.0226900354,
0.0152329402,
-0.0086954925,
-0.0150177432,
-0.012942615,
-0.0655930117,
0.0751046985,
0.0261345878,
-0.0188886877,
0.0052533168,
-0.0286265891,
0.086307399,
-0.0518856123,
0.0694092289,
-0.0495880768,
-0.0091710389,
0.0026182351,
-0.02880775,
-0.0290917866,
0.0244440734,
-0.0162127074,
-0.0214640014,
-0.0295761172,
0.0453119911,
0.0110164424,
-0.0379193947,
0.004892698,
0.0063014547,
-0.028705569,
-0.0037731968,
-0.0406809039,
0.0433195829,
-0.0027490132,
0.034181837,
-0.0113087771,
-0.0336522795,
-0.0117265396,
0.0111088911,
0.0194243696,
-0.0358358882,
-0.0021779959,
0.0885029659,
-0.0164944008,
-0.04834418,
-0.016830558,
0.011403854,
-0.0220410638,
-0.0455416255,
0.0529197827,
-0.0388432071,
0.023483444,
0.0352701098,
-0.016251957,
-0.0261472128,
-0.0170387775,
0.0065641538,
-0.0394373387,
0.0483408608,
0.0353783891,
0.0240632873,
0.0366326869,
-0.0058360742,
-0.0039902483,
0.0339284614,
0.0417357124,
0.0414262116,
-0.007456894,
-0.0392804481,
0.0394250676,
0.0754615813,
-0.010759132,
-0.0079795662,
-0.0499665663,
-0.0408501886,
-0.047348883,
0.0076900078,
0.0693584681,
0.051584173,
-0.0314800106,
-0.0064587523,
0.0454182811,
0.0980352759,
0.0319430716,
0.0593550839,
0.0518351234,
-0.0016498051,
-0.0235852543,
0.0221281648,
-0.0668779984,
0.0308665726,
0.0483203083,
-0.0054653925,
-0.0199327674,
0.0358954631,
-0.0444211066,
-0.0038686136,
-0.0165743567,
-0.0212135278,
-0.0130828666,
-0.0390229411,
0.0580315776,
0.0505308919,
-0.0222513732,
0.125748232,
0.0096708676,
-0.033815518,
0.0096314037,
-0.0503723174,
-0.0022344049,
-0.0033656911,
-0.021413343,
0.0127430661,
-0.0004262344,
0.0130731547,
0.0355538614,
0.0660345256,
0.0125521552,
0.0046830964,
0.0137366364,
-0.0039796033,
-0.0271255113,
0.000460132,
0.0175430588,
0.0041306573,
-0.0158999413,
-0.0274733305,
-0.0091003217,
-0.044822894,
-0.0240454897,
0.0469177589,
-0.0410018563,
-0.0096021183,
-0.0095651569,
0.0268681347,
0.0146453911,
-0.0206062961,
0.0033075528,
0.0133347763,
0.0098401336,
-0.0620473772,
-0.0345404707,
0.0352501273,
-0.0627353564,
0.0009859039,
0.0303973537,
-0.022320291,
0.0209244136,
0.0266207755,
-0.0161293373,
-0.0367759801,
0.066436559,
-0.0399831235,
0.0596860684,
-0.0122944908,
0.0158058815,
-0.032549534,
0.0327476412,
-0.018250661,
0.0572934002,
-0.0217308961,
0.0121503985,
-0.0113373064,
0.02982012,
-0.0134162093,
0.0392005555,
-0.0256638546,
0.0196837429,
-0.0221841373,
0.0116400262,
-0.0137760649,
-0.0295319483,
-0.0145599609,
-0.0599080622,
-0.0149469124,
-0.0451813899,
-0.0069311946,
0.0297251195,
0.0256116316,
-0.0203303341,
0.0125942025,
0.0039733434,
-0.013905867,
0.0170606542,
0.0155395614,
0.0256882142,
0.0397818722,
-0.0059525487,
-0.0285217296,
-0.0224618502,
-0.0306325722,
-0.0184418689,
0.0804676786,
-0.015867047,
0.0227112882,
-0.0604921393,
0.0114054661,
0.0091375839,
0.0166630652,
-0.0290416069,
-0.0473096371,
0.0207543392,
-0.0336964987,
0.0529964752,
0.0146683147,
0.01797661,
0.0609622262,
0.0519349277,
-0.0064659179,
-0.0246036164,
-0.0079525867,
-0.0568033159,
0.0039771707,
-0.0728908107,
-0.0094526494,
0.007113623,
0.1125338003,
-0.0023556773,
0.0000850169,
-0.0355881713,
0.0124221342,
-0.0243851859,
0.0086193802,
0.0228655767,
-0.023040574,
-0.058910545,
0.0471993499,
-0.0297960248,
-0.0361094251,
0.0277208034,
-0.0230046511,
-0.0187604018,
-0.0046376097,
0.0145406071,
-0.0380523615,
0.0775608271,
0.017352527,
0.0106149167,
-0.0178901423,
-0.0596892983,
-0.0481076352,
0.0408122875,
-0.038410753,
-0.0116319815,
0.0842953473,
-0.0231775735,
-0.0317367204,
-0.0177166574,
-0.0211968832,
-0.0306377541,
0.0097593367,
-0.0199027676,
0.0358893164,
0.0224857815,
-0.0287873577,
-0.0357618742,
0.0061877025,
0.0006070539,
0.0576280691,
-0.0136886099,
-0.0152452337,
-0.0204038583,
0.0266955644,
0.0287631676,
0.0728576332,
0.0306206085,
-0.0326441973,
0.0310487319,
0.026027577,
-0.0013869361,
0.0135173351,
-0.0439890847,
0.0245551635,
0.053461425,
0.0025399523,
-0.0180171337,
0.0188078713,
0.0055251722,
-0.0259463955,
-0.0529660918,
0.0923713893,
-0.000787059,
-0.0484116152,
0.0275445972,
-0.0096225133,
0.0114781549,
0.056643907,
0.0194333158,
-0.0457165465,
0.0171922054,
-0.0047642239,
-0.0209915955,
0.0316555314,
-0.0141529692,
-0.0033145414,
-0.0913823992,
0.0196783803,
-0.022492718,
-0.0919387192,
-0.0377573073,
-0.0257728137,
0.012128246,
0.0220706202,
0.0290043857,
-0.0026810062,
-0.0388482846,
-0.0340800472,
0.0126115382,
-0.0159343053,
-0.0294486322,
0.0280921031,
-0.0381821021,
0.00997438,
0.0171348378,
0.0318631157,
-0.0400811471,
-0.0325301997,
0.0549031943,
-0.0012524275,
0.0290969629,
0.0209421162,
0.0364157967,
0.0440806374,
-0.0235409122,
-0.008907211,
0.0043490184,
-0.028466776,
-0.0193327721,
-0.0528562814,
0.0906255171,
-0.0262562428,
-0.0548938997,
-0.0035400889,
-0.007210074,
-0.0380218737,
0.0556757525,
0.0646754354,
-0.0500611365,
-0.0442972519,
-0.0030985163,
0.0158062428,
-0.0049997102,
-0.0790918842,
0.0303975288,
0.0373187177,
-0.0387549028,
-0.034484826,
0.0008704134,
-0.0102615654,
0.0156978928,
0.0154011212,
-0.0544364117,
0.0302359927,
-0.0422356836,
-0.0159337875,
0.0417172462,
0.0071799802,
-0.0001380526,
-0.0397460908,
-0.001990024,
0.0221212897,
-0.0436180197,
-0.0350999311,
-0.0003383407,
-0.0549648851,
-0.0452776924,
0.0538604334,
0.0314171351,
0.0159475319,
0.0178931206,
0.0146195153,
-0.0520609431,
0.0687597096,
-0.0063486621,
-0.010735603,
0.048308976,
0.0182131771,
-0.0006301767,
-0.0062557422,
0.0453672707,
-0.0141285108,
0.0543835908,
0.0473771989,
-0.0028109581,
-0.0010714757,
-0.0039120889,
-0.0122012477,
-0.030927768,
0.0954989865,
0.0013818174,
0.0186058953,
-0.0413363092,
0.0009596738,
-0.028204903,
0.0049437969,
-0.0372037441,
-0.0441719443,
0.0283835549,
0.0366358459,
-0.010579953,
0.0264260359,
0.0144959558,
0.0442183539,
-0.0034221681,
0.0079175821,
-0.0032078305,
-0.0056186472,
0.0232897103,
0.0557124726,
-0.0082865478,
-0.0062561515,
-0.0243489053,
0.0060770717,
-0.021687964,
0.0248837043,
0.001588064,
0.0175820719,
-0.0579602383,
0.0216662325,
-0.0635677278,
-0.0478250608,
-0.0049569434,
-0.0510131754,
0.030294029,
-0.0412219018,
0.0631050095,
-0.026432259,
-0.0043122191,
0.003807124,
0.0126343044,
-0.0218516421,
-0.0472775549,
0.0549020655,
0.0018314171,
-0.0092590749,
-0.041084487,
0.0499437079,
-0.0378513709,
0.0194873922,
-0.016844064,
0.0360682197,
-0.0066728666,
-0.0190773644,
-0.0019435446,
-0.0130098211,
-0.0211786479,
-0.0057614059,
-0.0481184386,
-0.0716444775,
0.0190747865,
-0.0291085187,
-0.0014612975,
-0.0261210185,
0.050555855,
0.0510952584,
-0.0032271002,
-0.0243993234,
-0.0113978013,
-0.010664911,
0.0315893926,
0.002827378,
0.0075891041,
-6.208307292e-33,
-0.0135925794,
-0.0105101066,
0.0077473144,
0.0413541421,
-0.0285738483,
0.0183979049,
-0.0419417843,
0.0291296858,
-0.0118385665,
0.0135933505,
-0.0160830133,
0.0111642815,
0.0020055901,
0.0544179082,
0.022560928,
0.0077046514,
0.0424432978,
-0.0381800681,
-0.0049142027,
0.027174456,
0.0346222706,
0.0292119067,
0.0799608603,
-0.0381451771,
-0.0140917338,
0.0050370093,
0.0283913258,
0.0405879803,
0.0094912881,
0.0537450314,
-0.0506433025,
0.0513249338,
0.0462657511,
-0.0877953619,
-0.0412400514,
0.0305689909,
-0.0506844483,
-0.0404998139,
-0.0115563143,
0.0326580741,
-0.0363299996,
-0.0259529781,
0.0712351799,
-0.0093688667,
-0.0279167611,
-0.0572436564,
0.0390184969,
-0.0027447836,
-0.0126174539,
-0.0262829699,
-0.0785143152,
-0.006973072,
-0.0072197574,
0.0005913269,
0.078323625,
0.0236112028,
-0.0092688799,
0.0097907474,
-0.0454031602,
0.0234705824,
-0.0199019704,
0.0348328725,
0.0461381003,
0.0486967899,
-0.0096200593,
0.0736029595,
0.0705297366,
-0.011992787,
-0.0469517335,
-0.0220044032,
0.0717856884,
0.0553587079,
0.0009669519,
-0.034911342,
0.0020103331,
-0.0075482596,
-0.0223957263,
-0.0170748252,
-0.0348777622,
0.0516277961,
0.0219851267,
-0.0220290888,
0.0069211135,
-0.0341959447,
-0.0403232351,
-0.016843861,
-0.0415176526,
-0.0056395736,
0.0077258139,
-0.0028736026,
0.0259522647,
-0.0158783756,
0.0224214867,
-0.0479973033,
-0.0385318473,
-0.0000123555,
-0.0264817104,
0.0402414426,
0.0044670082,
0.0271903351,
-0.0245283544,
-0.0677749664,
-0.0151621075,
0.0121052396,
0.016475711,
0.0113387657,
0.027178729,
0.0443670638,
-0.0625752062,
-0.015959749,
-0.0390616618,
-0.0327940956,
0.0260968804,
0.0805932954,
0.0283654872,
0.0114529291,
-0.02179165,
0.0191357695,
-0.0017084809,
0.0326227471,
-0.0153659722,
0.0335081853,
0.0010657609,
0.0540050305,
-0.0336990654,
-0.0199226364,
-0.0703918487,
0.0228652563,
0.0866141096,
-0.0794502944,
-0.0367625579,
0.0095806746,
2.725e-7,
0.0712227523,
0.0518919714,
0.0301742665,
0.028714383,
0.0477536358,
0.0038940699,
-0.003998186,
0.0521245413,
0.0104413722,
0.0224159639,
-0.0274217818,
0.0056657866,
0.0314925835,
0.0244480129,
-0.0878910497,
0.0281691514,
-0.0643907487,
0.0425838605,
-0.0333243422,
0.0411210582,
0.0353175737,
0.0908555463,
-0.0187537093,
-0.0322607271,
-0.0449510999,
0.0131287929,
0.0404231586,
-0.0434441641,
0.0761365891,
0.0351640843,
-0.0289814714,
0.0131516326,
0.0184962582,
0.0216391794,
-0.0109870564,
0.0494947173,
-0.009634411,
0.0554443263,
0.0063663642,
-0.0198049024,
0.0272993594,
0.001215472,
-0.0167907346,
-0.0550995506,
0.0422148034,
-0.131739825,
-0.0440166928,
-0.0430069976,
0.0037374205,
0.0135226278,
0.0008507391,
-0.0121065779,
0.0031687082,
0.0243210681,
0.0414100066,
-0.0249471385,
0.0092646247,
-0.0868123844,
-0.0242533535,
0.0239552241,
-0.0736846849,
-0.025083201,
-0.0395224243,
-0.0131533509,
0.0621763021,
-0.0035530878,
-0.0072314329,
2.504956421e-34,
-0.0216222703,
0.0157334637,
0.0211655628,
0.0331359394,
-0.0101385424,
-0.0441916846,
0.0659198761,
0.0242726747,
-0.0177077372,
-0.0950859934,
-0.0243204422
] |
http://arxiv.org/abs/2405.15454v1
|
http://arxiv.org/pdf/2405.15454v1
|
Linearly Controlled Language Generation with Performative Guarantees
| 2024-05-24T11:30:44 | 2024-05-24T11:30:44 |
Emily Cheng, Marco Baroni, Carmen Amo Alonso
|
cs.CL
|
cs.CL, cs.SY, eess.SY
|
The increasing prevalence of Large Language Models (LMs) in critical
applications highlights the need for controlled language generation strategies
that are not only computationally efficient but that also enjoy performance
guarantees. To achieve this, we use a common model of concept semantics as
linearly represented in an LM's latent space. In particular, we take the view
that natural language generation traces a trajectory in this continuous
semantic space, realized by the language model's hidden activations. This view
permits a control-theoretic treatment of text generation in latent space, in
which we propose a lightweight, gradient-free intervention that dynamically
steers trajectories away from regions corresponding to undesired meanings.
Crucially, we show that this intervention, which we compute in closed form, is
guaranteed (in probability) to steer the output into the allowed region.
Finally, we demonstrate on a toxicity avoidance objective that the intervention
steers language away from undesired content while maintaining text quality.
|
Linearly Controlled Language Generation with Performative Guarantees: The increasing prevalence of Large Language Models (LMs) in critical
applications highlights the need for controlled language generation strategies
that are not only computationally efficient but that also enjoy performance
guarantees. To achieve this, we use a common model of concept semantics as
linearly represented in an LM's latent space. In particular, we take the view
that natural language generation traces a trajectory in this continuous
semantic space, realized by the language model's hidden activations. This view
permits a control-theoretic treatment of text generation in latent space, in
which we propose a lightweight, gradient-free intervention that dynamically
steers trajectories away from regions corresponding to undesired meanings.
Crucially, we show that this intervention, which we compute in closed form, is
guaranteed (in probability) to steer the output into the allowed region.
Finally, we demonstrate on a toxicity avoidance objective that the intervention
steers language away from undesired content while maintaining text quality.
|
[
0.0308135655,
0.079222396,
-0.0250221621,
0.0290273447,
-0.0370994434,
-0.0052845203,
-0.0017518409,
0.0082938205,
-0.0213555116,
-0.0173483063,
0.0264744032,
-0.0674363077,
-0.0180899277,
0.0751510113,
0.0627483502,
-0.0441677831,
0.0327365957,
-0.0276962351,
0.0053638215,
-0.0097721899,
-0.0108186519,
-0.0193423927,
0.0239918046,
0.0255112574,
-0.0195836611,
-0.0242843758,
0.0233681723,
0.0038772405,
-0.0150498375,
-0.0073246355,
0.0084991856,
0.0132362749,
0.0041336319,
0.020932911,
0.0000021561,
-0.0050842953,
-0.0070379926,
-0.02086417,
0.0617748871,
0.0135001028,
0.0652597472,
-0.0287041273,
-0.0319995098,
0.0124029247,
-0.0114871226,
0.019597495,
0.0604198836,
0.0098333443,
0.0548985749,
0.0653089508,
-0.0131517164,
0.0059812111,
-0.0029586216,
-0.0056217075,
0.0415136777,
0.0146819009,
0.0180913173,
-0.0154790543,
-0.0269597787,
-0.0161340833,
-0.0460700989,
0.0525240637,
0.0024776957,
-0.0397365279,
0.0253134724,
0.047962606,
0.0102200033,
-0.0596156456,
0.0027475115,
0.0482805148,
-0.0533406623,
-0.0358379632,
0.0375586152,
-0.0369362421,
-0.026671743,
-0.0291648488,
-0.015521829,
0.0679574236,
0.0137885353,
0.0087387888,
0.0134083321,
-0.0053987899,
0.0400827266,
0.0274853036,
-0.0102760829,
0.0442778245,
0.0367437564,
-0.0462545827,
-0.0289236195,
0.0068200896,
-0.0467205346,
-0.0515237749,
-0.0045079691,
0.0147173265,
-0.0340131558,
0.0339271948,
-0.0763604045,
-0.053792499,
0.0049883318,
-0.0709867477,
0.058302734,
0.0304994658,
0.039328631,
-0.0068283631,
-0.0191647261,
0.0421624668,
-0.0457916744,
0.004930492,
-0.0168775991,
0.0559855737,
0.0364678428,
-0.0271955878,
-0.0473812111,
0.0441717543,
-0.0137867033,
-0.0247271918,
-0.0407121405,
0.076282464,
-0.0358891264,
-0.007488674,
-0.0336349942,
0.0426531173,
-0.0568159074,
0.0232836194,
0.0064077508,
-0.0395582914,
0.0092112059,
-0.0064465282,
0.0184356198,
-0.0242231991,
-0.0220617931,
0.0379196443,
0.0051590372,
-0.0093577681,
-0.0387905911,
0.0749539584,
-0.0196052399,
0.0021530665,
-0.0297865402,
-0.0711186379,
-0.0333115123,
-0.027141422,
0.0094046127,
-0.0298245326,
-0.0149136521,
-0.0031012346,
-0.0174409095,
-0.0715092644,
0.0230852533,
0.0301357787,
-0.0434588455,
0.0486229695,
-0.0004184622,
-0.0029489032,
0.045641467,
0.0157651361,
0.0013787224,
0.0546497926,
0.0638451874,
0.0662787184,
0.0315550864,
0.0226895455,
-0.0080960384,
0.0367968418,
0.0576243959,
0.0168317165,
-0.0670410916,
-0.0419603176,
-0.0053666583,
0.053519208,
0.0450977236,
0.0825753883,
-0.0325049311,
0.0235784408,
0.0609316006,
0.0208280422,
0.0120005738,
0.0457141027,
0.1013640463,
-0.0001320569,
-0.0557614118,
0.0201455317,
-0.0209966563,
0.0065609845,
-0.0160478428,
0.0057507246,
0.0507614091,
0.0068607517,
-0.0010366257,
-0.009600413,
0.0242822338,
-0.011383418,
0.0324731879,
-0.0523689128,
0.0226104502,
0.0238955449,
-0.0698567331,
0.0320918113,
0.0300555341,
-0.0426321886,
0.0482141413,
-0.0490462147,
0.0429099277,
0.0595147572,
0.0764620826,
0.0105443364,
0.0058061155,
-0.0020380099,
-0.002026336,
0.0765579864,
0.0116460025,
-0.0037749102,
-0.0166764427,
-0.0332015604,
-0.0023652711,
-0.011899096,
0.010295962,
-0.0016163994,
-0.0232315529,
0.0191865806,
0.0184955765,
0.0132251568,
-0.0550734699,
0.0431698859,
-0.0112464381,
-0.0374200903,
0.0422967263,
-0.0094031831,
0.0480621681,
-0.0217058752,
-0.0403149948,
0.0405724235,
0.0077432017,
-0.0445728898,
-0.0532182641,
0.0041245921,
-0.0426721796,
-0.0168919004,
-0.0022657972,
-0.0252625979,
0.017707862,
0.0174475666,
-0.0007388502,
-0.0397779681,
0.0703172013,
-0.0727881566,
0.0169412643,
-0.030420417,
0.0092246998,
-0.0022846076,
-0.0056069437,
-0.0231825616,
0.0429347903,
-0.0149561642,
0.0157100316,
0.0368107259,
0.0493120141,
-0.0158099253,
0.0027424432,
-0.0166475605,
0.0143323839,
0.0161647685,
0.0082214475,
0.0565185323,
0.0443520024,
0.0128419483,
-0.0441797376,
-0.0005771023,
0.0026474898,
-0.0265708286,
-0.0179882478,
0.0420888513,
0.0123954564,
0.0172127429,
0.0351657234,
-0.0488006808,
0.0215495154,
0.0486823656,
0.0240071677,
0.0259390753,
0.0111978017,
-0.0397124663,
-0.0155550037,
0.0106651699,
0.0115724336,
0.0331874341,
0.0334598795,
-0.0173585583,
-0.0328702331,
0.0435355902,
-0.0329482518,
0.0163936727,
0.0024501325,
-0.0330295824,
0.0039556497,
-0.015946351,
0.0404616147,
0.0228106063,
0.0385664701,
0.0524210706,
0.0617651753,
0.0173531566,
-0.0258539617,
-0.0060224794,
-0.0662270039,
-0.0017334668,
-0.0024777299,
-0.0372533873,
0.0451758988,
0.0850959569,
0.0532683954,
-0.0121845491,
-0.0966000855,
0.0186355114,
-0.0376575552,
-0.0563669875,
-0.0099756615,
-0.0364772156,
-0.0480519421,
0.0159941968,
-0.0463648327,
-0.0702558383,
0.0663888156,
0.0117862783,
0.0273490902,
0.008303714,
0.0205819402,
0.0059490968,
0.0153569365,
0.0259743128,
0.0329069272,
0.0258907098,
-0.0083250254,
-0.0368113182,
0.0384091549,
0.0374644287,
-0.0262950454,
0.0208031274,
-0.0602302253,
0.0269737542,
0.0241291877,
-0.0318799652,
-0.0563501157,
-0.0249877255,
0.0645669699,
-0.0155035872,
0.0594669431,
-0.0050326637,
0.0046070809,
0.006910258,
0.0164073296,
0.0201173034,
-0.0276352596,
-0.0185376126,
-0.0101421699,
0.0103794811,
0.0701436624,
-0.0348893479,
0.0035177527,
-0.0450361334,
-0.0016761633,
0.0653070286,
-0.049630139,
0.0146003934,
0.0123922005,
0.0269673187,
0.0448732898,
0.0171386972,
-0.0419491716,
0.0279844794,
0.0015403128,
-0.0192002542,
-0.0770485774,
0.0857791007,
0.025058046,
-0.028884273,
0.0417437591,
0.0105587738,
0.0629031658,
-0.0077437432,
0.0183501635,
-0.0789768547,
0.0280616134,
-0.0195520613,
0.0371158719,
-0.0533424914,
-0.0235801227,
-0.035556443,
-0.0617378987,
0.0272276867,
0.0297112912,
-0.0348164178,
-0.0280843116,
-0.0116563942,
-0.0389053375,
0.0118073672,
-0.0126158213,
0.0223439131,
0.004933991,
0.0125072934,
0.0223899912,
-0.0022924631,
-0.0419015251,
-0.0675408915,
0.0061087203,
0.0401247367,
0.0586718284,
0.0181969721,
-0.000803136,
-0.0178407393,
0.0350055061,
-0.0240767412,
-0.0145495245,
-0.0267401617,
0.0136074359,
0.0044355094,
-0.015939882,
-0.0032532446,
0.0399863571,
-0.0051496453,
-0.0155777875,
-0.0258420836,
0.0522618219,
-0.0409740135,
-0.0473486036,
-0.035265483,
0.016233718,
0.0652288273,
0.0562975779,
0.0322460197,
-0.0754240379,
-0.0489149578,
-0.0414361991,
-0.0161263943,
-0.0442506894,
-0.0631759241,
-0.0597916655,
-0.0095430436,
-0.0541527011,
-0.1100158021,
-0.0049810358,
-0.00473222,
0.0310580358,
-0.0043226974,
-0.0316568427,
0.005128229,
-0.0438554063,
-0.0465671383,
-0.0183241367,
-0.0034672457,
0.0053445091,
-0.0467691608,
-0.0331306867,
0.0067049447,
-0.044287432,
-0.0202238858,
0.0329924226,
-0.0562182106,
-0.1058330238,
0.0190868322,
0.0495123975,
0.0212624595,
0.0286742952,
-0.0020244766,
-0.0613827072,
0.0964319929,
0.0403419174,
-0.0343771577,
0.0458532274,
0.004976721,
0.0104476642,
0.0005299978,
0.0508164577,
-0.0524551496,
0.0255475603,
0.0379349291,
0.0445788056,
-0.0065979511,
-0.0008152834,
0.0308328774,
0.0100059211,
0.0224039182,
-0.0782473385,
-0.0158798043,
-0.0514673814,
-0.0532014966,
-0.0411610417,
-0.0341166891,
-0.0583131537,
0.0217783265,
0.0039258008,
-0.0100529445,
-0.0406543612,
-0.0066740448,
-0.0232867859,
0.0337307416,
-0.0008372144,
-0.0366496108,
-0.0179932527,
0.0175431892,
0.0747806206,
0.0241482276,
-0.0122829899,
0.025506435,
-0.0560145825,
-0.0094645889,
-0.017404139,
0.0008297736,
0.0578318276,
0.0170814376,
-0.0091678547,
0.0048843464,
-0.0316809602,
-0.0018202791,
-0.0334619917,
0.0191851407,
0.0510698035,
0.0210561827,
0.0603395738,
-0.0693545341,
-0.0154109495,
-0.0169868246,
0.0252946485,
0.0312227663,
-0.0779523328,
0.0259668808,
-0.0004475149,
-0.0163905639,
-0.0294980425,
0.0917058066,
0.0032596234,
0.0306297783,
-0.034825813,
0.013016386,
-0.0278717112,
-0.0106645124,
0.0177891701,
0.0285253841,
0.0000501236,
0.0377422869,
-0.0583197884,
-0.0599021614,
0.0070571974,
0.0226762593,
0.0309098121,
-0.019478973,
0.0524688661,
0.0331647806,
-0.0026887699,
0.023789214,
-0.0552434511,
0.016306933,
-0.0168797262,
-0.0332850218,
-0.0613092668,
-6.636600128e-33,
-0.0353736952,
-0.0521288663,
0.0076219672,
0.0625743866,
-0.0413039215,
-0.0001475465,
-0.0474261008,
0.0216432884,
0.0490245074,
0.022362262,
0.0267936569,
-0.0084101092,
-0.0108040133,
0.046548333,
-0.0237788148,
0.0374003053,
0.045705162,
-0.0471093878,
-0.0459562317,
0.0692283511,
0.0427710675,
0.0185071528,
-0.0432673208,
-0.0399583466,
0.0204762537,
-0.0138604045,
-0.0389318652,
-0.0081930608,
0.0161043219,
0.0117434701,
-0.045506034,
0.0480280183,
0.0160805192,
-0.0204879716,
-0.0229306743,
0.0190082565,
0.0192725677,
-0.0337152071,
0.0014615942,
-0.0323857628,
-0.0584546626,
-0.0713147968,
0.0665452257,
-0.0123901619,
-0.0356948972,
0.0208465382,
0.0188139863,
-0.0115097789,
-0.0385994315,
-0.0811208114,
-0.0981647149,
0.0242960248,
-0.0060541788,
-0.0069078701,
0.0045776428,
-0.0054279454,
-0.0074166604,
-0.0035485232,
-0.0130512128,
0.0411528647,
-0.0077531915,
0.0320931561,
0.0392343849,
0.0422369167,
0.0292941798,
0.0391152054,
0.0082782796,
0.0238654017,
-0.0309891421,
0.0206174776,
0.0269581955,
-0.0160772018,
0.0062937634,
0.0102486387,
0.028696537,
-0.0635976344,
-0.0559410974,
-0.0038507283,
-0.0289353728,
-0.0110781332,
0.036436554,
0.0154652344,
-0.0143705094,
-0.0356507525,
-0.0430713296,
-0.0135677094,
-0.0171072166,
-0.0272300206,
0.0277877077,
0.0368869826,
-0.030360274,
-0.058938656,
-0.0132164946,
-0.0197145082,
0.0165132191,
-0.0014821913,
-0.0465431474,
0.0040653325,
-0.0063649658,
0.0105004245,
-0.0406170897,
-0.0249221046,
0.0010920549,
-0.0067528775,
0.0080090035,
0.0265432298,
0.0006906851,
0.0430098288,
-0.0524242558,
0.0127546927,
-0.0618099123,
0.0271069854,
0.0261925664,
0.0004000842,
0.0081841014,
-0.0119049456,
0.0069048693,
0.0221121013,
0.0167667009,
0.0026143079,
0.0126338759,
0.0120024625,
-0.0076608965,
0.0659369528,
-0.0859382376,
-0.0055628913,
-0.0199757647,
0.0775047466,
0.0678971335,
-0.0884101242,
-0.0303230081,
-0.0175535344,
2.936e-7,
0.0120742628,
0.0406867824,
-0.0026623548,
0.0673283413,
0.040936403,
0.0065842047,
0.0337146595,
0.0447960161,
-0.0114476979,
0.0041398117,
-0.002188164,
-0.0079115052,
0.0151144098,
0.0025175894,
-0.0497104377,
0.0615756176,
-0.0603085868,
0.0164164454,
-0.0105007356,
-0.0174525678,
0.0617468655,
0.046948947,
-0.0044842819,
-0.0223260503,
0.0133970436,
-0.0148869008,
0.0817901865,
0.0062231505,
0.0272127092,
0.0147439931,
0.0197890978,
-0.0010307141,
-0.014248942,
-0.0275644101,
-0.0371389985,
-0.0187543519,
-0.0182158444,
0.0767108724,
0.0063442788,
0.0004994811,
0.0271905158,
-0.0174764376,
-0.0009772906,
-0.02679337,
0.0518308207,
-0.0455073416,
-0.0219860915,
-0.0104142027,
0.0118124848,
0.000931329,
-0.0077255238,
-0.0454855487,
-0.0705126822,
-0.0006238983,
0.0203433614,
0.0359234698,
0.026382817,
-0.0517517664,
-0.0108855888,
-0.0216891523,
-0.0603351891,
-0.0359800421,
-0.0432490595,
0.0565091744,
0.0042941612,
-0.0338608846,
-0.0099100098,
3.115621604e-34,
0.0318771601,
-0.0515662096,
0.0006750738,
-0.0355831943,
0.0036751973,
0.0028539854,
0.025771305,
0.0140853683,
0.0456046946,
-0.0348398983,
-0.0492403284
] |
http://arxiv.org/abs/2405.15453v1
|
http://arxiv.org/pdf/2405.15453v1
|
Benchmarking Pre-trained Large Language Models' Potential Across Urdu NLP tasks
| 2024-05-24T11:30:37 | 2024-05-24T11:30:37 |
Munief Hassan Tahir, Sana Shams, Layba Fiaz, Farah Adeeba, Sarmad Hussain
|
cs.CL
|
cs.CL, cs.AI, I.2.7
|
Large Language Models (LLMs) pre-trained on multilingual data have
revolutionized natural language processing research, by transitioning from
languages and task specific model pipelines to a single model adapted on a
variety of tasks. However majority of existing multilingual NLP benchmarks for
LLMs provide evaluation data in only few languages with little linguistic
diversity. In addition these benchmarks lack quality assessment against the
respective state-of the art models. This study presents an in-depth examination
of prominent LLMs; GPT-3.5-turbo, Llama2-7B-Chat, Bloomz 7B1 and Bloomz 3B,
across 14 tasks using 15 Urdu datasets, in a zero-shot setting, and their
performance against state-of-the-art (SOTA) models, has been compared and
analysed. Our experiments show that SOTA models surpass all the encoder-decoder
pre-trained language models in all Urdu NLP tasks with zero-shot learning. Our
results further show that LLMs with fewer parameters, but more language
specific data in the base model perform better than larger computational
models, but low language data.
|
Benchmarking Pre-trained Large Language Models' Potential Across Urdu NLP tasks: Large Language Models (LLMs) pre-trained on multilingual data have
revolutionized natural language processing research, by transitioning from
languages and task specific model pipelines to a single model adapted on a
variety of tasks. However majority of existing multilingual NLP benchmarks for
LLMs provide evaluation data in only few languages with little linguistic
diversity. In addition these benchmarks lack quality assessment against the
respective state-of the art models. This study presents an in-depth examination
of prominent LLMs; GPT-3.5-turbo, Llama2-7B-Chat, Bloomz 7B1 and Bloomz 3B,
across 14 tasks using 15 Urdu datasets, in a zero-shot setting, and their
performance against state-of-the-art (SOTA) models, has been compared and
analysed. Our experiments show that SOTA models surpass all the encoder-decoder
pre-trained language models in all Urdu NLP tasks with zero-shot learning. Our
results further show that LLMs with fewer parameters, but more language
specific data in the base model perform better than larger computational
models, but low language data.
|
[
0.0066707735,
0.0539870597,
0.0115428092,
0.0413976833,
-0.0473270752,
0.0340310447,
0.0394652858,
-0.0023707482,
-0.0235145353,
-0.0258921981,
-0.056100361,
-0.0437401272,
0.0126589257,
0.0011261629,
0.0196999162,
-0.021398237,
0.0195552036,
-0.0239022225,
-0.0077673225,
0.0057354742,
0.0383009203,
0.0216879975,
-0.0317020677,
0.0504572317,
-0.0091420207,
-0.0191063043,
-0.0004876255,
-0.0060414686,
-0.0309927911,
-0.0359446742,
-0.0122847985,
0.0691493675,
-0.0009574425,
0.036575757,
0.0000019212,
-0.0539903156,
-0.0325074196,
-0.0161108654,
-0.0288118068,
0.0060876328,
0.0581516996,
-0.0428704433,
-0.0443371497,
0.0202820096,
0.0024373878,
-0.0217789076,
0.0946677774,
0.0567552559,
0.0236433502,
0.103543669,
-0.0075042499,
-0.0879879892,
0.0353109874,
-0.0169810727,
-0.0229067318,
-0.0501897037,
0.0137615697,
0.00215191,
0.0070007453,
-0.0110832704,
-0.0508742072,
-0.0018084975,
0.0078700194,
-0.0138028767,
-0.0283600409,
0.0131586939,
-0.0379484631,
-0.0523566902,
0.0211093109,
0.0686928853,
-0.1090793386,
-0.037352547,
-0.0187240373,
-0.0163842849,
-0.0547550097,
-0.0140931215,
-0.0411153957,
0.0390048996,
0.0010429899,
0.0480290614,
0.0474039093,
0.0234326422,
0.0131582208,
-0.0126827229,
-0.0222805757,
0.0444320254,
0.0228460822,
-0.0298175775,
0.0245297085,
0.0271493439,
-0.0071617104,
-0.0181894172,
0.0051055332,
0.0057497923,
-0.0308910906,
0.018203726,
-0.0060280254,
-0.0368115231,
0.022713948,
-0.0730870515,
0.0517676733,
0.092227079,
-0.0070178541,
0.0186448954,
-0.0446686968,
0.0316301659,
0.0125068501,
0.0140175587,
-0.0336766429,
-0.0130471447,
-0.0259370599,
-0.0186702255,
-0.0337998569,
0.0368102118,
-0.0646615252,
-0.0294367038,
-0.0518281385,
0.0130665731,
0.0118678557,
-0.001210768,
0.0463220328,
-0.0326528847,
-0.0110528739,
-0.0133208334,
0.0110087302,
0.0529114231,
-0.0291896984,
0.0319501236,
-0.0452296026,
-0.1394578069,
-0.017408954,
0.0228721518,
0.0098017175,
-0.0243852641,
-0.0039033196,
0.0929397121,
-0.0238204263,
-0.0824916139,
-0.0537029281,
0.0557372831,
0.0115307644,
-0.0753080845,
-0.0255899075,
-0.0586605147,
0.0780121759,
0.0153262196,
0.0326666422,
-0.0092199976,
0.0093673505,
0.0307633039,
0.0079439674,
0.0771679506,
0.0368165374,
0.0215136539,
0.0361452587,
0.013914736,
0.0526860133,
0.0981552452,
0.0009516404,
-0.0069000465,
0.0174119417,
-0.02348,
0.0104636969,
0.0556284674,
0.003399539,
-0.0047202408,
0.0075547248,
0.0001542107,
0.0024218336,
0.0371228755,
0.0180708598,
0.0054483898,
-0.0485534556,
0.0388751999,
0.0421760939,
0.0651800185,
0.0326980501,
0.0175174717,
0.0071584419,
0.0233240891,
0.0492123999,
0.001628824,
-0.0027065168,
0.0039165453,
0.0021834637,
-0.0017591642,
0.0395742282,
0.0201360676,
-0.0443970188,
-0.0013155284,
-0.0355547033,
-0.0425333567,
0.0201355834,
-0.0160858184,
0.0750222355,
-0.0029070452,
-0.0699444637,
0.0149017554,
0.0421239249,
-0.0112064481,
0.0219637733,
-0.0425268784,
-0.0172007494,
0.0466598421,
0.0143876821,
-0.0137564018,
-0.022161046,
0.008528349,
-0.036284484,
0.0391672775,
0.0062518371,
-0.0033359888,
0.021702785,
-0.0429718532,
0.0058503472,
-0.0240515172,
0.0060155643,
-0.0010571008,
0.0119626829,
0.0384992585,
-0.0149036758,
0.0319734849,
-0.0360144041,
0.0227746498,
-0.0638646781,
-0.0095407637,
0.003069524,
-0.0132882334,
0.0291594844,
0.0148707768,
-0.0371428058,
0.001964642,
0.0232028421,
-0.0207014214,
-0.0080745639,
0.0093836682,
-0.0403165631,
-0.0789569244,
0.0019481276,
-0.0073097623,
0.0442477316,
-0.0030926575,
-0.0442113094,
-0.051900804,
0.0391116627,
0.005440583,
0.0293278545,
-0.0571780354,
0.0585023537,
-0.0094482061,
0.0188459866,
-0.0301383622,
0.032937713,
0.0062802522,
0.01767621,
-0.0367443115,
0.0067847739,
0.052803427,
-0.0210124869,
-0.0432499014,
0.0898224786,
0.0193198677,
0.0183027051,
0.0329937972,
0.0299230739,
-0.017705543,
-0.0260551032,
-0.0833457783,
0.0127457166,
-0.0229886137,
0.0212671664,
-0.0023812861,
0.0101987645,
-0.0132285515,
-0.01122142,
-0.0184135586,
-0.0140598798,
0.0476800799,
0.0319720507,
0.0419770926,
-0.0277543571,
0.0135379378,
-0.031923756,
-0.0435529761,
0.0199544728,
0.0744438618,
-0.0200929306,
-0.024036143,
0.0185391698,
0.0108789392,
0.0117950635,
0.0181243885,
0.0041364776,
-0.0599214621,
0.008081127,
-0.0168161243,
0.0186363775,
0.0307560135,
-0.0267787334,
0.0104792509,
0.0194192231,
-0.0024517551,
-0.0166427474,
-0.011851633,
-0.0125108501,
-0.0435896628,
-0.0396544673,
0.0086753005,
-0.0123592913,
0.0448956862,
-0.0324289203,
-0.0079102591,
-0.0707409605,
-0.018061094,
-0.0399374664,
0.0027090043,
-0.0208942946,
0.0195471235,
-0.0007948389,
0.0471279137,
-0.0352677591,
-0.0209167898,
0.0062018568,
-0.0402431823,
-0.0120415632,
-0.0276594386,
0.0228528176,
-0.0646208301,
0.0070629069,
0.001192298,
0.0286737122,
-0.0416577607,
-0.0656997561,
-0.0094463639,
0.0206731502,
-0.0043483945,
-0.0431311391,
-0.0037366985,
-0.0358922184,
0.0335823596,
-0.0358482003,
-0.0254157521,
-0.0148789482,
-0.0083978735,
-0.0226002578,
0.0048832661,
0.0039772168,
-0.0052421107,
0.0315360241,
0.0393004119,
-0.0022810984,
0.0350057073,
0.0001375885,
0.0272500552,
-0.0119952643,
-0.0089127757,
0.0245477036,
-0.0239204429,
0.003295253,
-0.0109056244,
0.0249235407,
-0.0035708109,
-0.0017973363,
0.0144059816,
-0.0216460023,
0.0060239374,
0.0441091135,
-0.0478336327,
-0.0271486398,
-0.030583676,
-0.0001744746,
-0.02494107,
0.0041650031,
0.1024145782,
0.0274176393,
-0.0116851237,
0.0103681078,
0.0156485438,
-0.0200731084,
-0.0181308221,
-0.0133931525,
-0.1388737559,
-0.0285770185,
-0.0004147831,
-0.0472524576,
-0.0071804314,
-0.0131416572,
0.0001559144,
-0.1665243357,
-0.0453039408,
-0.0182140227,
0.0064567695,
0.002614812,
-0.0391983651,
-0.0298830885,
-0.0432627723,
0.0288210902,
-0.0175780505,
-0.0247192048,
-0.0099094603,
0.0459378883,
0.0249674451,
-0.09987358,
0.0021496865,
-0.0285462029,
0.0390607789,
0.0558687188,
-0.0021806732,
-0.0312881134,
0.0188477095,
0.0275134332,
-0.0235575624,
-0.0347455814,
0.0019582845,
0.0171752945,
-0.0112957144,
-0.0068311524,
-0.0026490886,
-0.0409856811,
-0.0246794429,
0.0027716144,
-0.0564208291,
0.0090475921,
-0.0048006773,
-0.0182005372,
0.0355360769,
-0.0227791555,
-0.0282481965,
0.0685202032,
-0.0072172703,
-0.046665851,
-0.0634313673,
0.0118568623,
-0.0294518638,
-0.0569698364,
-0.0618228503,
0.0415088199,
-0.0072519421,
-0.0493901633,
-0.0324804001,
-0.0378429145,
0.0275459588,
0.0004281356,
0.0104154469,
0.0386605039,
0.0316084214,
-0.0002930413,
-0.0041507185,
-0.012310843,
-0.0122155109,
0.0006442165,
-0.0831601694,
0.0257888008,
0.0459859483,
-0.0139322653,
0.0078501198,
-0.0186858512,
-0.0842043981,
-0.0342485458,
0.018074492,
0.0174191874,
0.0125578251,
-0.0158363096,
-0.0284187514,
-0.0719746873,
0.0617760643,
-0.0035150819,
-0.0273453686,
0.0041417084,
0.0285844356,
0.0115263872,
-0.0052961735,
0.0454004481,
0.0074495724,
0.041417215,
-0.0113012632,
0.0022971085,
0.0166729949,
0.0002365997,
0.0152789205,
0.0088188741,
0.0165822003,
-0.0451916307,
-0.0277869888,
-0.023268763,
0.0055950386,
-0.0414825417,
0.0109865684,
-0.0212762859,
-0.0396536812,
0.0695449188,
0.0415224135,
-0.0181171037,
0.0069495332,
-0.0054303426,
0.0498745888,
-0.0066893776,
0.0355245471,
0.0178411417,
-0.0125079071,
0.0172823071,
0.0203345232,
0.005009986,
-0.0113414582,
-0.0047642989,
0.0116843153,
-0.0342890136,
0.0131780859,
-0.0097913174,
0.0257497989,
0.0401637591,
0.0678895488,
0.0322308838,
-0.0272734985,
-0.0017653016,
0.0590659343,
0.1143459007,
-0.028434962,
0.0640042201,
0.0102643557,
0.0144918822,
-0.0385589264,
0.0269214846,
-0.0049951165,
0.0001879607,
0.0114854248,
-0.0249768719,
-0.0271244347,
-0.0061974567,
0.0795684755,
0.0347637907,
0.0257807337,
0.017788304,
0.0236352235,
-0.0007202761,
-0.0124120023,
-0.0222433768,
-0.0075148451,
-0.0153543511,
0.0268684253,
-0.0417550728,
-0.0459185131,
-0.0253132619,
-0.0580153428,
-0.0008029618,
0.0007316387,
0.027973529,
0.0289631207,
0.0190036912,
0.0027690765,
-0.0429507606,
-0.0077838884,
0.0067194225,
-0.0368061401,
0.0183842946,
-6.535706714e-33,
-0.0127183739,
0.0146749094,
0.0081013683,
-0.0344079137,
-0.0084247775,
0.006920625,
-0.0230561998,
-0.0095563717,
-0.0263019614,
-0.009583598,
-0.0371591374,
-0.0002367976,
0.0186393354,
0.0226104632,
0.0234749038,
-0.0147018125,
0.0540457033,
-0.0032195109,
0.0083140079,
0.0086213378,
0.0492099896,
0.0250477437,
0.0956976265,
-0.0046015149,
-0.0033274842,
-0.0340638906,
-0.0249658898,
0.0021740245,
0.0245332792,
0.0529389232,
-0.029850604,
0.0423255861,
0.0370371714,
-0.0615810119,
-0.0027066912,
-0.0141254282,
-0.0822251365,
-0.0028118275,
0.0359724425,
0.0646637827,
0.0095114782,
-0.032430727,
0.0072579132,
-0.0206629541,
-0.0656751618,
0.0499590002,
-0.0386925191,
0.0157776102,
0.0136071136,
0.0029850709,
-0.0883120894,
0.037414562,
-0.001157428,
0.0154654514,
0.0747760907,
0.0373824798,
-0.0153748654,
0.036994379,
-0.047129266,
-0.0095814215,
0.0269915015,
0.0514813215,
0.0531830452,
0.0191854741,
-0.0038841711,
0.0260865819,
0.0127666164,
-0.0348077267,
0.0197415799,
0.0165406428,
0.0286269039,
-0.0015996787,
-0.0143989986,
0.0437794887,
0.0704491958,
-0.051290147,
-0.089751713,
0.0136857927,
0.0152304145,
0.0441183932,
0.039942503,
0.021300735,
-0.0231132172,
-0.0441089496,
0.0024346071,
-0.022375783,
-0.0139815649,
-0.0308759566,
0.0333500803,
0.0171622224,
0.0209705774,
0.0618549585,
0.0530108102,
-0.0253773965,
0.0087417522,
0.0242318977,
-0.0654842257,
0.0202193633,
-0.0188792925,
0.0630159155,
-0.0998610333,
-0.0615172125,
0.0132739581,
0.0091004614,
0.0106285252,
-0.0038261195,
0.0367477983,
0.0234837066,
-0.0780639797,
0.0228164252,
-0.0369319469,
0.0283275787,
0.0204565767,
0.0765785649,
0.0324930772,
-0.0037602196,
-0.024255652,
0.0443661362,
0.0325347148,
0.0204027668,
0.003684479,
-0.015152026,
0.0140777407,
-0.0002647598,
0.0043984638,
0.0469363593,
-0.0563889071,
-0.0277222097,
0.019343907,
-0.0457881019,
-0.0114361364,
0.0073802648,
2.73e-7,
0.0394687615,
0.0391223356,
0.0342162661,
0.025389215,
0.0022886395,
0.0010935139,
-0.0300894324,
0.0396839194,
-0.0305088293,
0.0589633137,
-0.0269527435,
0.0213256571,
0.0464608744,
0.0069805253,
-0.0886923298,
-0.0262990817,
-0.0483622812,
0.0322479308,
-0.0807549953,
-0.0225236528,
0.0131814834,
0.1229091659,
-0.0297866389,
0.0200167727,
-0.0084301224,
-0.0344196633,
0.0730526745,
0.0123354979,
-0.0379595347,
0.0439836867,
0.006915174,
0.0268152356,
0.002228891,
0.0018400269,
-0.0255809911,
0.016429998,
0.0335999094,
0.0842546895,
-0.0128567591,
-0.0624450967,
0.0471839644,
-0.005319749,
-0.0012137013,
-0.0367611535,
0.0926247835,
0.0085987924,
-0.0644515455,
0.0250520743,
-0.0439681783,
-0.0034308017,
0.0701304823,
-0.0281665679,
0.0013743276,
0.0164954402,
0.050213404,
-0.0135217728,
-0.0205838028,
-0.0324792936,
-0.012859812,
-0.0049055186,
0.0229745638,
-0.0353165008,
0.0028413874,
-0.0446410105,
0.0102711804,
-0.0352619775,
0.0013450322,
2.265574114e-34,
0.0529458039,
0.016012948,
0.0393467173,
0.0246625915,
0.0308707412,
0.0037169387,
-0.0182225909,
0.030245902,
-0.0192114543,
-0.0645373687,
0.0012781027
] |
http://arxiv.org/abs/2405.15452v2
|
http://arxiv.org/pdf/2405.15452v2
|
Leveraging Logical Rules in Knowledge Editing: A Cherry on the Top
| 2024-05-24T11:30:00 | 2024-05-27T11:24:59 |
Keyuan Cheng, Muhammad Asif Ali, Shu Yang, Gang Lin, Yuxuan Zhai, Haoyang Fei, Ke Xu, Lu Yu, Lijie Hu, Di Wang
|
cs.CL
|
cs.CL, cs.AI, cs.LG
|
Multi-hop Question Answering (MQA) under knowledge editing (KE) is a key
challenge in Large Language Models (LLMs). While best-performing solutions in
this domain use a plan and solve paradigm to split a question into
sub-questions followed by response generation, we claim that this approach is
sub-optimal as it fails for hard to decompose questions, and it does not
explicitly cater to correlated knowledge updates resulting as a consequence of
knowledge edits. This has a detrimental impact on the overall consistency of
the updated knowledge. To address these issues, in this paper, we propose a
novel framework named RULE-KE, i.e., RULE based Knowledge Editing, which is a
cherry on the top for augmenting the performance of all existing MQA methods
under KE. Specifically, RULE-KE leverages rule discovery to discover a set of
logical rules. Then, it uses these discovered rules to update knowledge about
facts highly correlated with the edit. Experimental evaluation using existing
and newly curated datasets (i.e., RKE-EVAL) shows that RULE-KE helps augment
both performances of parameter-based and memory-based solutions up to 92% and
112.9%, respectively.
|
Leveraging Logical Rules in Knowledge Editing: A Cherry on the Top: Multi-hop Question Answering (MQA) under knowledge editing (KE) is a key
challenge in Large Language Models (LLMs). While best-performing solutions in
this domain use a plan and solve paradigm to split a question into
sub-questions followed by response generation, we claim that this approach is
sub-optimal as it fails for hard to decompose questions, and it does not
explicitly cater to correlated knowledge updates resulting as a consequence of
knowledge edits. This has a detrimental impact on the overall consistency of
the updated knowledge. To address these issues, in this paper, we propose a
novel framework named RULE-KE, i.e., RULE based Knowledge Editing, which is a
cherry on the top for augmenting the performance of all existing MQA methods
under KE. Specifically, RULE-KE leverages rule discovery to discover a set of
logical rules. Then, it uses these discovered rules to update knowledge about
facts highly correlated with the edit. Experimental evaluation using existing
and newly curated datasets (i.e., RKE-EVAL) shows that RULE-KE helps augment
both performances of parameter-based and memory-based solutions up to 92% and
112.9%, respectively.
|
[
0.0715360418,
0.0478204601,
-0.0131961042,
0.0274272915,
-0.0726553425,
-0.0066687032,
-0.0013598196,
-0.0323446207,
0.0152843557,
-0.0708479211,
-0.0122738136,
0.001598854,
-0.0481605083,
0.0173243806,
-0.0002263598,
-0.0006169747,
0.0647951141,
-0.0471004285,
0.000751523,
-0.0161529314,
0.0008371099,
0.0167694855,
-0.0062676193,
0.0144323045,
-0.0222753081,
-0.0497211739,
-0.0340077467,
-0.009592914,
-0.0088765342,
0.0253498368,
0.0270156451,
0.049605988,
-0.0066514867,
0.0245313942,
0.000002142,
-0.0781990141,
-0.0451512635,
0.0031759029,
-0.0300315935,
0.0371847153,
-0.0408782251,
-0.0039412919,
0.0032549249,
0.024804011,
-0.0340981111,
0.0018370919,
0.0290538203,
0.0468455702,
0.1018040851,
0.0608834997,
0.0086423652,
-0.0349422023,
0.0240583587,
0.0042378935,
-0.000135505,
-0.0460697785,
0.0821324289,
0.0520306639,
-0.0168613903,
-0.0110428957,
-0.0319156088,
0.0261398423,
0.0277915969,
-0.0004317651,
-0.0277470537,
0.0046654148,
-0.0405377373,
0.0145148449,
-0.0085508358,
0.0769541264,
-0.0033387872,
-0.0064065931,
0.0066864742,
0.0401772596,
0.020836005,
0.0404710807,
-0.0472078137,
0.0076444768,
-0.0037933786,
-0.0321369171,
0.0010003675,
-0.0538939349,
0.0208208822,
0.0248800796,
-0.0422258526,
0.1069914401,
0.0196341984,
-0.0385731906,
-0.0440048091,
-0.0253344923,
-0.0342341848,
-0.0677092299,
0.0081707463,
0.0690740496,
-0.024852844,
0.0339893438,
-0.0185812879,
0.0228573531,
0.0166651383,
-0.0435925052,
0.0102970116,
0.0516908281,
0.0314687416,
0.0703162625,
-0.0261480696,
0.0128999241,
-0.0109770652,
0.0093962029,
-0.0520803742,
-0.0119844349,
-0.071116358,
-0.038983807,
-0.0588681027,
0.0727901533,
-0.049078282,
-0.0041424707,
-0.0165179558,
0.0495163314,
-0.0022116087,
-0.0000220849,
0.004065285,
-0.0041190223,
0.0100204926,
-0.0221144389,
-0.0417034291,
0.0601012483,
-0.0326650329,
0.0230340511,
-0.0036939164,
-0.0245880242,
-0.0254249796,
-0.016016515,
0.0187757071,
-0.0094357766,
0.0072610606,
0.0616272166,
-0.0266909655,
-0.0039818827,
-0.0237310082,
-0.0816077814,
-0.0100331334,
-0.0492191203,
-0.0110445274,
-0.0686754063,
0.0488655344,
0.0035027983,
-0.0361537561,
-0.0399584882,
0.0097589353,
0.0176608097,
-0.0425754376,
0.0450522006,
0.0249024983,
0.0272576287,
0.0384858064,
0.0319196843,
-0.0360285267,
0.0682787225,
0.0211280901,
-0.0193656627,
0.0486303493,
0.0037904026,
-0.0020344607,
0.0369997323,
0.0477753915,
0.005793212,
0.0238801539,
0.0458020903,
0.0468384661,
0.0217338726,
0.0033492108,
0.0327038653,
-0.0227067191,
-0.0189532787,
0.0549703278,
-0.0112864794,
0.0168177262,
0.0209479909,
0.0180033222,
-0.0083178533,
0.0101886829,
0.0095083136,
-0.0310218036,
0.0324614681,
-0.0475572906,
0.0182411782,
0.0065596616,
0.0531160347,
-0.0219806712,
-0.0402347818,
-0.0030006259,
0.0299153179,
-0.0133259045,
0.0473330468,
0.0457456633,
0.0146084642,
-0.0237303488,
0.0625148714,
0.0183217712,
0.0197187643,
0.0202311631,
-0.0552150868,
0.0478031859,
0.070734337,
0.0716265738,
-0.0344883949,
-0.043722041,
-0.0135756824,
0.0342778265,
-0.0071992548,
0.0252876431,
-0.0383380689,
-0.003029488,
-0.0213721618,
-0.0027843895,
-0.0443069786,
0.0105955238,
-0.0112823425,
0.0662292838,
-0.0115344645,
-0.0067554424,
-0.0099652447,
0.0188458599,
-0.0052245683,
-0.0112051116,
0.0100714518,
0.027817864,
-0.0151012605,
0.0029372049,
-0.0039416864,
-0.0165484827,
0.0204076879,
0.0159864724,
-0.0275643431,
-0.0409074612,
-0.0130159799,
0.0051970915,
-0.0636823103,
0.0658316165,
0.0068179532,
-0.0196756199,
0.0196030512,
-0.0370608568,
-0.0240079071,
0.0084824534,
-0.0305785313,
0.0551590174,
-0.064922139,
-0.0096687665,
0.0365462638,
-0.0268192794,
-0.0413984619,
0.0443676338,
-0.0193386059,
-0.0266718399,
0.010509463,
-0.0215549413,
-0.0062962449,
-0.008599923,
0.0164506473,
0.0107000098,
0.0008880844,
0.0121171828,
-0.0121415434,
-0.0011394616,
-0.0519774929,
-0.0257588252,
-0.0473705158,
0.0315146111,
-0.0285711605,
0.0252887271,
0.0182172898,
0.043782793,
-0.0065403841,
-0.0154899843,
-0.0375125818,
-0.0102916574,
0.018322004,
0.0327955894,
0.0198779646,
0.0332252011,
-0.010079124,
-0.015188897,
-0.0027108069,
-0.0178873446,
0.0307889599,
-0.0330825895,
0.0201954562,
-0.0422183499,
0.0118385535,
-0.072620146,
0.0148126641,
0.0129124681,
-0.0133576496,
-0.0407132208,
-0.0124070402,
0.0691432133,
0.0269755274,
-0.0324714743,
-0.0129583953,
0.0676491335,
-0.0595400482,
-0.0256579164,
0.0493380763,
-0.0330290236,
0.0037825196,
-0.0144876139,
-0.0188638065,
0.027930757,
0.054833468,
0.004499835,
-0.0619014874,
-0.0396638811,
0.0574335642,
0.0110580316,
-0.0081381639,
-0.0749873668,
0.0059588044,
0.0206959005,
0.0223372765,
-0.0330150723,
-0.0087391306,
0.1092189029,
0.0241290107,
0.0197880976,
-0.0842287242,
0.0140672252,
-0.0493313633,
0.0506922603,
-0.0164093021,
0.0389560983,
-0.0434271656,
0.0166837387,
-0.0478813387,
-0.009433277,
0.0164112374,
-0.0468791947,
0.0423287004,
-0.0981284529,
0.0184369795,
-0.0029850558,
-0.038812045,
-0.0322976708,
-0.0274447352,
-0.016000485,
0.0128066503,
0.0000362333,
-0.0058503449,
0.0020020646,
0.026470542,
0.0429274999,
0.0218071528,
-0.0119291293,
-0.0048931572,
-0.0089644538,
0.0444671586,
0.0975269526,
0.0342013575,
0.0061403266,
0.013632034,
0.0612769909,
0.0719753206,
0.014386951,
-0.0126117039,
0.0514078103,
-0.0023440267,
-0.0070316712,
-0.0444716513,
0.0131586939,
-0.0031999138,
0.0503030568,
-0.0612054169,
-0.0155653907,
0.0841395408,
0.0735177994,
0.0146333333,
-0.0086496435,
-0.0480918065,
0.0119150681,
0.0031372188,
0.0763111115,
-0.0845702589,
-0.0413793065,
0.0024274348,
-0.0025968987,
-0.0179167688,
-0.0430648699,
-0.0095564974,
-0.08766523,
0.0074881944,
-0.0190895852,
-0.0687639192,
-0.0399689302,
-0.04302302,
0.0023427475,
0.0156136127,
-0.0074251457,
0.0401620194,
-0.004003711,
0.0177361351,
-0.0188102834,
-0.010768665,
-0.010167825,
0.0458935015,
-0.0191238914,
0.0282613263,
0.0771750435,
-0.0014363027,
0.0160158779,
-0.0183796864,
-0.0187151544,
-0.0092030074,
-0.0015600323,
-0.0411289111,
-0.0137455482,
-0.0016959844,
-0.0055743745,
-0.0295226779,
0.0447170772,
-0.0187481493,
-0.0287197269,
0.0694658831,
0.1074624211,
-0.019238608,
-0.0049180281,
-0.0122495713,
0.0045162947,
-0.0360049047,
0.0965005383,
0.0138158407,
-0.0333350636,
-0.0131266695,
-0.048200991,
-0.0284529123,
-0.0563445054,
-0.0092556998,
0.0068573952,
-0.0109930644,
0.0551268123,
-0.0820124224,
-0.0277951229,
0.083431989,
0.0168600399,
-0.0001491403,
0.0068717226,
0.036824137,
0.0185742341,
-0.0525427349,
-0.0137086883,
-0.0570552275,
0.0041842172,
-0.014116752,
0.0002867845,
0.0199712571,
-0.0340713076,
-0.0199401714,
0.0196784623,
-0.0655378923,
-0.0472199507,
-0.0013030259,
0.061058633,
0.0195268933,
0.0546182767,
-0.0099251391,
-0.0240650456,
0.0027276028,
-0.0063163913,
-0.0323387571,
0.021021314,
-0.0024565461,
0.0046911994,
-0.007741441,
0.0357846171,
-0.067620784,
-0.0087798983,
0.0090543674,
0.0900252089,
-0.0568819046,
-0.0237881802,
0.0431012772,
0.0001370775,
-0.0319720469,
0.0429648459,
-0.0249973293,
-0.0021190755,
0.063194409,
-0.0355305113,
-0.0315843932,
-0.017110562,
0.0101946285,
0.0016388965,
-0.0406408422,
-0.0019192178,
0.000813671,
0.0096261594,
0.0222868919,
-0.00925007,
-0.041658897,
0.006983351,
-0.0249995664,
0.0627929121,
-0.0262387674,
-0.017896248,
0.0159916766,
0.0229221862,
0.0264133681,
-0.0077617546,
0.0558393374,
0.0465807952,
-0.0315594524,
-0.0522736683,
-0.0071829031,
-0.063360028,
-0.0536235496,
0.0122692166,
0.0317450911,
0.0262282155,
0.0379059538,
0.0037596466,
-0.0246958416,
-0.0586897396,
-0.0199322309,
0.0064021354,
0.0279252883,
0.0049191485,
0.0490005761,
0.0256118495,
-0.0144564044,
0.0100286342,
0.0743329674,
-0.0054717874,
0.0931710973,
-0.0277230907,
-0.0365864411,
-0.0296773501,
-0.0560184903,
-0.0332593769,
0.0294017326,
-0.0237396639,
0.0055321995,
-0.0835934132,
-0.0368088372,
-0.0280975923,
0.0130159175,
-0.0171263367,
-0.0222949106,
0.0099351471,
0.0493844338,
0.0350405723,
-0.063315168,
-0.0398277827,
0.0154087199,
-0.0323255062,
-0.0394364893,
0.028987417,
-6.35297979e-33,
-0.006860821,
-0.0146548226,
0.0640664324,
-0.0041180761,
-0.0178830121,
-0.0083125886,
-0.0329365954,
-0.0300441384,
-0.0188671984,
-0.0072935214,
0.0004776771,
0.0071273916,
-0.0183179863,
0.0124983452,
0.0097371722,
-0.0249161851,
0.0432415456,
0.005931647,
-0.0158234276,
0.0131984167,
0.021024568,
0.0196577664,
0.0230068639,
0.0590736791,
0.0015051789,
0.0173266158,
0.0006311789,
0.0763723403,
-0.075781554,
0.0568396077,
-0.0231702458,
0.0037233545,
0.0220334549,
-0.0421753116,
0.0051481831,
-0.0166373048,
-0.081071347,
-0.0128847696,
0.0442580469,
-0.0010609938,
-0.0278336331,
-0.0654705539,
0.035816256,
0.0294130947,
0.0349380299,
-0.0134748518,
-0.0065835007,
-0.0157054123,
-0.0094306711,
-0.0354745314,
-0.0345237479,
0.0054631429,
-0.0248496942,
-0.0566182546,
-0.0466488786,
0.0285139866,
0.0281611793,
-0.0041442136,
-0.0807701573,
0.0189307667,
0.0110962791,
0.0365614891,
0.0550020672,
-0.0119532133,
0.0072389147,
0.0128082158,
-0.0086984951,
-0.0018321658,
-0.0235088747,
0.0258193761,
-0.0052604331,
0.0251264982,
0.0185749847,
0.0326119587,
0.0092931399,
0.0122841951,
-0.1022561491,
0.0226774625,
-0.0010585644,
0.0271683689,
0.0139080239,
0.007674762,
-0.0171515234,
-0.0104488069,
-0.0364970975,
0.0748660862,
-0.0429114364,
-0.0220719948,
0.0359361023,
0.0208549704,
-0.0235300306,
-0.0145990886,
0.0325496048,
-0.0495468602,
-0.0035399217,
0.0186834522,
-0.0156723,
0.0470195189,
-0.0158607587,
0.0484514795,
-0.0370149389,
0.0422246605,
0.0356384367,
-0.0235195477,
-0.0173303783,
0.019816393,
0.0174729973,
0.0035126996,
-0.0497587286,
0.0085910102,
-0.0419033244,
0.0423387215,
-0.0237919539,
-0.0408133417,
-0.0025860583,
0.0383303501,
0.0072400942,
0.0124121355,
0.0477845222,
-0.0066768117,
-0.0145345973,
0.0091259433,
0.0212333947,
0.0005517577,
-0.0086915046,
0.0290326234,
-0.0024628006,
-0.009710581,
-0.0513219088,
-0.0497154519,
-0.0082735056,
0.0553131662,
2.866e-7,
0.0528253876,
0.0508483797,
0.0103965411,
0.0385233648,
0.0271582454,
0.0059374338,
-0.0023294671,
-0.0026028974,
0.0327160135,
0.0126013421,
-0.0048796902,
0.0571742952,
0.0176390801,
0.013838972,
-0.0862653181,
-0.0080143912,
-0.0115964105,
-0.0493129417,
-0.0783239976,
-0.0243098252,
0.0460875705,
0.0452809669,
0.0119370455,
-0.004992154,
-0.0206394438,
-0.0022311478,
0.0550498255,
0.0044576181,
-0.0550977886,
0.0256451778,
0.0654910579,
-0.0284357667,
0.0280337129,
0.0323565044,
-0.0199022777,
0.0233337358,
-0.0325177386,
0.0798382759,
0.0098603191,
-0.0071433815,
-0.03289745,
-0.0323131196,
0.0180460382,
-0.053964287,
0.1083661318,
-0.0419737026,
-0.0089703957,
-0.0199476965,
0.0060196919,
-0.0272327513,
-0.0262459107,
-0.0351231769,
-0.0128338858,
0.0265438259,
0.0068097254,
0.0100050857,
0.0343620218,
-0.0341694318,
-0.0099793728,
0.0248185918,
-0.0219528824,
-0.0584050752,
-0.0216166172,
-0.0543317832,
0.0225798152,
-0.0784623101,
-0.0034666737,
2.800301902e-34,
0.0355685018,
0.0002189136,
0.0179822929,
0.0749461427,
0.043034263,
-0.0220561866,
-0.0278586894,
-0.0205228962,
0.0046322527,
-0.1231754869,
-0.024232123
] |
http://arxiv.org/abs/2405.15374v1
|
http://arxiv.org/pdf/2405.15374v1
|
Leveraging Large Language Models for Semantic Query Processing in a Scholarly Knowledge Graph
| 2024-05-24T09:19:45 | 2024-05-24T09:19:45 |
Runsong Jia, Bowen Zhang, Sergio J. Rodríguez Méndez, Pouya G. Omran
|
cs.IR
|
cs.IR, cs.AI, cs.CL, H.3.3; I.2.4; I.7.5; I.2.7
|
The proposed research aims to develop an innovative semantic query processing
system that enables users to obtain comprehensive information about research
works produced by Computer Science (CS) researchers at the Australian National
University (ANU). The system integrates Large Language Models (LLMs) with the
ANU Scholarly Knowledge Graph (ASKG), a structured repository of all
research-related artifacts produced at ANU in the CS field. Each artifact and
its parts are represented as textual nodes stored in a Knowledge Graph (KG).
To address the limitations of traditional scholarly KG construction and
utilization methods, which often fail to capture fine-grained details, we
propose a novel framework that integrates the Deep Document Model (DDM) for
comprehensive document representation and the KG-enhanced Query Processing
(KGQP) for optimized complex query handling. DDM enables a fine-grained
representation of the hierarchical structure and semantic relationships within
academic papers, while KGQP leverages the KG structure to improve query
accuracy and efficiency with LLMs.
By combining the ASKG with LLMs, our approach enhances knowledge utilization
and natural language understanding capabilities. The proposed system employs an
automatic LLM-SPARQL fusion to retrieve relevant facts and textual nodes from
the ASKG. Initial experiments demonstrate that our framework is superior to
baseline methods in terms of accuracy retrieval and query efficiency.
We showcase the practical application of our framework in academic research
scenarios, highlighting its potential to revolutionize scholarly knowledge
management and discovery. This work empowers researchers to acquire and utilize
knowledge from documents more effectively and provides a foundation for
developing precise and reliable interactions with LLMs.
|
Leveraging Large Language Models for Semantic Query Processing in a Scholarly Knowledge Graph: The proposed research aims to develop an innovative semantic query processing
system that enables users to obtain comprehensive information about research
works produced by Computer Science (CS) researchers at the Australian National
University (ANU). The system integrates Large Language Models (LLMs) with the
ANU Scholarly Knowledge Graph (ASKG), a structured repository of all
research-related artifacts produced at ANU in the CS field. Each artifact and
its parts are represented as textual nodes stored in a Knowledge Graph (KG).
To address the limitations of traditional scholarly KG construction and
utilization methods, which often fail to capture fine-grained details, we
propose a novel framework that integrates the Deep Document Model (DDM) for
comprehensive document representation and the KG-enhanced Query Processing
(KGQP) for optimized complex query handling. DDM enables a fine-grained
representation of the hierarchical structure and semantic relationships within
academic papers, while KGQP leverages the KG structure to improve query
accuracy and efficiency with LLMs.
By combining the ASKG with LLMs, our approach enhances knowledge utilization
and natural language understanding capabilities. The proposed system employs an
automatic LLM-SPARQL fusion to retrieve relevant facts and textual nodes from
the ASKG. Initial experiments demonstrate that our framework is superior to
baseline methods in terms of accuracy retrieval and query efficiency.
We showcase the practical application of our framework in academic research
scenarios, highlighting its potential to revolutionize scholarly knowledge
management and discovery. This work empowers researchers to acquire and utilize
knowledge from documents more effectively and provides a foundation for
developing precise and reliable interactions with LLMs.
|
[
0.067084603,
0.0391055718,
0.010148122,
0.0256229751,
-0.038710054,
0.0096623255,
0.0041065933,
0.0077765668,
0.0166152492,
-0.0609683059,
-0.0143459244,
0.0219582673,
-0.0219723564,
0.0430082269,
-0.0279601365,
0.0020915426,
0.0236213915,
-0.0107469773,
0.0549341217,
0.0029660668,
-0.003873233,
0.0537528172,
0.0003136749,
-0.0050804131,
-0.0037552065,
-0.0202661939,
-0.0303228423,
0.0106049962,
-0.0330851339,
-0.0220950525,
0.0084171398,
0.0340536088,
0.0343861841,
0.0686515644,
0.0000019051,
-0.0513644703,
-0.0549613722,
0.0490602739,
-0.0095866956,
0.0162546057,
0.0936396942,
-0.0074785841,
0.0220436472,
-0.0300242882,
-0.0145835215,
-0.014266789,
-0.0056783571,
0.0461908989,
-0.0157002993,
0.0354600362,
-0.0101574725,
-0.0706033781,
0.0798553824,
-0.0019005751,
0.0202956684,
-0.0552248545,
0.0418186672,
0.0178685579,
0.0686944351,
0.0746372044,
0.0172068719,
0.0123117724,
-0.0330932215,
-0.0714436024,
0.0504824892,
0.0092636766,
-0.0012904813,
0.0041658264,
0.044741787,
0.0292948894,
0.0408386141,
0.0049700318,
0.0693277642,
0.0164454505,
-0.032568451,
0.0856714919,
-0.0742557794,
-0.0065187411,
-0.0164031666,
-0.0265590213,
-0.0024146449,
-0.0177589282,
0.0385852382,
0.053917028,
-0.0250973534,
0.0738237277,
0.0158171933,
-0.0205589104,
0.0084208371,
-0.0223205984,
0.030018948,
-0.0367726646,
0.0289663747,
0.0392811149,
0.0091082659,
0.0223110672,
0.0129926968,
-0.024881389,
0.0646869913,
-0.032863792,
-0.0412498042,
-0.0044623436,
-0.001328251,
0.0890421867,
-0.0089621963,
-0.0137545299,
-0.0502859056,
-0.0022178867,
0.0092433542,
-0.0404686593,
-0.079023622,
-0.0330795683,
-0.0108011179,
0.045797281,
-0.0468153916,
-0.0389195792,
-0.0136199258,
0.0288474914,
0.0237937756,
-0.0020634874,
0.0140680298,
0.0245473105,
-0.0396046825,
-0.0078447955,
-0.0641545653,
0.0470021814,
0.0204669721,
0.0082447827,
-0.0128482636,
0.0071168295,
0.0508696996,
-0.0068628001,
0.0373757556,
-0.0249094795,
0.0261215866,
-0.0041830749,
0.0322959572,
-0.020837022,
-0.0275585912,
-0.0627098605,
0.0273687113,
-0.0313687921,
0.0312892497,
-0.0740857124,
0.0620361865,
0.0058445157,
-0.018611813,
-0.0269383639,
0.0063774958,
0.0267865416,
-0.0449881926,
0.0298848301,
-0.0449388325,
-0.0131595591,
0.0641512871,
0.0157115702,
0.0114022056,
0.022469135,
0.0011282159,
-0.00132481,
0.0648163036,
0.0073053883,
-0.0465065166,
-0.0377333127,
0.0049392143,
-0.0058357082,
-0.0033541296,
0.0384369232,
-0.034017764,
0.0412703156,
0.0101334285,
0.0341003053,
-0.0305644684,
-0.0231282655,
-0.0012112653,
0.0645793378,
0.0029157682,
0.026596643,
-0.0020200883,
0.0294216108,
0.013486959,
-0.0446339585,
-0.0043210387,
0.0310339108,
0.0278047025,
-0.002217225,
0.0141122444,
0.0244781189,
-0.0076328274,
-0.00972973,
-0.018816907,
0.0290213414,
-0.0162476897,
0.0127076861,
0.0757535025,
-0.0408324897,
-0.0001158005,
0.0128383655,
0.0284398049,
-0.0159880761,
0.0547369346,
0.0096877683,
0.0154422885,
0.0726290569,
0.0709880367,
-0.078757368,
-0.0083571244,
0.0183322039,
-0.0126140043,
0.0568436235,
0.0148398438,
-0.015467491,
-0.0297039598,
-0.02284199,
-0.0505688787,
-0.0216014683,
-0.0134726241,
-0.0016971705,
-0.0046303286,
0.0269523803,
0.0258538406,
-0.0686549619,
-0.0342242904,
0.0254505128,
-0.0342099518,
0.0168180168,
-0.0072618611,
-0.0222399291,
-0.0049417084,
0.0350437537,
0.0017086599,
0.035370484,
0.0673446581,
-0.0293471888,
-0.0443627723,
-0.0567477271,
0.0334857479,
0.0125503484,
0.03650143,
-0.0300266147,
0.0314111337,
0.0038470791,
-0.0127662113,
-0.0373913683,
-0.0092507685,
0.0183061305,
0.039413292,
-0.0121924989,
-0.0253407937,
-0.0146920932,
-0.0451427177,
-0.0392697491,
0.038678959,
0.0104902899,
0.0274183173,
0.0229732487,
-0.0459701642,
0.0069765402,
0.013910423,
-0.0374104306,
0.0081447875,
0.0137447529,
0.0197094679,
0.0613604076,
-0.0490541048,
-0.0454237685,
-0.0045414898,
-0.0583152175,
0.0399623811,
0.0112489723,
-0.0036045886,
-0.0101465089,
0.0028640239,
0.0088724783,
0.0023513352,
-0.0215727799,
-0.0030336166,
-0.0064182035,
-0.0010421438,
0.0315613896,
0.0449986011,
0.0051169135,
-0.0392020829,
-0.0580048747,
-0.0245000143,
0.0345973335,
0.0371270254,
-0.0050261109,
0.0270546079,
0.0283278264,
-0.042045977,
0.015120185,
-0.0276340768,
0.0398031548,
-0.0078688692,
-0.0638173074,
-0.0362524167,
0.06771788,
-0.0117565021,
0.0255579781,
0.0259871054,
-0.0286703892,
-0.0612952299,
0.0090053305,
-0.0414514355,
0.0321424566,
-0.0192317441,
0.0312722996,
0.0297818519,
-0.0467878133,
0.0008576338,
-0.0117603019,
0.0009909277,
0.0034625684,
0.0007732271,
0.0199083965,
-0.0249912571,
-0.0076151453,
0.0403022058,
0.0552934334,
0.0067606256,
-0.0005684856,
0.0555877984,
-0.0298783313,
0.0231558401,
-0.0370665751,
0.001636614,
-0.0571776703,
0.0558323339,
-0.0022813855,
0.0001628925,
-0.043880228,
-0.0050031771,
-0.0060182274,
0.0269030947,
0.0203922261,
-0.0422836691,
-0.0075971759,
0.0256156474,
0.0438430756,
-0.0226282682,
0.0317142084,
0.0014508154,
-0.0112074614,
-0.0289200507,
0.0570446253,
0.0271409098,
-0.0570060723,
-0.0231209546,
0.0129959779,
0.0338502564,
0.0193316415,
-0.0401342474,
-0.0275127161,
-0.0280519985,
0.0315045416,
0.0763095468,
-0.073711209,
-0.0032286646,
0.0446012318,
0.0093344646,
0.0488746911,
-0.0302735735,
-0.0059897122,
0.003803035,
-0.0474432707,
0.0071213222,
-0.0780599266,
-0.0079533784,
-0.031015845,
-0.0563681349,
-0.079941906,
-0.0027072416,
0.0920876041,
0.0324335732,
-0.0754157975,
-0.0504708253,
-0.0596830323,
0.0092334589,
0.0234811176,
0.0417969972,
-0.0960563868,
0.0148968501,
-0.0093152421,
0.0174452532,
0.0093127983,
-0.0523973107,
0.0472587571,
-0.0695853233,
0.0325342715,
-0.0347431153,
-0.0057718856,
0.0107216286,
-0.0659620836,
0.0070241475,
-0.0014498596,
-0.030641716,
0.0264514238,
-0.0288213938,
-0.0099376105,
0.0234699808,
0.0211759266,
-0.0301023107,
0.0093208803,
-0.039104078,
-0.0276528746,
-0.0083250096,
0.0802628696,
-0.0204929467,
-0.0304848626,
-0.0207534302,
0.0029997963,
-0.0170886889,
-0.0204719771,
0.0175046157,
-0.0213191882,
0.0080398498,
-0.0312936343,
-0.0116721401,
-0.0436949283,
0.0246736277,
-0.0090124672,
0.106676124,
-0.0101168659,
0.0398352891,
-0.0869059861,
-0.0408401601,
-0.0312421508,
0.0713721663,
-0.0154961152,
0.0164640266,
-0.0231383555,
0.0134804565,
0.0180412307,
0.0061615468,
-0.0153017789,
0.0341013074,
-0.042876903,
0.0094048064,
-0.029440755,
-0.033473663,
0.0218150057,
-0.0242925081,
-0.006247309,
0.042896878,
0.0178863853,
-0.0300785266,
-0.0766576976,
0.024708394,
-0.0206220876,
0.0240554027,
-0.0098246662,
0.0217197984,
0.0819875672,
-0.0095439386,
-0.0396311879,
-0.0479306541,
0.0090051237,
-0.0615392216,
-0.0335366689,
-0.0220086072,
0.0084769549,
0.0339117944,
-0.0246328153,
-0.062211819,
0.0250223558,
0.0337267071,
-0.0624778494,
-0.04706835,
-0.0013383598,
-0.0330776609,
-0.0440388322,
0.0142316818,
-0.0819936171,
-0.0005036987,
-0.0480401106,
0.0257902816,
-0.0462776162,
-0.0916747525,
0.0301725212,
0.0336594544,
0.0280302297,
0.0111544123,
0.0286885034,
-0.0177393109,
0.0197491739,
-0.0234861877,
-0.0146712419,
-0.014101469,
0.0127473474,
-0.0029020701,
-0.026108183,
0.0216954239,
0.0126609933,
0.0043481169,
0.0324362107,
-0.0220819172,
0.0226724334,
-0.0038457895,
0.0142385531,
0.0374204591,
-0.0062044263,
-0.0259719491,
0.0024639042,
-0.0448849946,
0.0136541938,
-0.0439422764,
-0.0044747028,
0.0315374546,
-0.0130053842,
-0.0095349029,
-0.0306289326,
-0.0957000107,
-0.0767341852,
0.0331761539,
-0.0292138755,
0.0619432926,
-0.0009287153,
0.0095380787,
-0.037246041,
-0.0398271345,
-0.0245464277,
0.01657575,
0.0232820548,
0.0173737835,
0.0152168162,
0.0445320681,
-0.0037804926,
0.0015036874,
0.0426270217,
0.0457256138,
0.0059492257,
0.0222454425,
-0.0655116513,
0.01552229,
-0.0130095473,
0.023903491,
0.0594000332,
-0.0293418225,
-0.0041229329,
-0.0164895188,
-0.0086939884,
0.0441901945,
-0.0034684476,
-0.0434917994,
-0.0499975272,
0.0378962606,
0.050787963,
0.0387832113,
0.0018132282,
-0.041465912,
0.0681344345,
-0.0813442618,
-0.0316111296,
0.016454462,
-6.432398393e-33,
-0.0067727705,
-0.0434364304,
0.0000975423,
0.0154298944,
-0.0270082727,
0.0208536256,
-0.0296436455,
-0.054627154,
-0.0304156803,
0.048951637,
-0.0436671935,
0.0156983472,
-0.0137969647,
0.0037702653,
0.0347240865,
0.0139434841,
0.0507216528,
0.0166617148,
0.038738776,
0.0365605503,
0.0806776285,
0.0450342596,
0.0291040018,
-0.046870701,
0.030301461,
-0.0265984833,
0.0538649,
0.0289903078,
0.0271316767,
0.0900362879,
-0.0144913979,
0.0220812634,
0.0114308205,
-0.0278570093,
0.0221514199,
-0.0088413423,
-0.0982853696,
0.0061097126,
0.0076586436,
0.0010092799,
-0.00154834,
-0.0084392847,
0.0599254556,
0.0088637061,
0.0319286659,
-0.033206135,
0.0594403818,
0.0035009005,
-0.0629027933,
-0.1244770959,
-0.0335443243,
0.0135639319,
-0.0067417203,
0.0463594347,
0.0179850794,
-0.0545632355,
0.0202631745,
0.0753177181,
-0.0442613177,
-0.0026791862,
-0.0452648103,
0.0292677693,
-0.0081746085,
-0.0123898014,
0.0234074015,
0.0281125866,
-0.0236447081,
0.0297712758,
0.0342822932,
0.0162925087,
0.035945043,
0.0589735806,
0.0336964726,
0.0360581651,
0.0339902379,
0.0011841959,
-0.0870729163,
0.0129169533,
0.0058950959,
0.0423157625,
0.0105302222,
0.0371517017,
0.0009506666,
-0.0152289132,
-0.011786852,
-0.0115564764,
-0.0336497836,
-0.0027184482,
-0.01207639,
-0.0166885033,
-0.0066738436,
0.0246839523,
0.0301286057,
-0.0481543429,
-0.0462962352,
0.0908676013,
0.0155668017,
0.0183161628,
-0.0348463431,
0.0143866455,
-0.0629586428,
-0.0214094482,
-0.0263303891,
-0.0510936156,
-0.0350318514,
0.0186567772,
0.0132484883,
-0.0228719916,
0.0005373522,
-0.0218145195,
0.0110604726,
-0.0296977591,
-0.0296290163,
0.001957776,
-0.0064862515,
0.0511937849,
0.0346511267,
0.065103218,
0.0543322749,
0.0554460026,
-0.023449406,
0.0028077811,
0.0346968174,
-0.0049116216,
0.0130196959,
-0.0333782546,
-0.0245346576,
0.0003589335,
0.0326074436,
-0.0533149503,
0.0071195294,
0.0001909488,
2.864e-7,
0.0642098039,
0.0376833268,
-0.0073251389,
0.0200631786,
0.0051044114,
0.0004648897,
0.0201777965,
0.0246107467,
0.0142565276,
-0.0206607059,
-0.000806498,
0.0055009956,
-0.0141052073,
0.0529884808,
-0.1068496853,
-0.0488114879,
-0.0473271795,
-0.0914139822,
-0.0730250552,
0.0109888716,
0.0083049834,
0.010411432,
-0.000809681,
-0.0175244175,
-0.0367278084,
-0.0512878224,
0.0429560877,
0.0140779922,
0.0081815682,
0.0354171284,
0.0591503568,
0.0158505552,
0.0024269505,
-0.0059523517,
0.0101122577,
0.0573434457,
-0.0101966653,
0.0332167372,
-0.0451171435,
0.0227443222,
0.0343284272,
-0.0653577372,
0.0260586496,
-0.0294516459,
0.0445725545,
-0.0177888703,
0.0046423008,
-0.0072865696,
-0.0390100479,
-0.0430648923,
-0.0054278122,
-0.0287274849,
-0.0101320995,
0.0138926208,
-0.0084545445,
-0.0136867538,
0.0136714764,
-0.0675652698,
0.0318061672,
0.0506671518,
-0.0149938641,
-0.0097323079,
-0.0086866245,
-0.0296744853,
0.0101877255,
0.0002225113,
0.0254046172,
2.631743818e-34,
0.0728438869,
0.0410005189,
-0.0364666283,
0.0430600978,
0.0131784659,
-0.0038169634,
-0.0410294533,
0.0041812928,
0.0029318125,
-0.1093045846,
-0.0099074114
] |
http://arxiv.org/abs/2405.15370v1
|
http://arxiv.org/pdf/2405.15370v1
|
Large Language Models can Deliver Accurate and Interpretable Time Series Anomaly Detection
| 2024-05-24T09:07:02 | 2024-05-24T09:07:02 |
Jun Liu, Chaoyun Zhang, Jiaxu Qian, Minghua Ma, Si Qin, Chetan Bansal, Qingwei Lin, Saravan Rajmohan, Dongmei Zhang
|
cs.CL
|
cs.CL
|
Time series anomaly detection (TSAD) plays a crucial role in various
industries by identifying atypical patterns that deviate from standard trends,
thereby maintaining system integrity and enabling prompt response measures.
Traditional TSAD models, which often rely on deep learning, require extensive
training data and operate as black boxes, lacking interpretability for detected
anomalies. To address these challenges, we propose LLMAD, a novel TSAD method
that employs Large Language Models (LLMs) to deliver accurate and interpretable
TSAD results. LLMAD innovatively applies LLMs for in-context anomaly detection
by retrieving both positive and negative similar time series segments,
significantly enhancing LLMs' effectiveness. Furthermore, LLMAD employs the
Anomaly Detection Chain-of-Thought (AnoCoT) approach to mimic expert logic for
its decision-making process. This method further enhances its performance and
enables LLMAD to provide explanations for their detections through versatile
perspectives, which are particularly important for user decision-making.
Experiments on three datasets indicate that our LLMAD achieves detection
performance comparable to state-of-the-art deep learning methods while offering
remarkable interpretability for detections. To the best of our knowledge, this
is the first work that directly employs LLMs for TSAD.
|
Large Language Models can Deliver Accurate and Interpretable Time Series Anomaly Detection: Time series anomaly detection (TSAD) plays a crucial role in various
industries by identifying atypical patterns that deviate from standard trends,
thereby maintaining system integrity and enabling prompt response measures.
Traditional TSAD models, which often rely on deep learning, require extensive
training data and operate as black boxes, lacking interpretability for detected
anomalies. To address these challenges, we propose LLMAD, a novel TSAD method
that employs Large Language Models (LLMs) to deliver accurate and interpretable
TSAD results. LLMAD innovatively applies LLMs for in-context anomaly detection
by retrieving both positive and negative similar time series segments,
significantly enhancing LLMs' effectiveness. Furthermore, LLMAD employs the
Anomaly Detection Chain-of-Thought (AnoCoT) approach to mimic expert logic for
its decision-making process. This method further enhances its performance and
enables LLMAD to provide explanations for their detections through versatile
perspectives, which are particularly important for user decision-making.
Experiments on three datasets indicate that our LLMAD achieves detection
performance comparable to state-of-the-art deep learning methods while offering
remarkable interpretability for detections. To the best of our knowledge, this
is the first work that directly employs LLMs for TSAD.
|
[
-0.0170470364,
0.0611804649,
-0.0345533304,
0.069993265,
-0.0520959496,
0.008366554,
0.0466540195,
-0.0011103007,
-0.0069518681,
0.0253447667,
0.0407741815,
-0.0270746965,
-0.0306209717,
0.101980783,
-0.0092732152,
0.0122012813,
0.040645998,
-0.0443258993,
0.063317962,
0.0103612402,
0.0061134091,
-0.0299704242,
0.0124880029,
0.0188428499,
-0.0022524623,
-0.0107599059,
-0.0199535843,
-0.0177120008,
0.0666110516,
-0.0426273867,
0.0035631014,
-0.0269688927,
0.035750933,
0.095344387,
0.0000022465,
-0.0121847847,
-0.0436996222,
-0.0309806224,
-0.0170079097,
0.0198802408,
-0.0221523587,
-0.0422126539,
0.0369404405,
0.0214098543,
-0.0626577064,
0.0020891489,
0.0211141333,
0.0668575317,
0.0537891164,
0.0587092675,
-0.0133402776,
0.0299408808,
-0.0021043532,
-0.0449997596,
-0.0361773372,
-0.0289522745,
0.0196001399,
-0.0399967879,
0.007016927,
0.0029368338,
-0.0276208762,
-0.0138352988,
0.0365939587,
-0.0243329089,
-0.0042578867,
0.0291256923,
-0.0171181764,
0.0183315128,
0.0090313172,
0.0469161756,
-0.03948972,
-0.0146570662,
0.0346451476,
-0.018486008,
-0.0029935159,
0.0049522771,
-0.0319469459,
-0.0011960525,
0.021047309,
0.0460077003,
0.0398913026,
0.0516444407,
0.0337336287,
0.0444692634,
-0.0634541884,
0.0622688867,
-0.0271703266,
-0.035995014,
-0.0824602246,
-0.0168834012,
-0.0078973304,
-0.0102056973,
0.0133636603,
0.0237973817,
-0.0529269092,
-0.0284234565,
0.0216187127,
-0.0171021223,
0.0013662652,
-0.0199586879,
0.0140454154,
0.0064984863,
-0.0411535986,
0.0400899425,
-0.012948595,
0.0389048271,
-0.0380915031,
-0.0380401313,
-0.0074941828,
-0.0283853132,
-0.0483995117,
-0.0435933992,
0.0357306302,
-0.0183043182,
-0.0260110777,
0.0009161194,
-0.010849257,
0.0174999144,
-0.0405206904,
-0.0036934193,
-0.0300970189,
0.0198216066,
0.0157566685,
-0.0289839227,
-0.0032764273,
-0.0086365184,
-0.0303605385,
0.0499177463,
-0.0151551925,
-0.0213288143,
-0.0381803028,
0.0257975291,
0.0018519898,
0.0227126349,
-0.0028573975,
0.019503152,
0.0557007603,
0.0063539161,
-0.0973082632,
-0.0554549284,
-0.0100400085,
0.0140356459,
-0.0477475524,
-0.0049581942,
-0.0405964255,
-0.0309565105,
0.0327668376,
0.022466993,
-0.0031522182,
-0.0385727398,
-0.0180261694,
0.0193872526,
-0.0675140992,
0.0149318418,
-0.0204152241,
-0.0023543304,
0.0322340243,
0.0215648394,
0.0481664501,
0.0009824978,
0.0363141112,
-0.0257921144,
0.0520830117,
-0.0160246287,
0.02361398,
-0.0404101945,
-0.0488463826,
0.0073183016,
-0.0629841164,
-0.027618004,
0.0332363658,
-0.0307510179,
0.0068101403,
-0.0341739617,
0.0365332328,
0.0021287533,
-0.036864318,
0.0034082257,
0.0304763652,
0.0051330561,
-0.0721404925,
-0.0372557603,
-0.0034686245,
0.0273577794,
-0.051165238,
-0.01246709,
-0.0349320136,
-0.010360606,
-0.0337870978,
-0.0504063778,
-0.0260846484,
-0.015255576,
-0.01862639,
-0.0134838289,
0.0325208567,
-0.0474565625,
0.0007528921,
-0.0131329698,
0.0288962312,
0.0554873198,
0.002406728,
-0.0422440767,
0.03589724,
0.1123057082,
0.0034912536,
-0.0293841325,
-0.0015465465,
-0.0412902385,
-0.0058275862,
0.0043422198,
-0.0140442187,
0.0228493251,
0.0034919032,
-0.0435875989,
-0.0167880468,
0.01160035,
0.0348402672,
0.0050351652,
-0.0674017593,
0.0349018648,
0.0191340744,
-0.0440473743,
-0.0301863365,
-0.0011195036,
-0.0034938464,
-0.0575625002,
0.0343131311,
0.0233726725,
0.0204374474,
-0.035628289,
-0.0049171597,
0.013241454,
0.0230267048,
-0.0641987845,
-0.0676531047,
0.0340759233,
-0.055534333,
-0.000162968,
-0.0159938578,
-0.0133346114,
0.0056951377,
0.0386698991,
-0.0568439253,
-0.0620319359,
0.0714464933,
-0.0505029336,
0.0029014193,
-0.0687741488,
-0.0432719029,
0.0105359377,
0.0606059991,
0.0138267064,
0.0517687388,
-0.0374165773,
-0.0503977463,
0.0180680864,
0.0254883505,
0.0382459648,
0.0039348765,
0.0071916184,
-0.0137496758,
-0.0122367321,
-0.0095005017,
0.0217142999,
0.0005358708,
0.025960926,
0.0509105437,
-0.0299640168,
-0.0283736996,
-0.0398827083,
0.023701692,
-0.00718698,
-0.0120923622,
-0.0131411785,
-0.0200332943,
0.0120233977,
-0.0063352389,
0.0037249175,
0.0133365132,
0.076917842,
-0.0110759269,
-0.0582588576,
0.0014171981,
-0.014567866,
-0.0318210907,
0.073274523,
0.0217429046,
0.0011602311,
-0.0345390923,
-0.0438903496,
-0.015838787,
-0.0105197858,
-0.0151500609,
-0.0134902624,
-0.0227197148,
-0.0314998366,
-0.0024118728,
0.0368742868,
0.0171539653,
0.022100145,
-0.0001492957,
-0.0173399653,
-0.0269269794,
0.0100486567,
-0.018332988,
0.0675172135,
-0.031267833,
0.0029649949,
-0.0032328847,
0.0638299733,
0.0082727987,
-0.030637864,
-0.03670137,
0.0202414412,
0.0032774643,
-0.016747402,
-0.0383249894,
-0.0451833494,
-0.0322035812,
0.0724070966,
-0.0100087095,
-0.0456579775,
0.0504559837,
0.0093301823,
-0.0026391814,
0.0017783314,
-0.0774246454,
-0.0698298663,
0.0154260658,
0.0015076526,
0.0097667286,
-0.0208126437,
-0.0398459621,
0.006847491,
-0.0055040102,
0.0017853621,
-0.0080507901,
-0.0175291579,
-0.0486550927,
0.012059194,
0.0319778509,
0.0012518816,
-0.0602357015,
-0.0272628404,
-0.0184520837,
0.0038129508,
-0.0111572258,
-0.0307421368,
-0.0201562755,
0.029731635,
-0.0477927811,
0.0287593286,
0.0139628043,
-0.0097559039,
0.0091266064,
0.0153477071,
0.0349719152,
-0.0395760201,
0.0646736324,
-0.0199403223,
0.0381724611,
0.005067925,
0.005455967,
0.0091272164,
-0.0632312223,
0.0466089249,
-0.0329319388,
-0.0159550719,
-0.0039018313,
0.017446775,
-0.0261034537,
-0.0546788014,
-0.0443279184,
0.1184936911,
0.0145425517,
0.0235382468,
0.0258953311,
0.0016465042,
0.0239132885,
-0.0056464197,
-0.0460234992,
-0.0695258453,
-0.008767386,
0.0280295778,
0.03069916,
-0.0266206414,
-0.0686252266,
0.0118967723,
-0.0670918301,
0.0263080578,
0.0166051425,
-0.0263592135,
-0.047350958,
-0.0670135245,
-0.0046526492,
-0.0436274372,
0.049722448,
-0.0174715407,
0.0303007495,
0.0242071413,
0.0469528921,
-0.0903142616,
-0.0749160945,
0.0132397581,
-0.0033616205,
0.0308744982,
0.0261267107,
0.0066735228,
-0.0292364359,
-0.0363557823,
-0.0510221533,
0.0183467809,
0.0520960465,
-0.0228206068,
0.004139462,
0.0403682664,
-0.0408693142,
-0.0261262432,
0.0969418064,
-0.1222056374,
0.0356593132,
0.0041067717,
0.0800408572,
-0.0542734973,
-0.0347289704,
-0.0511323847,
-0.0238543712,
0.0492477529,
0.0632312521,
-0.0201150645,
-0.0505201928,
-0.0230521467,
0.0168957692,
-0.0469852388,
-0.0293947384,
-0.0009472638,
-0.0047073252,
0.077051498,
-0.0328185894,
-0.011581827,
-0.0550938062,
-0.0342664085,
0.0102706905,
-0.0049118912,
0.0293851532,
-0.0156608783,
0.0034405349,
-0.0160542373,
-0.0058368258,
0.0076958658,
0.0525444858,
-0.0005413251,
0.0711349472,
0.0167244356,
0.0079928134,
-0.0570959896,
0.0211702678,
-0.0780757368,
-0.01288688,
0.0392101035,
-0.0221787505,
0.013307102,
0.042639818,
-0.0010392483,
0.0209201276,
0.0499498285,
0.0539577529,
0.0024935601,
-0.020527577,
0.0290135276,
-0.0349578187,
0.0282473247,
0.0362419896,
0.0015017941,
0.0607299954,
0.0008629724,
-0.0008836578,
-0.0778791532,
-0.0128323231,
0.0102914562,
-0.0197500195,
0.0470356755,
0.0296200756,
-0.0050151753,
-0.0146138826,
0.0070298458,
0.0190709103,
0.0115294494,
-0.0237745438,
0.0358302593,
0.0362243392,
-0.0197529029,
-0.0204518326,
0.0064491853,
0.052466169,
-0.0404780693,
-0.0212222971,
0.039948754,
0.0149017181,
0.0784353614,
-0.0044515152,
0.0180414896,
0.0238059741,
0.0145255551,
0.0113444235,
0.030888008,
0.0243486892,
0.0422313213,
0.0015663088,
0.0239106696,
-0.0250497703,
-0.0288977921,
-0.0583739318,
-0.0181327518,
0.0444036685,
-0.0028307815,
0.0384568796,
-0.0358381867,
-0.0181120113,
-0.0384842306,
0.0501378439,
0.0117208557,
0.0583090596,
0.0176912621,
0.0482401736,
-0.0153179634,
-0.010301481,
-0.0554475263,
0.0143046752,
0.084196195,
0.0116419448,
0.0297509246,
-0.0124723632,
0.0006540341,
-0.015341402,
0.0101330327,
0.0401163734,
0.0255130455,
0.0050366442,
0.0115573183,
-0.0378068537,
-0.0062400606,
-0.0057038795,
-0.0281510577,
0.0036267829,
0.0209087115,
-0.0036310328,
0.0931239575,
-0.0352638662,
0.0400814153,
0.0213337336,
-0.035207741,
-0.0081308903,
-0.0287264418,
-0.0596138239,
-6.07800007e-33,
0.0382054225,
-0.0517324433,
-0.0183218904,
-0.0365870818,
0.0272379965,
0.0314710625,
0.0138881318,
-0.0338193588,
0.017862089,
0.0250274111,
0.0234807208,
0.0326046497,
0.0016289012,
0.02345277,
0.0391973779,
-0.0271997061,
0.0880957171,
0.0366983525,
0.0742063522,
0.042725984,
0.099286586,
0.044392325,
0.0350436196,
-0.0813168436,
0.031428352,
-0.0301481616,
-0.0074997456,
0.0051251361,
0.0179004967,
0.0329292752,
-0.0125741195,
0.0530236885,
0.0211673472,
-0.0415867604,
-0.0337052494,
-0.0242938437,
0.0056809839,
-0.0334465504,
-0.01388474,
-0.0102070775,
0.0131278308,
-0.0556825548,
0.0165178981,
0.0241533238,
0.0022750238,
0.0433649868,
0.0277385935,
-0.0445222259,
0.0247564279,
0.0038747187,
-0.010390209,
0.0252744686,
-0.0740852952,
0.0385163054,
0.0088231759,
0.0184989944,
-0.003623398,
0.0086062672,
-0.0523838699,
0.0227819271,
0.0377602093,
-0.0488311835,
0.0309755299,
0.0066137607,
-0.0221561007,
0.0380122028,
0.0081949057,
-0.0396530069,
0.0040066643,
0.0408791862,
0.0198200885,
0.0756051838,
0.0415478684,
0.026838718,
0.0139180236,
-0.0567721874,
-0.1012191325,
0.0292674731,
0.0039847512,
0.0219982993,
0.063346684,
-0.0093921367,
-0.069117561,
-0.0064229709,
0.032920491,
-0.0162087772,
-0.0029552968,
0.0196112413,
0.0305736847,
0.0171960946,
-0.0434310958,
0.0293633752,
0.0396796428,
-0.0621152595,
-0.0025766813,
0.033892218,
0.0114926528,
-0.0420146734,
0.0296983048,
-0.009381582,
-0.0327072926,
-0.050849814,
0.0033503214,
0.0028700035,
0.0316876546,
-0.0567693971,
0.034467008,
0.0584079698,
-0.0397517495,
-0.0020672227,
-0.0043322244,
0.0210255105,
-0.0362127982,
0.0246567465,
0.0239307098,
-0.0393956378,
-0.0206897724,
0.0490396433,
-0.0161965881,
0.0186902434,
0.0041681686,
0.0072581652,
-0.054614041,
-0.0018484291,
-0.0342021585,
-0.0292263217,
-0.0077113239,
-0.0398425087,
0.0540813617,
-0.0512798056,
0.0066066394,
-0.0113147385,
2.892e-7,
-0.0132267782,
0.0527305156,
0.0180787668,
0.0090960432,
0.0565181263,
0.0570157506,
-0.0015445866,
0.017336715,
-0.0458746813,
0.0192144904,
0.0637983605,
-0.007924716,
-0.0053221867,
0.0225071702,
0.0088617923,
-0.0292189717,
-0.0759592131,
-0.0097242929,
-0.0842568651,
-0.0033193475,
0.0849032477,
0.1158587784,
0.0440428443,
-0.0460451022,
0.0878648385,
-0.0891980156,
0.0272151753,
-0.0167198479,
-0.001736806,
0.0132732093,
0.0432343856,
0.066402413,
0.0197173413,
0.0196014997,
0.0177968405,
0.0249304175,
-0.046540387,
0.0107787708,
-0.001232335,
0.0083537251,
-0.0272247661,
-0.0107959453,
0.016307652,
-0.0545910932,
0.0544382259,
-0.0032354461,
-0.0253351294,
-0.0162911192,
-0.0210800711,
0.0079954406,
0.0099857245,
0.0039924057,
0.0335969217,
0.0123326443,
-0.0181286,
-0.0325227901,
0.0331291631,
-0.0258054193,
0.0188451372,
-0.0216752421,
-0.0495724045,
0.0071998946,
-0.0159849375,
0.0112533076,
0.0736892596,
0.0329747573,
-0.0021668246,
2.753494957e-34,
0.0429750532,
-0.0003347888,
-0.0328629725,
0.0244089998,
0.0073905438,
0.0196606163,
-0.0643685088,
0.0229926016,
0.0201566741,
-0.0342191719,
-0.0225587934
] |
http://arxiv.org/abs/2405.15362v3
|
http://arxiv.org/pdf/2405.15362v3
|
Pipeline Parallelism with Controllable Memory
| 2024-05-24T08:54:36 | 2024-06-10T11:24:06 |
Penghui Qi, Xinyi Wan, Nyamdavaa Amar, Min Lin
|
cs.LG
|
cs.LG, cs.CL, cs.DC
|
Pipeline parallelism has been widely explored, but most existing schedules
lack a systematic methodology. In this paper, we propose a framework to
decompose pipeline schedules as repeating a building block and we show that the
lifespan of the building block decides the peak activation memory of the
pipeline schedule. Guided by the observations, we find that almost all existing
pipeline schedules, to the best of our knowledge, are memory inefficient. To
address this, we introduce a family of memory efficient building blocks with
controllable activation memory, which can reduce the peak activation memory to
1/2 of 1F1B without sacrificing efficiency, and even to 1/3 with comparable
throughput. We can also achieve almost zero pipeline bubbles while maintaining
the same activation memory as 1F1B. Our evaluations demonstrate that in pure
pipeline parallelism settings, our methods outperform 1F1B by from 7% to 55% in
terms of throughput. When employing a grid search over hybrid parallelism
hyperparameters in practical scenarios, our proposed methods demonstrate a 16%
throughput improvement over the 1F1B baseline for large language models.
|
Pipeline Parallelism with Controllable Memory: Pipeline parallelism has been widely explored, but most existing schedules
lack a systematic methodology. In this paper, we propose a framework to
decompose pipeline schedules as repeating a building block and we show that the
lifespan of the building block decides the peak activation memory of the
pipeline schedule. Guided by the observations, we find that almost all existing
pipeline schedules, to the best of our knowledge, are memory inefficient. To
address this, we introduce a family of memory efficient building blocks with
controllable activation memory, which can reduce the peak activation memory to
1/2 of 1F1B without sacrificing efficiency, and even to 1/3 with comparable
throughput. We can also achieve almost zero pipeline bubbles while maintaining
the same activation memory as 1F1B. Our evaluations demonstrate that in pure
pipeline parallelism settings, our methods outperform 1F1B by from 7% to 55% in
terms of throughput. When employing a grid search over hybrid parallelism
hyperparameters in practical scenarios, our proposed methods demonstrate a 16%
throughput improvement over the 1F1B baseline for large language models.
|
[
0.0188507251,
0.0879717991,
-0.0284317341,
0.0305724833,
-0.0559822246,
0.0200829785,
0.0463646576,
-0.0356261991,
0.0247780923,
-0.0016452124,
0.0028291373,
-0.0764117688,
-0.0128206667,
0.0444285348,
-0.0140529964,
-0.0195490681,
0.0260659903,
-0.0205100849,
0.0009702014,
-0.0283240415,
-0.0309740063,
0.0333829522,
-0.0252387263,
0.0061367163,
0.0202827062,
-0.0344607644,
-0.0104390876,
0.0270556416,
0.0037613884,
0.0104939966,
0.0285091922,
0.0280349646,
0.0095710335,
0.0243084151,
0.0000023137,
-0.0004304836,
-0.0004016939,
-0.0007856122,
0.0091140363,
0.0618400089,
0.0279554538,
-0.0265670884,
-0.0071020536,
0.0123277148,
0.0118125966,
-0.0141275134,
0.05958445,
-0.0130898971,
0.0339603089,
0.0346895456,
-0.0118883457,
-0.0073764245,
0.02995806,
-0.0196722466,
-0.0083075203,
-0.0739338398,
0.0020630821,
-0.0258679651,
-0.0013628819,
-0.0211970396,
0.0020004648,
0.035814777,
0.0160783939,
-0.00938717,
-0.0038389554,
0.0084183821,
-0.0176855791,
-0.0350745693,
-0.0026042636,
0.0505404621,
-0.0982880741,
-0.0525841452,
0.0102175586,
-0.0451776274,
-0.0552776381,
-0.0334528089,
-0.0214805324,
0.0316213146,
-0.0098412391,
0.0101020141,
-0.0114787193,
-0.0247058943,
0.0607439689,
-0.0206503626,
-0.0363341682,
0.0421820879,
0.0326633453,
0.0018115857,
-0.0163942873,
0.0079068039,
-0.0195807442,
-0.0185640175,
0.0316232443,
-0.0056857052,
-0.0550353974,
0.0019329134,
-0.0347420312,
-0.0693738163,
-0.0106331399,
-0.0315690897,
0.0514386408,
0.0591352694,
0.0399862453,
0.0392783657,
-0.0444394536,
0.0385296047,
-0.085421212,
0.022392381,
-0.0393636897,
-0.0269503519,
-0.0086381109,
0.0079942327,
-0.0444140695,
0.009284432,
-0.0314244553,
-0.0154638933,
-0.0208386052,
0.0247987173,
-0.0051987669,
-0.0110167805,
-0.0094716176,
-0.0400217772,
0.0039429944,
0.0172884334,
0.0580281541,
0.0382112674,
-0.0335915238,
-0.0442419611,
-0.0397009701,
-0.0267030522,
-0.0521121845,
0.0208085291,
0.0019702292,
-0.0313421637,
0.0182242226,
0.0267681982,
-0.0327195749,
-0.0340708606,
-0.0625389889,
-0.0085050547,
0.0166897122,
-0.0435061119,
0.0235444624,
-0.0188061018,
0.0431458093,
0.0074680275,
0.0110121779,
0.0177104603,
0.006240542,
0.0372391865,
-0.0383497328,
0.0407225005,
0.0551452413,
-0.0170760266,
0.0219120327,
-0.0102132997,
-0.0045445263,
0.0772341713,
0.0181786306,
0.0261202026,
0.0203980617,
-0.0156921595,
-0.0184329897,
0.0354836397,
0.02743949,
-0.0024097862,
-0.0006550765,
-0.0087511213,
0.0297301169,
0.0040306365,
0.0784940347,
0.0431176163,
-0.0395721085,
-0.0244509783,
0.0090589039,
0.0007588621,
-0.0002163204,
-0.0149467709,
0.0597522557,
0.0123764146,
-0.0100108646,
-0.0023733343,
-0.0279046651,
0.0384698212,
0.0546390824,
0.015409871,
0.0197568312,
0.067532368,
-0.0469219983,
0.0484167561,
-0.0061448412,
0.0301959086,
-0.0302898549,
0.0238646772,
0.0570876226,
0.0082615316,
-0.0633382425,
0.0859955102,
0.0095121367,
-0.0300855152,
-0.0558478534,
-0.025872441,
-0.0254928563,
0.0553699099,
0.0401366539,
-0.0097892405,
-0.0110018691,
0.0362874307,
0.0090015326,
0.0078974506,
0.0319193229,
0.0061360602,
0.0400853753,
-0.0436000042,
-0.0036888218,
-0.0103613315,
0.0000646858,
0.0130660217,
0.0742684677,
0.0027833579,
0.0196037777,
0.029903762,
-0.0105155511,
0.0173237734,
-0.0306937639,
0.0084852045,
0.0502889343,
-0.0071935803,
0.0156042641,
-0.0175645184,
-0.0202078447,
0.0045121992,
0.0585830659,
-0.0357230715,
-0.0085240407,
0.0032290858,
-0.0602920689,
-0.0180588346,
0.000505896,
0.0426606797,
0.0042189704,
0.0399364308,
-0.0583554283,
-0.0421165638,
0.0681753382,
-0.0065005035,
0.0459060073,
-0.0109888045,
0.0528518595,
-0.0185931846,
0.0198952928,
-0.0453023985,
0.0252355598,
-0.0354224406,
0.0262651294,
-0.0052732406,
0.0282212105,
-0.0232451372,
-0.019767493,
-0.0276704375,
0.0433942191,
-0.0007111264,
-0.003552946,
-0.0350312926,
-0.0412616991,
0.0071961386,
-0.0197083391,
-0.0483347699,
0.0301948395,
-0.0329604223,
0.0127723394,
0.0166159756,
0.019585561,
0.0610749647,
0.0001277807,
0.005092246,
0.0386923477,
-0.0062763458,
0.0493495353,
-0.0386776403,
-0.0093161985,
-0.0580072738,
-0.0180685893,
-0.0931718573,
0.0068503027,
0.0572866797,
-0.0302765518,
-0.0454557091,
-0.0204889197,
-0.0339420177,
0.0499865897,
0.0338792205,
-0.0152056655,
0.0051832907,
-0.0082681905,
-0.034087915,
-0.0090399776,
0.0596693046,
-0.0221753903,
0.0022368957,
0.0333624929,
0.0140771158,
-0.010961188,
0.0502051488,
-0.021446174,
-0.0201263241,
0.0043774676,
-0.0125612561,
-0.0146636609,
0.0740314573,
-0.0050625587,
-0.0089602731,
-0.0344381854,
-0.0333362408,
-0.0067439508,
-0.0174720939,
-0.0024377662,
-0.0294327252,
-0.0139865223,
-0.0023342436,
-0.0187636893,
-0.0744322687,
0.026891768,
-0.0323004946,
-0.0301884282,
-0.0331091806,
-0.0126282265,
-0.0415855609,
0.0416320786,
0.0269123763,
0.025644239,
-0.0578452274,
-0.0209952816,
0.0138016967,
0.0175735522,
-0.0306994542,
-0.0077798078,
0.0463180952,
-0.0167254936,
0.0263201371,
-0.0246272385,
-0.035353031,
-0.027112199,
-0.0080700535,
0.0020442179,
0.0427346863,
0.0912415683,
-0.071685791,
-0.0553945042,
0.0202893149,
-0.0290936735,
0.0796629563,
-0.0064318506,
-0.004047269,
-0.0308485851,
0.0445723645,
0.0127946641,
-0.0354141854,
0.0297155958,
0.0109149497,
0.0304123908,
0.0269022509,
0.064593479,
0.0042842622,
-0.0389468111,
0.0028388561,
0.0371643938,
-0.0087595899,
-0.0355328172,
-0.0249221753,
0.0043883752,
-0.0525737442,
-0.0473149791,
0.0471823923,
0.0301731825,
-0.0495087318,
-0.0050671515,
0.0479123071,
0.0562240183,
0.0217661038,
0.0257377233,
-0.0325418636,
-0.0051122406,
0.0088537112,
0.0249088928,
0.0186297111,
0.0075814719,
-0.0146481274,
-0.0903325081,
0.022387784,
0.0098599177,
-0.0320847891,
0.0054902546,
-0.0442580879,
-0.0379293561,
0.0017859966,
0.0238995459,
0.0279309079,
0.0145920208,
-0.018953139,
-0.0148454895,
0.0624872036,
-0.007688479,
0.0316077359,
-0.0071804486,
0.0181574654,
-0.0017059414,
-0.0271988306,
0.0285859127,
-0.0405967273,
0.0431563668,
-0.0367017351,
-0.0007373132,
-0.0361733362,
0.0264704712,
0.022610452,
-0.045814544,
-0.0020597999,
0.0015859578,
-0.0522258021,
-0.0112841446,
-0.030851461,
0.0920057744,
-0.0046741832,
-0.0217605401,
-0.0214004405,
0.0037723815,
0.0124342302,
0.0819583088,
0.0034748369,
-0.0300661717,
-0.0421109349,
-0.0041362536,
-0.049661234,
-0.1182285547,
0.0000074516,
0.0153985647,
0.0347948037,
-0.0122059043,
-0.0120883277,
0.061842598,
0.0162962209,
0.064603731,
-0.0096728383,
-0.0338648409,
0.002627892,
-0.0098409392,
-0.0946698263,
0.0165441949,
-0.01778844,
0.0269918796,
-0.0461318195,
-0.0356727801,
0.0215796512,
0.0324885361,
-0.0186036862,
0.0750843361,
-0.1449220926,
-0.0727957711,
-0.0047136308,
0.0098765474,
0.0153867388,
0.041103147,
0.0289784428,
-0.0165316258,
0.00895382,
-0.0123240706,
-0.0321623459,
0.0467274152,
0.0018377033,
-0.0254470371,
-0.0381422751,
0.0295016393,
-0.0131284948,
-0.0027702979,
0.0404138826,
0.0420870632,
0.0094607584,
0.0169715229,
-0.0111349812,
-0.0161441285,
0.0092457514,
0.0391145572,
0.0418278091,
-0.0058546313,
0.0646590963,
0.0158690326,
0.0257695559,
-0.0488015302,
-0.0082600182,
0.0124662509,
-0.0328930244,
-0.0359029733,
-0.0289043989,
0.0625542328,
0.0091328053,
-0.0014473038,
-0.0610024035,
-0.0431991182,
-0.0208946876,
0.0325139649,
-0.0011406783,
-0.0248140227,
0.0474763326,
-0.0063051516,
0.0306590982,
-0.0145895369,
0.0254122596,
-0.0006304001,
0.0306439437,
-0.0686242357,
-0.0152048673,
-0.0541501604,
-0.0773632377,
-0.0237258691,
-0.0352076925,
0.0631667599,
-0.0413169898,
0.0556060188,
0.0005918715,
-0.0213359147,
-0.0061754254,
-0.0322036371,
0.0083826715,
-0.0331521966,
-0.000081695,
0.0264341272,
-0.0091441181,
-0.0143551417,
0.0575270914,
0.0062384009,
0.0333653092,
0.0207488444,
-0.000214873,
-0.0118113905,
-0.0586541519,
-0.0300405584,
0.0734667405,
-0.0036828998,
-0.0376069248,
-0.0227998029,
-0.0355128311,
0.0130812321,
-0.0188337564,
-0.0063125552,
-0.004562492,
0.0117888246,
0.039703276,
0.0338459574,
-0.0351328775,
-0.0292556211,
0.0076561873,
0.0138516352,
-0.0246598031,
-0.0351527855,
-6.560179771e-33,
-0.0078218505,
-0.0113779344,
0.0044876109,
-0.000970145,
-0.0073074647,
0.0099706938,
-0.0504368618,
-0.0303760096,
-0.0204067975,
-0.0048523177,
-0.0192193482,
0.0199095272,
0.0363449976,
0.0299891438,
0.0099902367,
0.036438778,
0.0573636778,
-0.0252052881,
-0.0072226333,
0.0758373514,
0.0568577982,
0.0298306402,
0.0351317227,
-0.0238337386,
-0.0586943328,
0.0061394074,
0.0138369119,
0.0019788772,
-0.0093159983,
0.0204626489,
-0.0162568055,
0.0332606137,
0.0023144705,
-0.1038195565,
-0.0357750691,
-0.0006134093,
-0.0825903937,
0.0230322462,
0.0293897279,
0.010966992,
-0.0386758111,
-0.021990899,
0.0202884842,
0.002216388,
0.0256440714,
-0.0293186083,
-0.0439274646,
0.0285066962,
-0.0397402756,
-0.0314690396,
-0.0335832983,
0.0143155679,
-0.0041892128,
0.0698314607,
0.0588168241,
-0.0314386114,
-0.0040632361,
0.0440314747,
-0.0700506493,
-0.0056258407,
-0.0294301994,
0.0728555694,
0.0631761029,
0.0430179425,
0.0692643523,
0.0363265797,
0.0798000395,
0.0446758196,
0.025943175,
0.0364924408,
0.0737430602,
-0.0174353421,
0.0314272717,
-0.0434461087,
0.0488564484,
-0.0088672489,
-0.0605728887,
-0.057271231,
0.0217707176,
0.0068751201,
0.0558327101,
0.0226651803,
-0.0396515094,
-0.0393970609,
-0.0261565391,
-0.0040414426,
-0.0568783879,
0.0221521333,
0.0306187421,
0.0381325893,
-0.0273416415,
0.0076011983,
0.0391807221,
-0.0068818121,
0.0216311701,
0.0768751204,
-0.0154850753,
0.0331244357,
0.0044512167,
0.0375405699,
-0.0403226204,
-0.0344668664,
-0.0258470271,
-0.009759767,
0.0271734353,
-0.0010332719,
0.0181362964,
0.0306556467,
-0.0801751167,
-0.0224104133,
-0.0707706362,
0.0650110021,
0.006773497,
0.0930698961,
0.011140462,
0.0077859969,
0.0136717129,
0.0374000743,
0.0106073814,
-0.0330509357,
0.001263298,
0.0651003569,
0.0162478387,
0.0247112848,
-0.032399822,
0.0091312379,
-0.0411631614,
-0.0111654736,
0.0524692424,
-0.0940281972,
-0.0206360165,
0.0443498567,
2.831e-7,
-0.0253452901,
0.0050114505,
0.0009147676,
0.0882892609,
0.0529752187,
-0.0055089146,
0.042507071,
-0.0093532298,
-0.004802532,
-0.0053725992,
0.0132831745,
0.0286174715,
0.0296596326,
0.0113119679,
-0.0817326754,
-0.0057841423,
-0.025176743,
0.0491326898,
-0.033344008,
-0.0692502707,
-0.0296440683,
0.080216758,
0.0180773921,
-0.0447683781,
0.0077380422,
-0.0388804674,
0.0022893327,
0.0315295085,
0.0340052769,
0.0325403512,
-0.0243620388,
-0.009456519,
-0.0382401273,
0.0539201424,
-0.0246954393,
0.0118135391,
-0.0287349056,
0.1000280306,
0.004874012,
-0.0823544711,
0.0258240774,
-0.0188255627,
-0.0155468015,
-0.0587288998,
0.0646084398,
0.0002205496,
-0.0003752789,
-0.0501179285,
0.0187393017,
-0.0223704055,
0.0031397145,
0.0191370696,
0.0233264845,
0.0569674447,
0.0313107185,
-0.0258942209,
0.0133643663,
-0.0715514496,
-0.0383554995,
0.0015574491,
-0.0718827695,
-0.0195666999,
-0.0695974678,
0.0137155745,
0.0118874628,
-0.0305326525,
-0.0215930287,
3.31415686e-34,
0.0613402389,
-0.0245414209,
0.0347929038,
0.0310046375,
0.0297603458,
-0.0167862047,
-0.0339625441,
0.0174821392,
-0.0329096764,
-0.0969330966,
-0.0333816893
] |
http://arxiv.org/abs/2405.15349v1
|
http://arxiv.org/pdf/2405.15349v1
|
UnKE: Unstructured Knowledge Editing in Large Language Models
| 2024-05-24T08:42:40 | 2024-05-24T08:42:40 |
Jingcheng Deng, Zihao Wei, Liang Pang, Hanxing Ding, Huawei Shen, Xueqi Cheng
|
cs.CL
|
cs.CL
|
Recent knowledge editing methods have primarily focused on modifying
structured knowledge in large language models, heavily relying on the
assumption that structured knowledge is stored as key-value pairs locally in
MLP layers or specific neurons. However, this task setting overlooks the fact
that a significant portion of real-world knowledge is stored in an unstructured
format, characterized by long-form content, noise, and a complex yet
comprehensive nature. The "knowledge locating" and "term-driven optimization"
techniques conducted from the assumption used in previous methods (e.g., MEMIT)
are ill-suited for unstructured knowledge. To address these challenges, we
propose a novel unstructured knowledge editing method, namely UnKE, which
extends previous assumptions in the layer dimension and token dimension.
Firstly, in the layer dimension, we discard the "knowledge locating" step and
treat first few layers as the key, which expand knowledge storage through
layers to break the "knowledge stored locally" assumption. Next, we replace
"term-driven optimization" with "cause-driven optimization" across all inputted
tokens in the token dimension, directly optimizing the last layer of the key
generator to perform editing to generate the required key vectors. By utilizing
key-value pairs at the layer level, UnKE effectively represents and edits
complex and comprehensive unstructured knowledge, leveraging the potential of
both the MLP and attention layers. Results on newly proposed unstructure
knowledge editing dataset (UnKEBench) and traditional structured datasets
demonstrate that UnKE achieves remarkable performance, surpassing strong
baselines.
|
UnKE: Unstructured Knowledge Editing in Large Language Models: Recent knowledge editing methods have primarily focused on modifying
structured knowledge in large language models, heavily relying on the
assumption that structured knowledge is stored as key-value pairs locally in
MLP layers or specific neurons. However, this task setting overlooks the fact
that a significant portion of real-world knowledge is stored in an unstructured
format, characterized by long-form content, noise, and a complex yet
comprehensive nature. The "knowledge locating" and "term-driven optimization"
techniques conducted from the assumption used in previous methods (e.g., MEMIT)
are ill-suited for unstructured knowledge. To address these challenges, we
propose a novel unstructured knowledge editing method, namely UnKE, which
extends previous assumptions in the layer dimension and token dimension.
Firstly, in the layer dimension, we discard the "knowledge locating" step and
treat first few layers as the key, which expand knowledge storage through
layers to break the "knowledge stored locally" assumption. Next, we replace
"term-driven optimization" with "cause-driven optimization" across all inputted
tokens in the token dimension, directly optimizing the last layer of the key
generator to perform editing to generate the required key vectors. By utilizing
key-value pairs at the layer level, UnKE effectively represents and edits
complex and comprehensive unstructured knowledge, leveraging the potential of
both the MLP and attention layers. Results on newly proposed unstructure
knowledge editing dataset (UnKEBench) and traditional structured datasets
demonstrate that UnKE achieves remarkable performance, surpassing strong
baselines.
|
[
0.0433371142,
0.044094272,
-0.0015468691,
0.0366191901,
-0.0158489589,
-0.0154420314,
0.0033968689,
-0.0326267295,
-0.0406862944,
-0.0534511246,
-0.0392375179,
-0.0041725803,
0.0020278667,
0.0430026129,
0.0223658867,
-0.0394282565,
0.0542656928,
-0.0102726202,
0.0479340069,
0.028242616,
-0.0099576786,
-0.023680212,
0.0111705083,
0.015068558,
-0.0292696562,
-0.04183615,
-0.0506102666,
0.0361303538,
0.0010695407,
-0.0043310504,
0.0222252905,
0.0558938608,
0.0068762233,
0.0252176039,
0.0000022734,
-0.022639744,
-0.0256884433,
0.0032186816,
-0.0227996446,
0.0444603413,
-0.0013801417,
-0.0163390934,
-0.0549862906,
-0.0202316269,
-0.017420467,
-0.0068410262,
0.0542300791,
0.0576539971,
0.0673199669,
0.1218078136,
-0.0176385492,
-0.0156594273,
0.0287358444,
-0.0176660959,
-0.0120447055,
-0.0224267282,
0.0468884222,
-0.0275758617,
-0.0986084491,
0.0165644828,
-0.0238218028,
0.0397678651,
0.0127164125,
-0.0490566567,
0.0733134523,
0.0047235419,
-0.0632770285,
-0.0008053539,
0.012567264,
0.0675944164,
-0.0290957782,
-0.0230786391,
0.0217812005,
-0.0146801528,
0.0207539108,
-0.0076359017,
-0.0141547555,
-0.0138500556,
0.0137304151,
0.0110186776,
0.0026305143,
-0.0819482878,
0.0505074263,
-0.0001526166,
-0.0181618333,
0.079062216,
0.0374757536,
-0.0428445302,
-0.0278820973,
-0.0592744239,
-0.025078278,
-0.0973276272,
0.027760338,
0.0450950451,
-0.0106177544,
0.0222469177,
-0.063085936,
-0.0176648721,
-0.0371294245,
-0.0983861461,
0.0573855676,
0.0405010656,
-0.0041325279,
0.0764439479,
-0.037669953,
-0.0008310965,
-0.0089497566,
0.0315268189,
-0.0213905629,
0.0175874587,
-0.0443133041,
-0.0161792524,
-0.0578009598,
0.0175175834,
-0.082000263,
0.0133639593,
-0.0265457444,
0.0573141426,
-0.0163393188,
0.0027928483,
-0.0013607816,
0.0079349652,
0.0221318062,
0.0030928128,
-0.026337089,
0.0217603613,
-0.0075748935,
-0.0045004888,
0.0222413465,
-0.0155373374,
-0.0104361176,
0.0030032338,
0.024332704,
-0.0450131111,
0.0237250682,
0.0470847897,
-0.0160210785,
-0.0185610224,
-0.0747309923,
-0.0569005907,
-0.0081315832,
-0.0372484848,
0.0146875577,
-0.0513900556,
0.0562376045,
-0.0410993248,
-0.0743966475,
-0.0668521598,
0.0055495626,
0.0231069922,
-0.0581058227,
0.0827857181,
-0.0413977206,
0.0305178706,
0.0259222444,
0.00659326,
-0.0130657824,
0.016241299,
0.0161687694,
0.010546837,
0.0697918609,
-0.0029415509,
-0.0015315313,
0.0189666599,
0.0792375579,
0.0200196039,
-0.0339753591,
0.0265655946,
-0.0216478445,
0.037131872,
0.0396559909,
-0.0013750355,
-0.039489869,
0.0182599854,
0.0844657347,
-0.0166057013,
0.0430620015,
0.0529135726,
-0.0010691803,
-0.0199390762,
0.007006635,
-0.0421088152,
-0.0662933514,
0.0304151662,
-0.0330936909,
0.0083430614,
0.0219965335,
-0.0122278677,
-0.0262476895,
-0.0425336324,
-0.0220501591,
0.0327791609,
0.0122329267,
0.0063529154,
0.0264299065,
0.0397153795,
-0.0460761935,
0.1081834733,
0.0245968979,
-0.03171359,
-0.006154398,
-0.0319475196,
0.0117174434,
0.0695923716,
0.0682765841,
-0.0472509563,
-0.0880754367,
0.0027604916,
-0.0049700509,
0.0031855681,
0.0133683309,
-0.0273645818,
0.0205974746,
-0.0294428784,
-0.0174822807,
-0.0454661921,
0.0146860769,
-0.0118584037,
0.042131789,
-0.009329279,
0.0179211441,
-0.0524199679,
-0.0034652958,
0.0064086393,
-0.0120230066,
-0.0262470096,
-0.0017328787,
0.0203848947,
0.0259718467,
0.0104686338,
-0.0055851848,
-0.0035168985,
0.0177021865,
-0.006180306,
-0.0382430181,
-0.0272611603,
-0.0526920035,
-0.0361144431,
0.0680144653,
-0.0022947125,
-0.0341888182,
0.0112505108,
-0.0339700989,
-0.0206886102,
0.0302058179,
-0.0120458389,
0.0778278857,
-0.0338402018,
0.0214651991,
0.0205176454,
-0.0215036515,
-0.0172020029,
0.0644895434,
-0.0039484375,
-0.0352491587,
0.0274777394,
0.001420091,
-0.0121123027,
0.0007353677,
0.0068708472,
0.0188171118,
-0.0085212057,
-0.0103773279,
0.0001087361,
-0.0144442115,
-0.0466988906,
-0.0151831154,
-0.0680741891,
-0.0025061406,
-0.0111663183,
0.0239243358,
0.0527680293,
0.0310821999,
-0.0142915091,
-0.0005024792,
-0.0288050994,
0.0073013487,
-0.034340851,
0.0282845497,
0.0275149085,
0.0149665996,
0.0137193119,
-0.0419801995,
0.0141958343,
-0.0090430668,
0.0287167821,
-0.0294302497,
-0.03247419,
-0.0165635701,
0.0636611357,
-0.0548023954,
0.0124449646,
0.0068098186,
-0.0455466434,
-0.0376662202,
-0.0063898885,
0.0609982684,
0.059622094,
-0.0368957222,
0.0195128396,
0.0447818786,
-0.0681847706,
-0.0008413674,
0.0500503331,
-0.0252166148,
0.0341613777,
-0.0206906144,
0.0313062072,
0.03965085,
0.0924307853,
-0.0035906935,
-0.0472864471,
-0.0332625285,
0.0326506644,
0.0128956418,
-0.014688611,
-0.0590301678,
-0.005006372,
-0.0043273503,
0.0219427422,
-0.0683527589,
-0.0067722304,
0.0471850745,
-0.0151262721,
0.016036639,
-0.0220215339,
0.0009044391,
-0.026463015,
0.0533232503,
-0.0263860598,
0.0370589755,
-0.0552326143,
0.0064419219,
-0.0482476838,
-0.009999752,
0.0193594396,
-0.0011427192,
0.0359897763,
-0.0497943871,
-0.0007547464,
-0.0185285211,
-0.0152529953,
-0.0216218811,
0.0222925022,
0.0173020978,
0.0189510789,
0.0479298234,
-0.0515945442,
-0.0024376838,
-0.0132375723,
0.0234671626,
0.0902915373,
-0.0001990428,
-0.0110597238,
-0.0259227566,
0.0205728449,
0.0442011617,
-0.0049407436,
0.0249204151,
0.0122117866,
0.0094876559,
0.0746263936,
0.0153279612,
-0.03569673,
0.0120319659,
-0.0127521977,
-0.0041793589,
-0.0704917461,
-0.0046548448,
-0.0352610014,
0.0590755939,
-0.071823895,
-0.0213961266,
0.1095601693,
0.0472756438,
-0.0286173597,
0.0255763549,
-0.0286923181,
-0.0174132381,
0.0395337045,
0.0474456325,
-0.0642835945,
-0.023085095,
0.0288458113,
0.0019815569,
0.0068766675,
-0.0482684597,
-0.017088471,
-0.0608325414,
0.0020333475,
0.0090337517,
-0.0446114764,
-0.0525424592,
-0.0395173021,
0.0029579056,
0.0149188759,
0.0077203326,
0.025707759,
0.0306830704,
-0.0140478443,
-0.0651288182,
0.000362997,
-0.0567023829,
0.0280506387,
-0.032942228,
0.0739748999,
0.0423349291,
0.0363043733,
0.0155612454,
-0.0194064081,
0.0562382303,
0.0590069368,
0.0139748715,
-0.00838969,
-0.0056792228,
-0.017286811,
-0.0055301422,
-0.0221189819,
0.0376291238,
-0.0519836098,
-0.0354223214,
-0.0031220384,
0.0945146382,
-0.0034553451,
0.0061279284,
0.0034564079,
-0.0570488349,
-0.0114427833,
0.0698160976,
0.0434430204,
-0.0173969716,
-0.0213332791,
-0.0414266437,
-0.0192835685,
-0.0526194982,
0.0104529643,
-0.0044858297,
-0.0227689072,
0.0242246911,
-0.0707056299,
-0.0488120355,
0.0692169815,
0.0180030409,
0.0128425183,
0.0036965511,
0.0373654962,
-0.0384429246,
-0.0690445676,
0.0098044407,
-0.0421441309,
-0.0096148066,
-0.0513033867,
-0.0407868475,
-0.0231926739,
-0.0424155071,
-0.0207633264,
-0.0058666239,
-0.0417223498,
-0.0766321048,
0.0099808713,
0.0418250635,
0.0072466214,
0.0753799677,
-0.0210246723,
-0.0432359427,
0.0822342262,
-0.0008843719,
-0.0403663777,
0.0245529376,
0.0038830149,
-0.003316005,
0.0098343622,
0.056475915,
-0.054674495,
0.0035198443,
0.0568205826,
0.0541476682,
-0.0149050886,
-0.0246756282,
0.0259386878,
0.0152831869,
0.0563832745,
0.0539959073,
0.0049407673,
-0.0522964634,
0.0531457514,
-0.0116803339,
-0.0118263029,
-0.0308719128,
0.0149738193,
0.0009830267,
-0.0134765608,
-0.0158508401,
0.0143627627,
0.0035497083,
0.0145502584,
0.0147245359,
-0.0193761773,
0.0073441495,
0.0071597854,
0.0577468164,
0.0201606229,
-0.0052602687,
0.0564540662,
-0.0157912485,
0.0121139772,
-0.0500357188,
0.0223847199,
0.0251700208,
-0.0323906653,
-0.049013257,
0.0288060661,
-0.0441469476,
-0.0217611138,
-0.0205753278,
0.0008909036,
0.0612153076,
-0.0028125274,
0.0171462912,
-0.0236843564,
-0.0504526682,
-0.0025887855,
0.0366728604,
0.0303818602,
-0.0343907699,
-0.01239012,
0.0025154948,
-0.0115163298,
-0.0023559083,
0.0566370226,
0.0065709762,
0.1097952574,
-0.0174942259,
-0.021002546,
-0.0348220207,
-0.0074155754,
0.0005436425,
0.0393998139,
-0.0413411073,
0.0257878769,
-0.0097894678,
-0.0531220362,
-0.0103338677,
0.0039799395,
-0.0335763246,
0.0105105545,
0.0171703063,
0.0315436386,
0.0065516378,
-0.0096657211,
-0.0236549508,
0.0095728096,
0.0001842721,
-0.0305872504,
0.016448101,
-6.588688448e-33,
-0.014015859,
-0.0074664941,
0.0450939909,
0.016929619,
-0.0303156413,
0.0030191506,
-0.0531518236,
-0.016485557,
-0.0012209781,
0.0270247404,
-0.042566821,
0.0347684696,
-0.0102854986,
0.036729753,
-0.010655581,
0.0126290089,
0.0580611825,
0.0355342962,
-0.0041580983,
0.0210651029,
0.0383146405,
0.0414075628,
0.0853159726,
-0.0447141528,
0.0184442066,
0.011376177,
0.0126802465,
0.0428946614,
-0.0216140803,
0.0319022872,
-0.0124086365,
0.0150281005,
0.005009036,
-0.0450041518,
-0.0108856605,
-0.0439169034,
-0.0576849319,
-0.0189736038,
0.0096335215,
0.0422656834,
-0.0731829703,
-0.0940399989,
0.0368743353,
0.0248029009,
0.0265199151,
-0.0192153957,
0.0123088034,
-0.0182040036,
-0.0069676787,
-0.0601679049,
-0.0691193789,
0.0242811088,
0.0054814969,
-0.0172511172,
-0.0032003857,
-0.0242230147,
0.0165059268,
-0.0308493171,
-0.1286931187,
-0.0073502078,
0.0197712984,
0.0528287031,
0.0330162756,
-0.0009699259,
0.0179308187,
0.0410857759,
0.0308385585,
0.0225471575,
-0.0212389193,
0.0175869595,
-0.0044511752,
0.0501099825,
-0.0019314672,
0.0256357659,
0.0331203602,
-0.0206944719,
-0.0479011349,
-0.034108419,
-0.0434723347,
0.0047257575,
-0.0073672086,
0.0487526543,
0.021984756,
-0.0035026802,
-0.0294823088,
-0.0029810155,
-0.0655471459,
-0.0235712361,
0.033401031,
0.0341044143,
0.0034645451,
-0.0230629221,
-0.0014195245,
-0.0166488867,
0.0392869674,
0.0088761086,
0.0185955148,
0.0470291227,
0.0020791024,
-0.0102081895,
-0.0672094077,
0.0065988922,
0.0318481028,
-0.0344809964,
0.0008776097,
0.0437970981,
0.0608043596,
0.0475178286,
-0.0415049344,
0.0065639168,
-0.0219027083,
0.0154398484,
0.001220963,
-0.0069170734,
0.0124061387,
0.0064457804,
-0.0025581985,
0.0305799246,
0.0762978196,
0.0514528193,
-0.005702504,
0.0294297673,
0.0045683263,
0.0249171052,
0.0068924781,
0.0085272333,
-0.0396437272,
-0.011127702,
0.0141101228,
-0.0645186156,
-0.0109038353,
0.027774021,
3.04e-7,
0.0375339612,
0.0796386749,
0.0119289868,
0.0036823275,
0.0450664945,
0.0314891525,
0.0083306106,
0.0107717039,
-0.0324217491,
-0.0087386565,
0.019271696,
0.012628342,
-0.0169087909,
0.0058049499,
-0.1150605902,
-0.003940274,
-0.0401987508,
0.0207903404,
-0.0259358268,
-0.0204821713,
0.025303077,
0.0697762072,
0.0225218609,
-0.0080747502,
0.0183567405,
-0.0170899425,
0.0530826114,
-0.0037313567,
0.0053715031,
0.0192252994,
0.0059931907,
0.0177496858,
-0.017959971,
0.0197775885,
-0.0359815136,
0.008606392,
-0.0089375768,
0.0925049707,
-0.0298430026,
0.0026276456,
-0.0012525879,
-0.0049209655,
-0.0150543842,
-0.0440392233,
0.0609291829,
-0.0341484472,
-0.0175724681,
-0.0171086714,
-0.018762758,
0.026099205,
-0.023237789,
-0.0210247412,
-0.0013220756,
-0.0073461346,
-0.001353329,
0.0253558327,
-0.023085542,
-0.0754375979,
-0.0273124836,
-0.0077298563,
-0.0048186556,
-0.0090858247,
-0.0815300494,
-0.0328986794,
0.0751332864,
-0.0532007888,
0.0003774859,
2.850182122e-34,
0.004125495,
0.0121015636,
0.0082291514,
0.0026401682,
0.0452166125,
-0.0243043248,
-0.0238457862,
-0.0121346451,
-0.0046470198,
-0.0690468997,
-0.0360919796
] |
http://arxiv.org/abs/2405.15346v1
|
http://arxiv.org/pdf/2405.15346v1
|
BiSup: Bidirectional Quantization Error Suppression for Large Language Models
| 2024-05-24T08:39:27 | 2024-05-24T08:39:27 |
Minghui Zou, Ronghui Guo, Sai Zhang, Xiaowang Zhang, Zhiyong Feng
|
cs.CL
|
cs.CL, cs.AI, cs.LG
|
As the size and context length of Large Language Models (LLMs) grow,
weight-activation quantization has emerged as a crucial technique for efficient
deployment of LLMs. Compared to weight-only quantization, weight-activation
quantization presents greater challenges due to the presence of outliers in
activations. Existing methods have made significant progress by exploring
mixed-precision quantization and outlier suppression. However, these methods
primarily focus on optimizing the results of single matrix multiplication,
neglecting the bidirectional propagation of quantization errors in LLMs.
Specifically, errors accumulate vertically within the same token through
layers, and diffuse horizontally across different tokens due to self-attention
mechanisms. To address this issue, we introduce BiSup, a Bidirectional
quantization error Suppression method. By constructing appropriate optimizable
parameter spaces, BiSup utilizes a small amount of data for quantization-aware
parameter-efficient fine-tuning to suppress the error vertical accumulation.
Besides, BiSup employs prompt mixed-precision quantization strategy, which
preserves high precision for the key-value cache of system prompts, to mitigate
the error horizontal diffusion. Extensive experiments on Llama and Qwen
families demonstrate that BiSup can improve performance over two
state-of-the-art methods (the average WikiText2 perplexity decreases from 13.26
to 9.41 for Atom and from 14.33 to 7.85 for QuaRot under the W3A3-g128
configuration), further facilitating the practical applications of low-bit
weight-activation quantization.
|
BiSup: Bidirectional Quantization Error Suppression for Large Language Models: As the size and context length of Large Language Models (LLMs) grow,
weight-activation quantization has emerged as a crucial technique for efficient
deployment of LLMs. Compared to weight-only quantization, weight-activation
quantization presents greater challenges due to the presence of outliers in
activations. Existing methods have made significant progress by exploring
mixed-precision quantization and outlier suppression. However, these methods
primarily focus on optimizing the results of single matrix multiplication,
neglecting the bidirectional propagation of quantization errors in LLMs.
Specifically, errors accumulate vertically within the same token through
layers, and diffuse horizontally across different tokens due to self-attention
mechanisms. To address this issue, we introduce BiSup, a Bidirectional
quantization error Suppression method. By constructing appropriate optimizable
parameter spaces, BiSup utilizes a small amount of data for quantization-aware
parameter-efficient fine-tuning to suppress the error vertical accumulation.
Besides, BiSup employs prompt mixed-precision quantization strategy, which
preserves high precision for the key-value cache of system prompts, to mitigate
the error horizontal diffusion. Extensive experiments on Llama and Qwen
families demonstrate that BiSup can improve performance over two
state-of-the-art methods (the average WikiText2 perplexity decreases from 13.26
to 9.41 for Atom and from 14.33 to 7.85 for QuaRot under the W3A3-g128
configuration), further facilitating the practical applications of low-bit
weight-activation quantization.
|
[
-0.0142675182,
0.0481624976,
0.0186730493,
0.052145157,
-0.011340674,
0.0113452207,
0.0218527485,
-0.0099055888,
-0.0086791897,
-0.0347104929,
-0.0208185539,
-0.064167805,
-0.0007474151,
0.0122106709,
0.022921931,
0.0452634767,
0.0828546658,
-0.0489526801,
-0.0085585034,
-0.0370498635,
-0.0292980541,
-0.012652738,
0.0120231034,
0.0159608964,
0.0167433489,
-0.0241820849,
-0.0253980383,
0.0342799015,
-0.0013412777,
-0.0426813141,
0.0091237258,
0.0480707809,
-0.0125287613,
0.0521142296,
0.0000021695,
-0.002741588,
-0.0082889926,
0.0020752547,
-0.037585713,
0.0856833979,
-0.0006724991,
0.0144499009,
-0.0730275139,
0.0269134883,
0.0287325606,
-0.0024721869,
0.0851846114,
-0.0192842633,
0.044146277,
0.0364680924,
0.005197966,
0.00557849,
0.0303864796,
-0.0040277527,
-0.0088230558,
-0.0368344635,
0.0333716795,
0.0021469092,
-0.0263386704,
-0.012792917,
-0.0219709594,
0.0250990652,
0.0154080037,
-0.0330310091,
-0.0352397375,
0.0101274261,
-0.0112539371,
-0.0121542532,
-0.0493934192,
0.0460250676,
-0.0747142658,
-0.0416202322,
0.0018612272,
-0.0012468897,
-0.0353172049,
-0.118519932,
-0.022700686,
0.012519083,
0.0280396827,
0.0295715034,
-0.0205506459,
-0.0222942345,
0.0332365148,
-0.0123485522,
-0.0725393146,
0.0747305602,
0.0440927148,
-0.0085960422,
-0.0439467654,
-0.0264244396,
-0.0271950513,
-0.0555210561,
-0.0030199594,
-0.0029792611,
-0.0368853733,
0.0043075578,
-0.0433551781,
-0.032896474,
0.0201004427,
-0.0617532916,
0.0201126207,
0.0739027411,
-0.0288860183,
0.0099969497,
-0.0364002809,
0.0525051653,
-0.022546038,
0.057561826,
-0.0064809495,
0.0145468237,
0.025137065,
0.0057578795,
-0.0355971195,
0.042282898,
-0.0362757668,
-0.024919467,
-0.0495714173,
0.0127629535,
0.0117593249,
0.0058085821,
-0.0108813299,
-0.0001228969,
-0.0289642401,
-0.0144968592,
0.0354560502,
0.0331890471,
0.0070247315,
-0.0222019795,
0.0001012966,
-0.0500017963,
-0.0282974821,
0.0240026824,
0.008051876,
-0.0291580912,
0.0164875574,
0.0393379405,
0.0150988065,
-0.0494772755,
-0.0068851225,
-0.005310256,
-0.0208174158,
-0.0360914655,
-0.0257839933,
-0.0532108769,
-0.024414327,
0.0065017128,
0.0013118896,
-0.0174217038,
0.025700599,
0.0273877978,
-0.0035346441,
0.0546922646,
-0.0344418138,
-0.0103490511,
-0.0143827861,
-0.0033263259,
0.0506646298,
0.0841732472,
0.0267794114,
0.0059952894,
0.0154060507,
0.0073525133,
-0.0095776943,
0.0351538062,
0.0435406528,
-0.0061540692,
-0.0589353926,
-0.0096587259,
-0.0022709148,
0.0200218856,
0.0860434398,
0.090857774,
-0.0382286869,
-0.0049125832,
0.009998342,
0.0362868942,
-0.0026800465,
-0.0002600803,
-0.0180393383,
-0.0021322805,
0.0084835226,
-0.0394721739,
-0.0493997447,
0.0347527415,
0.0012143208,
-0.0007303432,
0.095843263,
0.0696481541,
-0.027846273,
-0.0125120087,
-0.0360709578,
-0.0364216566,
0.0097306771,
-0.0158357769,
0.0270348825,
-0.0247913841,
-0.037105836,
0.0989626348,
0.0263592508,
-0.0712009668,
-0.0251994431,
-0.0824596062,
-0.0278234556,
0.0722038299,
0.0367208123,
0.0276999455,
-0.0024606893,
0.0391631201,
-0.007516522,
0.0098739602,
-0.0252868738,
0.017083358,
0.01334491,
-0.0427341908,
-0.0224955101,
0.038782794,
0.0077917259,
0.0016892643,
0.0098428428,
0.0107084708,
-0.0024706454,
-0.0274066329,
0.0019654545,
-0.0011017371,
0.0367361046,
-0.0488511138,
0.0206505898,
-0.0107566928,
0.0317534767,
0.0285409894,
-0.0015245657,
0.0162774138,
0.0247807261,
-0.0012523924,
0.0068262652,
-0.0412318781,
-0.0330528691,
-0.0170049518,
0.0835329071,
0.0501994155,
0.0205052476,
0.034685418,
-0.0168145783,
-0.0590322986,
0.053491395,
-0.0243516844,
0.0357767828,
-0.0521512106,
0.0386424996,
-0.0145720663,
0.0402038619,
-0.0179925393,
0.0033208954,
-0.0407298431,
0.003898826,
0.014244101,
0.0235628802,
0.0149182109,
-0.015176693,
-0.0419761874,
0.0137063619,
0.0022896524,
0.0525903739,
-0.013060797,
-0.0047483016,
0.0082879346,
-0.0046678106,
-0.095504418,
0.0312155448,
-0.0573551543,
0.0393006243,
0.0383356512,
0.0583676174,
0.0075317537,
-0.0202469788,
0.0075584035,
-0.0176327452,
-0.0157183632,
0.0019467532,
-0.0198560897,
-0.0259768981,
-0.043449007,
-0.0335166343,
-0.0413690619,
-0.0008028126,
0.0140169309,
-0.0628079772,
-0.036860358,
-0.106437467,
-0.017616719,
-0.0194497854,
0.0434704274,
0.0102904141,
-0.052998092,
0.0149000082,
-0.0034457585,
0.0079265693,
0.0185701307,
-0.0228176266,
0.0010897442,
0.068442896,
-0.0013103373,
-0.0360381268,
0.0247113295,
0.0164403077,
-0.0585867912,
-0.0003835302,
0.020714201,
-0.01636854,
0.0691112652,
-0.0397187471,
-0.0196232647,
-0.022826843,
-0.0136368684,
0.0234876182,
-0.0130161121,
-0.0067032427,
-0.0309833214,
-0.0534008145,
0.0473572388,
-0.0123643279,
-0.0334218182,
0.0186931975,
-0.0398151092,
-0.0131145911,
-0.0372846089,
0.0260345414,
0.0164539441,
-0.0026225394,
0.0002196929,
0.0450955331,
-0.0425473899,
-0.0438088961,
-0.0167894606,
-0.0200857334,
-0.0289229751,
-0.0433497131,
0.0258080773,
-0.0122465389,
0.0494374372,
-0.0357006676,
-0.0105696823,
0.0066198632,
0.0357041992,
-0.0159076396,
-0.0251360163,
0.0313952267,
-0.0479108579,
0.0583104119,
0.0584849678,
0.0062115807,
0.0472765006,
-0.006756823,
0.0056339195,
-0.0453720018,
0.006370713,
0.0050994703,
-0.0119740954,
0.0630529001,
0.0112672457,
0.0009503372,
0.0376975611,
0.0410346612,
0.0507989414,
-0.025455188,
-0.0196288526,
0.0612500347,
0.0085486826,
-0.0218945239,
0.0392114818,
0.0819511488,
-0.0217452701,
-0.0122381318,
0.068157196,
0.0446600541,
-0.0568340383,
0.0274457131,
-0.0272881407,
0.0280996691,
0.0165543407,
0.022654416,
-0.1111190766,
-0.0149753159,
0.0376926064,
0.0123803783,
0.021217728,
0.018743895,
0.0039120684,
-0.100977689,
0.0276027061,
0.0539814532,
-0.0045102327,
-0.047730837,
-0.0293216836,
-0.0162668675,
0.0157203674,
0.1100622639,
0.0878899843,
0.0124912038,
-0.0137268277,
0.0176850297,
0.0136576341,
-0.0712386593,
0.0312091913,
0.0001829893,
0.074042812,
0.0488876328,
-0.0125165936,
0.0096000517,
-0.0147529887,
0.0531488508,
-0.0091280444,
-0.0066683749,
0.0129054878,
-0.0341409966,
0.0537715144,
-0.0279179104,
-0.0159737356,
-0.0104338806,
-0.0262012444,
-0.0126907406,
-0.0619898774,
0.0566472262,
-0.0152075579,
-0.0055988473,
0.0111452779,
-0.055106055,
-0.0068796435,
0.017428454,
0.0219508465,
-0.0107198767,
-0.0797925889,
-0.017300915,
-0.0144164357,
-0.0522480123,
-0.0037343006,
-0.0007447221,
0.0090359915,
-0.0691682249,
-0.0289554428,
-0.0037721146,
0.0041408734,
0.0323310383,
-0.0362596735,
-0.015765762,
-0.0079877377,
-0.0055450038,
-0.0533615761,
-0.0200689826,
-0.0452203453,
0.0288716089,
-0.0595131144,
0.0352145806,
-0.0053958241,
0.0456450544,
-0.0188696571,
0.0057391967,
-0.0990449786,
-0.0842068493,
0.0305738356,
0.0243381392,
0.0230537597,
0.0448752716,
-0.0232685655,
-0.0906231925,
0.0664415061,
0.0668050572,
-0.0260001272,
0.0576458052,
0.0181558058,
-0.0293332767,
-0.004557746,
0.015244497,
-0.0514279418,
0.0201329067,
0.0165221747,
0.0553144142,
-0.0416434258,
-0.0346196778,
-0.0087771136,
0.0307849478,
0.0583773889,
-0.014713401,
-0.0021542634,
0.0242658947,
0.0672688484,
-0.0451724567,
0.0359809361,
-0.0245134905,
0.0129078031,
-0.0128550092,
-0.013712815,
-0.0123633379,
-0.0382835902,
0.0185553972,
-0.0333138332,
0.0772542134,
0.0031242429,
-0.0391875617,
0.0524730682,
0.0765019059,
-0.0210569985,
-0.0334274583,
0.020064991,
0.0421419926,
0.0387962982,
-0.041214563,
-0.0231049992,
-0.008510014,
-0.0070363563,
-0.0355146974,
0.0517092049,
-0.0233623218,
-0.0338449888,
-0.0530253313,
-0.0019765627,
0.0784010142,
0.0831596032,
0.0588564314,
-0.0001479741,
-0.0110413739,
-0.0251878947,
0.0324437544,
0.0144846961,
-0.006614916,
0.0251735188,
0.0125660477,
-0.0134889707,
0.0117058204,
0.0167740546,
0.001031774,
0.0292117558,
-0.0154154375,
-0.0027414216,
-0.0055689653,
-0.0195704009,
-0.0592067689,
0.0258880965,
-0.0434162393,
-0.0247483142,
-0.0155007541,
-0.0571689755,
0.0291161723,
0.0550193861,
-0.0032543661,
-0.0292707253,
-0.0013153957,
0.039933715,
0.024113724,
-0.0180106238,
-0.0095665799,
-0.0637557283,
0.0144399935,
-0.0159070399,
0.0258085709,
-6.911809805e-33,
-0.0443980396,
0.0151911937,
0.0031935507,
-0.0404469408,
-0.0110889319,
0.0024274606,
-0.01987621,
0.0167470742,
0.0173823833,
-0.0286456682,
-0.02341347,
0.0165301207,
0.0383390076,
-0.0143597974,
-0.0065966933,
0.0438806415,
0.0642542168,
-0.0070238737,
-0.008931472,
0.0449494831,
0.0207977593,
0.0135402558,
0.0273439437,
0.0076261386,
-0.0062180711,
0.0209542047,
-0.0073401094,
-0.0056168265,
0.0045077209,
0.0268949401,
-0.0407610871,
0.0364763103,
-0.0204727892,
0.0252137892,
-0.0464111306,
-0.0326655395,
-0.0426017046,
-0.012925921,
0.0191374701,
0.031504117,
0.0252447389,
-0.0439736024,
0.053657189,
0.0379886664,
-0.0652547479,
-0.0063645095,
-0.0354897641,
0.0129505089,
-0.0369055606,
0.004713783,
-0.015338026,
0.0174719375,
0.0317128226,
0.0422193967,
0.0296044331,
-0.0543449819,
-0.0395188257,
-0.0156797767,
-0.0949709341,
-0.0153599791,
0.0414912775,
0.0688514262,
0.0573514923,
0.0391521901,
0.0020486608,
0.0295304675,
0.0710142627,
-0.0116662495,
-0.0015527291,
0.03458675,
0.0500109047,
-0.0037730909,
0.0490654446,
0.0205369759,
0.0941545889,
-0.0353333205,
-0.0711383522,
-0.0165614076,
-0.0015731052,
0.0034082413,
0.0256611388,
0.0454160683,
-0.0249385629,
-0.0265045743,
-0.0123767359,
0.0177562349,
-0.0473052822,
-0.0620147064,
-0.0044281548,
-0.0039709057,
-0.0347343236,
0.0176368672,
0.0321905389,
-0.0186341293,
-0.0437482893,
-0.0426133052,
0.0066183181,
0.0278847963,
-0.0125372838,
0.0428880341,
-0.0118033644,
-0.0084184948,
-0.0039045026,
-0.0059415423,
0.0163139589,
-0.0022653907,
-0.0070778783,
0.0700057298,
-0.1047119647,
-0.0025904742,
-0.0183296427,
0.0008229563,
0.0148796961,
0.0085913781,
0.0088910218,
-0.0134221958,
0.0274659786,
0.0254840087,
-0.0148711586,
-0.0161752775,
-0.0098115671,
0.0011756128,
0.0143512478,
0.0487498455,
-0.0545034334,
0.0106952749,
-0.058458779,
0.0441998467,
0.0328522995,
-0.0519798137,
-0.0514578335,
0.0143838003,
3.108e-7,
0.0391462222,
0.0486078337,
0.0495558642,
0.0116686961,
0.0086522754,
-0.016225446,
0.0129174534,
0.0141095538,
-0.0061053238,
0.0015572193,
-0.0092766583,
-0.0006257287,
-0.0100241741,
0.0068220948,
-0.0416083746,
0.0549133159,
-0.046438776,
0.0149052953,
-0.0427104682,
-0.0340992026,
0.0225319397,
0.1014870852,
-0.0256455932,
-0.0160854384,
0.0400181487,
-0.029185243,
0.0251399614,
0.029627271,
0.0182743594,
0.0218999237,
0.0340726525,
0.0109282862,
0.0006536801,
-0.0035182647,
0.011187898,
-0.0085776281,
-0.0495740362,
0.0568697862,
-0.0610778145,
-0.0367199071,
0.0161469243,
-0.0109978942,
-0.0297859237,
-0.0256316997,
0.055555854,
-0.0163264554,
-0.0048356429,
-0.1083691642,
0.0005801933,
-0.0138491923,
-0.0168222319,
-0.0071518025,
0.0128012802,
0.0281930063,
0.0264237914,
0.0092845298,
0.0231154542,
-0.09014263,
-0.0552313104,
0.0116360513,
-0.0194014851,
-0.0552470721,
-0.0492927879,
0.0065884087,
0.0143673345,
-0.0211217534,
0.0221948549,
3.365303262e-34,
0.0648611709,
-0.0436514132,
0.0475816056,
0.0434816852,
-0.0053948541,
-0.0292680375,
-0.0278461426,
0.025636211,
-0.0061566532,
-0.0384278744,
-0.033121638
] |
http://arxiv.org/abs/2405.15334v1
|
http://arxiv.org/pdf/2405.15334v1
|
Detection and Positive Reconstruction of Cognitive Distortion sentences: Mandarin Dataset and Evaluation
| 2024-05-24T08:17:20 | 2024-05-24T08:17:20 |
Shuya Lin, Yuxiong Wang, Jonathan Dong, Shiguang Ni
|
cs.CL
|
cs.CL
|
This research introduces a Positive Reconstruction Framework based on
positive psychology theory. Overcoming negative thoughts can be challenging,
our objective is to address and reframe them through a positive
reinterpretation. To tackle this challenge, a two-fold approach is necessary:
identifying cognitive distortions and suggesting a positively reframed
alternative while preserving the original thought's meaning. Recent studies
have investigated the application of Natural Language Processing (NLP) models
in English for each stage of this process. In this study, we emphasize the
theoretical foundation for the Positive Reconstruction Framework, grounded in
broaden-and-build theory. We provide a shared corpus containing 4001 instances
for detecting cognitive distortions and 1900 instances for positive
reconstruction in Mandarin. Leveraging recent NLP techniques, including
transfer learning, fine-tuning pretrained networks, and prompt engineering, we
demonstrate the effectiveness of automated tools for both tasks. In summary,
our study contributes to multilingual positive reconstruction, highlighting the
effectiveness of NLP in cognitive distortion detection and positive
reconstruction.
|
Detection and Positive Reconstruction of Cognitive Distortion sentences: Mandarin Dataset and Evaluation: This research introduces a Positive Reconstruction Framework based on
positive psychology theory. Overcoming negative thoughts can be challenging,
our objective is to address and reframe them through a positive
reinterpretation. To tackle this challenge, a two-fold approach is necessary:
identifying cognitive distortions and suggesting a positively reframed
alternative while preserving the original thought's meaning. Recent studies
have investigated the application of Natural Language Processing (NLP) models
in English for each stage of this process. In this study, we emphasize the
theoretical foundation for the Positive Reconstruction Framework, grounded in
broaden-and-build theory. We provide a shared corpus containing 4001 instances
for detecting cognitive distortions and 1900 instances for positive
reconstruction in Mandarin. Leveraging recent NLP techniques, including
transfer learning, fine-tuning pretrained networks, and prompt engineering, we
demonstrate the effectiveness of automated tools for both tasks. In summary,
our study contributes to multilingual positive reconstruction, highlighting the
effectiveness of NLP in cognitive distortion detection and positive
reconstruction.
|
[
0.0345912799,
0.0548428819,
-0.0585246868,
0.0476372913,
-0.0510750264,
-0.0087032514,
-0.0677099749,
-0.0276070889,
-0.0023484181,
-0.0468177907,
-0.0246651787,
-0.0262421276,
0.0200361293,
0.0147520136,
0.0239362177,
-0.0735402852,
0.0569251776,
0.0664896965,
0.0218608491,
-0.0006417101,
0.011935778,
0.0009828809,
0.0150606893,
0.0199956186,
-0.0285406411,
0.0113708135,
0.0165514629,
0.0388354249,
0.0367573686,
-0.0323810056,
-0.003071951,
-0.0127979284,
-0.0198224299,
0.0127816536,
0.0000020902,
-0.058862187,
-0.0543816946,
-0.0382770449,
0.0090471627,
-0.0281146187,
0.0668583214,
-0.0319867693,
-0.0316378772,
0.0013307885,
-0.0285107959,
0.0532963648,
0.0152462739,
0.0827787146,
0.1051866636,
0.0702929199,
-0.013179251,
-0.0199271291,
-0.0416574888,
0.0151341604,
0.025044065,
-0.009997746,
0.0618209727,
-0.0122041134,
-0.0171530955,
-0.0123766577,
-0.031870652,
-0.0110324845,
-0.0293264762,
0.0127975177,
-0.0010062788,
-0.0459224135,
-0.042122636,
-0.0155167384,
-0.0184311736,
0.0551064871,
-0.0153566813,
-0.0152637372,
-0.0157952886,
-0.0354503132,
0.0152357817,
-0.0071600545,
-0.0020662413,
0.0251659639,
0.0130207362,
0.0257873535,
-0.0110026095,
-0.0366444997,
-0.002702666,
0.0053100158,
-0.0081089856,
0.0265031233,
-0.0040352824,
-0.0713718385,
0.031252373,
0.019224735,
-0.0631779432,
-0.007975989,
0.036469359,
0.0203242935,
0.011055504,
0.0209440552,
-0.0119824493,
-0.0551980361,
-0.0136078754,
-0.0380062871,
-0.001627828,
0.0279001445,
-0.0443454608,
0.0128131853,
-0.0144710634,
0.0369484574,
-0.0004699647,
0.0210010465,
-0.0483315364,
-0.002677894,
-0.0601547174,
-0.0206331331,
-0.0450054705,
0.0385326631,
0.0301136151,
0.0198775548,
-0.0725333393,
-0.0339527018,
-0.0752993599,
0.003096371,
-0.0459687449,
0.0287812166,
-0.0312028229,
0.0011668028,
-0.0406021439,
0.0494555496,
-0.0182446651,
0.0311175976,
-0.0580208898,
-0.0263745915,
-0.0187125504,
0.0288463458,
0.0054198396,
-0.0184653867,
0.0130939987,
0.0024077806,
-0.0080019403,
-0.0069369967,
-0.0680597872,
-0.0174956042,
-0.0519824661,
-0.0288112946,
0.0065352926,
-0.0068788948,
-0.020429913,
-0.0484418049,
0.0218133517,
-0.0612928458,
0.0269677825,
-0.0005954122,
-0.0416807383,
0.0304615181,
-0.031770885,
-0.0099046379,
-0.0021140592,
0.0093094893,
-0.0445678458,
0.0190609172,
0.0296229739,
0.0163301192,
0.0719731152,
0.0054730503,
0.0126675442,
-0.0014074035,
0.0267808754,
-0.0485677198,
-0.0278494284,
-0.0144929131,
-0.0336402319,
0.0100873234,
-0.0138368718,
0.0064509995,
0.01136804,
0.0314440131,
0.0503641441,
0.0027844144,
-0.0808594301,
-0.0285685547,
0.0087136766,
0.0165908951,
-0.0090449145,
0.0243224073,
-0.0326802805,
-0.010775825,
-0.0494933091,
-0.0104005309,
-0.0493367389,
0.0209584348,
0.0058500632,
-0.0306526665,
-0.0215265285,
-0.0129009858,
0.0442813262,
-0.0175117962,
0.0388214774,
0.0339301787,
-0.0304863937,
0.0052035842,
0.0227090307,
-0.0698913112,
-0.0108052846,
-0.0323347263,
0.005313749,
0.0269604921,
0.0056230887,
-0.0568196699,
-0.0135904513,
-0.0176403467,
-0.0727626532,
0.042163007,
0.0513862744,
-0.0100394394,
-0.0211924259,
0.0044023842,
0.0153867519,
0.0125020752,
0.0292232409,
-0.0168937426,
-0.0093255024,
0.0327739269,
0.0150105916,
0.000671652,
-0.0390889086,
0.0497490801,
0.045652993,
-0.0078811403,
0.0100089107,
0.0578645132,
0.0191667285,
0.0312602706,
-0.0134307193,
0.0096948249,
0.032200601,
-0.0320929587,
-0.0761994049,
0.0676166713,
-0.0607837401,
-0.0521236211,
0.0038798379,
-0.0565103404,
-0.0179312658,
0.0602059662,
-0.0411200747,
-0.0489105433,
0.0539250374,
0.0071407044,
-0.0065849028,
-0.0123188784,
0.062304657,
-0.0183600858,
0.0355847739,
-0.018716896,
0.053818848,
0.0280690268,
0.0266362708,
0.0182428155,
-0.0466428772,
0.0328098685,
-0.0096800001,
-0.0478634201,
-0.0352044143,
-0.0192715079,
-0.018091809,
0.0394219011,
0.0083029019,
-0.047265619,
0.0151769426,
-0.0291332472,
0.0003284183,
-0.0206672028,
-0.0028982987,
0.0525650047,
-0.0279800501,
-0.0086888215,
0.0168608055,
-0.005841766,
0.0486711264,
0.012887218,
-0.003360535,
0.0367714316,
-0.0163284801,
0.0126514984,
-0.0179265626,
0.0297669806,
-0.0414468423,
0.0067796838,
-0.0263951905,
-0.0976713821,
-0.0483019724,
0.0146494824,
-0.0147766052,
0.0110456841,
-0.0101782624,
-0.0527978577,
0.0017364587,
0.0325664245,
0.0582493618,
0.0866363049,
0.0156850722,
0.0078990972,
0.0094434293,
0.010430349,
0.0014203208,
0.009153923,
-0.0040694997,
-0.0307650957,
-0.0082553104,
-0.0770940706,
-0.0000057251,
0.0568369329,
0.0482128002,
-0.0377976373,
-0.0609490015,
0.011381954,
-0.0602733269,
-0.033183109,
-0.0112461234,
-0.0380769186,
-0.0195832979,
0.0229930114,
0.0011708933,
0.0164132249,
0.0620350167,
0.0208092313,
0.0023412241,
-0.0208119154,
0.0420008898,
0.0147241987,
0.0512780659,
0.0272590835,
0.018550666,
0.1038719565,
0.0130185448,
-0.0428464264,
-0.0396321416,
0.0279048923,
-0.0198328588,
-0.0171765462,
-0.0745568499,
-0.0170715991,
-0.0317777656,
-0.0340568349,
-0.0716527402,
-0.0383270308,
0.0171889998,
0.006142647,
0.0392764658,
0.034272261,
-0.0171846244,
-0.0228777844,
0.012654488,
-0.0095117968,
-0.0224219989,
-0.0066535268,
-0.0126714902,
-0.007703925,
0.0386753455,
0.0526311994,
0.0708472878,
0.0149489855,
-0.0108103389,
0.0357792713,
-0.0372934006,
0.0408464558,
0.0452820063,
-0.0032022493,
0.0085174032,
-0.02398628,
0.0088602472,
0.0200259406,
0.0975839198,
-0.0093393642,
-0.0372739136,
0.0772371516,
0.052871149,
-0.0270938333,
0.0291740671,
-0.0031047179,
-0.042662181,
0.0386260897,
-0.0321806222,
-0.079799965,
-0.0279051494,
0.0342280008,
0.0145713938,
-0.0467512533,
-0.032528121,
-0.0243788809,
-0.1582974792,
0.0334579162,
0.012649022,
0.0305548441,
-0.0346619487,
-0.0070712906,
-0.0063867113,
0.0139736282,
0.0280338246,
-0.0118545741,
0.0107438406,
-0.0037603641,
-0.0189726297,
-0.0074815084,
-0.0393930897,
0.0208582226,
0.0058886614,
0.0354667902,
0.0299555752,
0.0669818074,
-0.003483057,
-0.0080409953,
-0.0169562679,
-0.0004422481,
0.0217514969,
0.0213565603,
0.0381760895,
0.0332844742,
-0.0132787973,
-0.0089086229,
0.0283676926,
-0.0342638157,
-0.0267099217,
0.0502180755,
0.0715664402,
0.0150580388,
-0.0252537206,
-0.0155742979,
0.0254360475,
0.0392312817,
0.0999832526,
-0.0205203705,
-0.0214622207,
0.0181667693,
0.0165678486,
-0.0146056656,
-0.0041108192,
-0.0413366221,
0.0672223046,
-0.0169159882,
0.0378794931,
-0.1161471158,
-0.0405193157,
-0.0220681783,
0.0164453909,
0.045686394,
-0.0284628011,
0.0254644901,
0.0134148924,
-0.0478944331,
0.0096290335,
0.0020296441,
0.0592227057,
0.012217951,
-0.0125618679,
0.0026472276,
-0.015175784,
0.0002667475,
-0.0303583983,
-0.0329796523,
-0.0200458039,
0.0163627435,
0.0261994116,
0.0151340486,
0.0364597477,
-0.019527873,
0.0145621793,
-0.0315384679,
0.0387367383,
-0.0323560648,
0.0204324797,
0.0525972098,
0.0366589613,
-0.0061238906,
0.0397159532,
-0.0156341586,
0.0291751698,
0.0581234917,
0.0113304798,
0.0161765423,
-0.0408279933,
0.047880169,
0.0395961925,
0.0510120355,
-0.0084343161,
0.0102189658,
-0.0188119207,
-0.0052284524,
-0.0237109214,
-0.0535447896,
-0.0106773339,
-0.0204244312,
0.0308850091,
0.0059683356,
-0.0408415683,
0.029944513,
-0.0482548326,
-0.0046028504,
-0.0448260866,
-0.0151315862,
0.0206016749,
0.0277894773,
-0.0156594105,
0.0238240529,
-0.0119861849,
0.0050736833,
0.0056052459,
0.0324115381,
-0.0004700263,
0.0647391304,
0.0297238752,
-0.0004077945,
-0.011834166,
0.0054795113,
-0.0005937833,
-0.0066380287,
-0.0156886764,
0.0022586302,
0.0632237047,
-0.007746445,
0.0621466748,
-0.0194633659,
-0.0092397016,
0.0024106938,
0.0286444668,
0.0112862,
-0.0403164476,
0.0332530886,
-0.0219352506,
-0.031411767,
-0.0344239399,
0.0337882265,
-0.0908892229,
0.0298672356,
0.0284971464,
-0.0214533806,
-0.0385577232,
-0.0048352471,
0.0486493856,
0.0185825024,
0.00125337,
0.0527121685,
-0.0769802853,
-0.0611289367,
0.0435918942,
-0.0339298286,
0.040953137,
0.014750639,
0.0089677386,
0.0420802832,
-0.0073387241,
-0.0236617681,
-0.0078462809,
0.0498244576,
-0.0244392231,
0.0160413124,
-0.0288599283,
-6.335779369e-33,
-0.018846117,
-0.0036347099,
-0.055609256,
-0.0130817713,
-0.0603565909,
-0.0343753658,
-0.0147467349,
0.005457547,
-0.0435659997,
0.0395793505,
-0.0238295756,
0.0208574533,
0.0022342273,
0.0285242684,
0.0044437507,
0.0270070489,
0.0626817718,
0.0154915899,
-0.0308659188,
-0.0372188017,
0.0665430501,
0.0337303467,
0.1341551244,
-0.0604611263,
0.0396343209,
0.0303819384,
-0.0318511017,
0.029572567,
0.0350031368,
-0.0191965792,
-0.0110291839,
0.0782230496,
0.0284815617,
-0.0322104692,
-0.0026081877,
-0.0115223164,
-0.0372593105,
-0.0463492237,
0.0361801498,
0.040621236,
-0.0121323084,
-0.0110730836,
0.0250966307,
-0.0150691355,
-0.0026780623,
0.0565198325,
0.0412617512,
-0.0341710299,
0.0573267043,
0.0035790417,
-0.1216626763,
-0.0117227295,
-0.0318216421,
-0.0038393338,
-0.0275762342,
0.0085768653,
-0.0218983199,
0.0295249447,
-0.068543151,
-0.0060996064,
-0.0183047894,
0.0172397438,
-0.0251384769,
-0.0305604655,
0.0165316369,
0.0413205586,
0.0064496449,
0.0289862026,
-0.0155817326,
-0.0029770343,
0.0348668061,
0.0049073026,
0.067203626,
0.1234447658,
0.0580765344,
-0.0790105164,
-0.0649529919,
-0.0123361973,
0.0394661017,
0.0149181895,
-0.0031559721,
0.0121587645,
-0.0069949529,
-0.0241049062,
-0.0185007192,
-0.0329879448,
-0.0072665126,
0.015601892,
0.0164155774,
0.0005573607,
-0.0638046116,
0.007397776,
0.0143716475,
-0.0638737157,
0.0019581704,
-0.0016979775,
-0.0385723561,
-0.0322005711,
0.0183253959,
-0.0015058229,
-0.0813495219,
-0.0192242842,
0.0160045959,
0.0502672419,
0.0284393914,
-0.0140593052,
0.0566735566,
0.0177094638,
-0.0357410572,
0.0181427728,
-0.0140477223,
0.0320918672,
0.007929123,
-0.0093196006,
0.0112187983,
0.0459215306,
-0.0347761326,
0.0031510284,
0.047290273,
-0.0099503603,
-0.0306866188,
0.0341823362,
0.0243691783,
0.0027542321,
-0.0605706535,
-0.0212382562,
0.0145943025,
0.0241145622,
0.0270334631,
-0.0768779442,
-0.0467264168,
-0.0497367792,
2.887e-7,
0.0009639623,
0.0177017823,
-0.0082861828,
0.0459974594,
0.0496784709,
0.0617302582,
-0.0304937791,
0.0163072105,
0.0278123487,
0.0060534067,
-0.002428913,
-0.0132678812,
0.0207755268,
0.0103924572,
-0.0512510277,
0.0421388745,
-0.0001162428,
-0.0321486406,
0.003938579,
0.0309711099,
0.1294686645,
0.0443587638,
0.0052208039,
-0.00758137,
0.0057398998,
-0.0206436161,
0.0292629227,
-0.0149981603,
0.0177802816,
0.0203876086,
0.0259220619,
-0.0272896029,
-0.0112968329,
0.0194412339,
0.0038597016,
0.0235817116,
-0.0120078623,
0.0178408884,
-0.0538224243,
-0.0110172229,
-0.0484424308,
0.007655934,
0.0189718902,
-0.0407067575,
0.0538647398,
-0.0966983363,
-0.0311285947,
-0.0558703765,
-0.052063182,
0.0031381245,
0.0535304211,
0.0200416781,
-0.0291299447,
0.0496775173,
0.0261584949,
0.0778531879,
-0.0373569466,
-0.0770958811,
-0.009191337,
0.0007736183,
0.0076144324,
-0.026430795,
0.0187786706,
0.0338823721,
-0.0255215093,
-0.069762148,
0.0394313447,
2.258898592e-34,
0.011105733,
0.0070989961,
0.0266305432,
-0.0407889411,
0.0348214433,
-0.0223287903,
-0.0198415481,
0.0110963704,
-0.0266760383,
0.0125329401,
-0.017036764
] |
http://arxiv.org/abs/2406.00025v1
|
http://arxiv.org/pdf/2406.00025v1
|
SCALM: Towards Semantic Caching for Automated Chat Services with Large Language Models
| 2024-05-24T08:16:22 | 2024-05-24T08:16:22 |
Jiaxing Li, Chi Xu, Feng Wang, Isaac M von Riedemann, Cong Zhang, Jiangchuan Liu
|
cs.CL
|
cs.CL, cs.AI
|
Large Language Models (LLMs) have become increasingly popular, transforming a
wide range of applications across various domains. However, the real-world
effectiveness of their query cache systems has not been thoroughly
investigated. In this work, we for the first time conducted an analysis on
real-world human-to-LLM interaction data, identifying key challenges in
existing caching solutions for LLM-based chat services. Our findings reveal
that current caching methods fail to leverage semantic connections, leading to
inefficient cache performance and extra token costs. To address these issues,
we propose SCALM, a new cache architecture that emphasizes semantic analysis
and identifies significant cache entries and patterns. We also detail the
implementations of the corresponding cache storage and eviction strategies. Our
evaluations show that SCALM increases cache hit ratios and reduces operational
costs for LLMChat services. Compared with other state-of-the-art solutions in
GPTCache, SCALM shows, on average, a relative increase of 63% in cache hit
ratio and a relative improvement of 77% in tokens savings.
|
SCALM: Towards Semantic Caching for Automated Chat Services with Large Language Models: Large Language Models (LLMs) have become increasingly popular, transforming a
wide range of applications across various domains. However, the real-world
effectiveness of their query cache systems has not been thoroughly
investigated. In this work, we for the first time conducted an analysis on
real-world human-to-LLM interaction data, identifying key challenges in
existing caching solutions for LLM-based chat services. Our findings reveal
that current caching methods fail to leverage semantic connections, leading to
inefficient cache performance and extra token costs. To address these issues,
we propose SCALM, a new cache architecture that emphasizes semantic analysis
and identifies significant cache entries and patterns. We also detail the
implementations of the corresponding cache storage and eviction strategies. Our
evaluations show that SCALM increases cache hit ratios and reduces operational
costs for LLMChat services. Compared with other state-of-the-art solutions in
GPTCache, SCALM shows, on average, a relative increase of 63% in cache hit
ratio and a relative improvement of 77% in tokens savings.
|
[
0.0655859858,
0.0061664619,
-0.0231694393,
0.0517042503,
-0.0285865888,
-0.0028460214,
0.0070740026,
0.0060864072,
0.0765680149,
-0.0402848609,
-0.0390777178,
-0.0726095587,
0.0239043497,
0.0045263963,
0.0049610878,
-0.0044838781,
0.0404406153,
-0.0225634202,
-0.0173437763,
-0.0012846127,
-0.0125114135,
0.0299073253,
0.0381914712,
0.0301313587,
-0.0302626975,
-0.0176964309,
-0.0073067225,
0.0163401887,
-0.0431846604,
-0.031755425,
0.0075213788,
0.0126939751,
-0.012250877,
0.0359930657,
0.0000022731,
-0.0557681732,
-0.0194644481,
0.0143995462,
-0.0538853854,
0.0561170317,
0.0567392744,
0.0394171439,
-0.0550825372,
0.045050092,
0.0115197608,
0.0105954669,
0.0547433123,
-0.0406537466,
0.0394474566,
0.0186998136,
-0.0205465052,
-0.0200100001,
-0.0474851131,
0.0186510794,
0.0370836444,
-0.0581864268,
0.0258994885,
-0.0138653694,
-0.0253316741,
-0.0209550392,
0.0071631526,
0.0245423429,
0.0070123686,
0.0021037546,
-0.0119186835,
0.0367445312,
-0.0715847462,
0.0193132646,
0.0236514527,
-0.0153397452,
0.0507847108,
0.0065368335,
0.0247942973,
0.0353522003,
-0.0379783809,
-0.0042567807,
-0.0254140571,
-0.0114578567,
-0.0288421102,
-0.0238410011,
-0.027758915,
-0.0396056883,
0.0131607521,
0.0114395004,
-0.0341280811,
0.0172048826,
0.0319166109,
0.0080293128,
-0.0573408566,
-0.029545499,
0.0552488938,
-0.0542796664,
0.0396127589,
0.0459868982,
-0.0275970288,
0.0251812097,
-0.0072094901,
0.0632766858,
0.0483986512,
-0.0845952332,
0.0119506037,
0.0587561615,
-0.0081208209,
0.0266964361,
-0.0017648917,
0.046704758,
-0.0069901273,
-0.0544321761,
-0.0292260908,
-0.048363138,
-0.0285801589,
-0.0338492282,
-0.0363151729,
0.0452332124,
0.0229510237,
-0.0279420055,
-0.0805101395,
0.0833201632,
0.0507126376,
0.0252555609,
-0.0058383578,
0.0002215172,
0.0067141531,
0.0536949858,
-0.0378901102,
0.0125467675,
-0.0305126216,
-0.073790729,
-0.0119988658,
-0.1145873964,
0.0254960712,
0.0131945284,
-0.0021181565,
0.0151693374,
0.0079685207,
0.1093098745,
0.0064817234,
-0.0295464229,
-0.0155852092,
-0.0341062211,
0.0498841628,
-0.0375855267,
-0.0280494839,
-0.0651384816,
0.011974006,
0.0256510451,
-0.0117553202,
0.0247045718,
-0.0266825072,
0.0324660949,
-0.0367751494,
-0.0090602655,
-0.0082980571,
0.0001878427,
0.0602083616,
0.0208496526,
0.0353103764,
0.0541247576,
0.0237489007,
0.0099654431,
0.0565685034,
-0.009653681,
-0.0096464306,
0.0201096628,
-0.0037493431,
0.0133036058,
0.008937045,
0.0311002899,
0.0361392684,
0.0277280081,
-0.039283473,
0.0409349427,
-0.0178851131,
-0.051326897,
0.0258412026,
0.0799095184,
-0.0378181748,
0.0126383724,
-0.0139030376,
0.0295667686,
0.0202368312,
-0.0249954034,
-0.0057392539,
0.0483086556,
0.0055430955,
-0.0204651002,
0.0652711913,
0.0340807997,
-0.0580309518,
-0.0230742339,
-0.0137098907,
-0.0153453639,
-0.0031846985,
0.0096511105,
0.0199521054,
-0.0331954807,
-0.0109883407,
0.0833231732,
0.0132152643,
-0.0129881967,
-0.0359754376,
-0.0316427313,
-0.0380491465,
0.1365586072,
0.0626659617,
0.0235699434,
-0.0663174391,
0.058650814,
-0.0267613064,
0.0269260649,
0.0423982255,
0.0273855124,
0.0014343972,
-0.0398503467,
-0.0041734735,
0.0397202373,
0.0060665323,
0.0278523546,
0.043061953,
0.0204878543,
-0.0109798228,
0.0412402526,
0.0075194137,
-0.0216268152,
-0.0255592056,
0.0049749538,
-0.0147360126,
-0.0185880698,
-0.0077078338,
-0.0017903334,
0.0071871569,
-0.0122781564,
-0.0291277468,
0.0057826196,
-0.0098557835,
0.0000937317,
-0.0467085205,
0.0601813979,
0.0340045877,
0.032273192,
0.021214284,
-0.0067166295,
-0.0019775424,
-0.0130076753,
0.0273064356,
-0.0556413457,
0.0918627307,
-0.0096520549,
0.0292405151,
-0.0214286875,
0.0224123076,
0.0178377088,
-0.0035422004,
-0.0078437449,
0.0074675893,
-0.0291207395,
0.0326563045,
0.0247582756,
-0.004282177,
0.0024722731,
0.0576791614,
-0.0053801476,
0.0497054718,
0.0088896975,
0.0245620497,
-0.0621097125,
0.0179166999,
-0.1431188881,
0.0052176779,
0.0006794924,
0.0426939018,
0.0177104995,
0.0596458837,
-0.0125308698,
-0.0239927843,
-0.0298170485,
0.032368131,
0.0038576436,
0.0648022145,
0.0043804902,
-0.0186896194,
-0.0649596676,
-0.0233569536,
-0.0216608178,
0.0171127766,
0.0567571744,
-0.0187013708,
-0.0308341514,
0.0077304719,
0.0341959931,
-0.0785201713,
-0.0211215448,
-0.0236734338,
-0.0416178852,
0.004944257,
-0.022869166,
0.0117907664,
-0.0036585706,
0.0304653328,
-0.0153927496,
-0.0289180651,
-0.0620193034,
-0.0735709593,
-0.0460848771,
-0.0115030361,
-0.0513594151,
-0.0328845568,
-0.0390806608,
0.0277900565,
0.0367492586,
-0.0288567524,
0.0154079581,
0.0127654532,
-0.0238302182,
-0.0333398022,
0.0227185152,
-0.0631351322,
0.0167414807,
-0.0276586097,
-0.0018919082,
0.0303157866,
-0.0593840331,
0.0323422588,
-0.029034799,
-0.0190997571,
-0.0395654254,
-0.0349568389,
-0.0136721171,
0.0312343203,
-0.0370309539,
0.0518653765,
-0.043681547,
-0.0542973541,
-0.0128698545,
-0.0453702398,
-0.004242741,
-0.089550592,
-0.023667464,
-0.0013696862,
0.0189811904,
0.0272175707,
-0.0072682807,
-0.0345095508,
0.0141280815,
-0.0308273062,
-0.0402274393,
0.0111380927,
-0.0482785366,
0.0112829627,
0.0237378217,
-0.0115855392,
0.0366956405,
-0.0088230036,
-0.0220180806,
-0.0088808686,
-0.0323001854,
-0.0004246073,
-0.022958139,
-0.0088791642,
0.0107626142,
0.0260468647,
0.1237935275,
0.0385106169,
-0.0679443628,
0.0239757784,
-0.0109081576,
0.07477507,
-0.0079943361,
-0.0291814283,
0.0285829827,
-0.0180453099,
-0.0502557717,
-0.0340748318,
0.081151247,
0.0451335646,
-0.0637001693,
-0.0167565346,
-0.0068878047,
0.0052651297,
0.0083880816,
0.0584830604,
-0.1154239923,
-0.0075371894,
0.0343748853,
0.0179698486,
-0.0313066579,
-0.0125437323,
-0.0290183313,
-0.0805875659,
0.0107640792,
-0.0143305901,
-0.0703727603,
0.0594956055,
-0.0687097013,
-0.0122872153,
0.0538939163,
0.0454045944,
-0.0047884802,
0.025076203,
-0.0357273743,
-0.0027735599,
-0.0425767042,
0.0096913623,
0.0525935404,
-0.0594255999,
0.0377052277,
0.0337278508,
0.0380283408,
0.0445311144,
-0.0187670831,
0.058527384,
0.0346040986,
-0.0063405531,
-0.0400394052,
0.0089064604,
-0.0058328919,
0.0037461477,
-0.041494891,
-0.0366274714,
-0.0272014905,
0.0083254464,
-0.0074106837,
0.02524212,
-0.059715163,
0.0134035405,
0.0356989317,
0.0100604948,
-0.0087955222,
0.0471708663,
-0.0077688312,
-0.0200900789,
-0.0262456331,
-0.0502350591,
-0.0092384052,
0.0368351005,
-0.0513811335,
0.0030239813,
-0.0328525603,
-0.067168586,
-0.0261587612,
0.015996566,
-0.0452440493,
-0.009505122,
-0.0175058786,
0.0255070962,
0.0061008548,
-0.0025355618,
-0.0173126776,
0.003745808,
-0.0855266228,
0.0191512201,
-0.0505752377,
0.0054380023,
0.0345850512,
0.0083591966,
0.0044241105,
0.0035443776,
-0.0472997278,
-0.0515989661,
0.0263798404,
0.018407451,
0.0260832366,
0.0208780486,
0.0005387034,
-0.0396598876,
0.0035454324,
0.0479944721,
-0.0991567448,
0.0239492729,
-0.0083354656,
0.0295446366,
0.0044717137,
-0.0042696986,
-0.0625109226,
0.0147527233,
-0.022709595,
-0.0258068964,
-0.0042240522,
-0.004122728,
-0.0147275403,
0.0452893227,
-0.0172611699,
0.0294736456,
-0.0410249606,
0.0605090521,
0.0720552877,
-0.0452299416,
-0.0047203177,
0.0268196799,
-0.0133069316,
-0.0332986079,
0.0035020823,
0.0116034066,
-0.0097514279,
0.0214188639,
0.0231356118,
0.0269154925,
-0.0257561635,
-0.0331523605,
-0.0283298008,
0.0328803211,
0.0234183259,
-0.0124099534,
0.0006025342,
0.0195378214,
0.0366650224,
-0.0605695546,
-0.0033002323,
-0.0017216476,
-0.0161223765,
0.0200511441,
0.0277902484,
-0.0517038666,
-0.0733907595,
0.0569257587,
0.0158415139,
0.0417309999,
0.0108813969,
0.030088922,
-0.0409866534,
0.0346678086,
-0.0387817323,
-0.0138203884,
-0.0015594169,
-0.0261214674,
-0.0429745652,
-0.0009394701,
-0.0017432324,
0.0013823827,
0.0334506147,
-0.0246422533,
0.0245657936,
-0.0320465676,
-0.0665836036,
-0.0269640461,
-0.004015699,
0.0079223076,
0.0245861914,
-0.0114386799,
-0.0781823471,
-0.0619887486,
-0.0346027799,
-0.0409618318,
-0.0160469897,
-0.0012189454,
-0.015010138,
0.0533533432,
0.0837212577,
0.0434464626,
0.007057826,
-0.0286628287,
0.018774176,
-0.0703401938,
-0.0097867632,
0.0433614515,
-6.851502536e-33,
0.0132333422,
-0.0427933335,
0.0415718593,
0.0065168384,
-0.0173292998,
0.0001128063,
-0.0355290398,
-0.0397546217,
0.0128041264,
-0.0487851873,
-0.0111031691,
0.0515252687,
0.0482314862,
0.0054721083,
-0.0234329849,
-0.0180419255,
0.0263045095,
0.0295683239,
-0.0069791526,
0.0609498061,
0.098592259,
0.0211424585,
-0.0084781367,
-0.0682157949,
-0.019452313,
-0.0418722704,
0.0781852081,
-0.0269302912,
0.040782854,
0.0053532878,
-0.0372674502,
0.043426156,
0.0021999038,
-0.0304909442,
0.0154712638,
-0.0153987082,
-0.0208972432,
-0.0056097484,
0.0263308771,
0.0441097505,
0.0498811007,
-0.0373768248,
-0.0460581146,
0.0350667275,
-0.0287342574,
0.049048081,
-0.0047167125,
0.00142887,
-0.0236623585,
0.0039765607,
0.013444853,
0.013092163,
0.0071061733,
0.0421441495,
0.0113905501,
0.0131393941,
0.009621039,
0.0331844725,
-0.0439103879,
0.0226017479,
-0.0611084439,
0.0656270385,
0.0323391668,
0.0475398973,
0.0037017383,
0.0358400941,
0.0244973786,
0.0407512523,
0.0112330914,
0.0423001498,
0.0243225563,
0.0129525485,
-0.0242843777,
-0.0281230882,
0.0064903731,
-0.0397316739,
-0.0136289354,
0.0072775129,
0.0286716409,
0.0044066585,
0.0440360382,
0.0004384882,
-0.0364856534,
-0.050965067,
-0.0059859017,
0.0881971195,
-0.0465459488,
-0.0327238217,
0.0093740355,
0.0040631937,
-0.0032581475,
-0.0135678565,
0.0401998758,
-0.0140251871,
-0.018396765,
0.0141219972,
0.0151693104,
-0.0038611747,
0.000085235,
0.0412373245,
0.0196096357,
-0.0305127986,
-0.0346732922,
-0.0330882631,
-0.0040253494,
0.0180471279,
-0.0244828202,
-0.0050479001,
-0.056472376,
0.0319943167,
-0.0353848524,
0.0102720056,
-0.0210031942,
-0.0506897829,
0.0035709324,
0.0152690476,
0.0145277968,
0.0849607512,
0.0426964536,
0.0022839727,
-0.0130465655,
-0.064819172,
0.0042600208,
0.0072537698,
-0.0130338278,
0.0404946692,
-0.0498984605,
0.0029439228,
0.0539599732,
-0.0336651243,
-0.0293346159,
-0.0241559427,
3.025e-7,
-0.0209626146,
0.0616124123,
0.0262449645,
0.0144580808,
0.0163217001,
0.0258954801,
0.0352117568,
0.0177795757,
0.0247524381,
0.0158213768,
-0.0413826965,
0.0300971884,
0.0433615893,
0.0530580021,
-0.0625517219,
-0.0049635349,
-0.0782645196,
-0.0559693649,
-0.0639356747,
0.0312776975,
0.0153651498,
0.0471735448,
0.0078695668,
-0.0422402471,
-0.0081975367,
0.0007879143,
0.0262630619,
-0.0715502277,
-0.0267786961,
0.0113435732,
0.0228906255,
0.0357521512,
0.0327936858,
-0.0063787494,
-0.0055305636,
-0.0381134637,
-0.0466873758,
0.0018372013,
0.0318134874,
-0.0103526264,
0.03765608,
-0.0467393845,
-0.0176746305,
-0.022297509,
0.0459425747,
0.0094066067,
-0.0327852555,
0.0157625787,
0.0261907261,
-0.0425114557,
-0.0116403643,
-0.0038452884,
-0.0392733254,
0.0596889332,
-0.0052803331,
-0.0194699243,
0.0263457205,
-0.0287831947,
-0.0183780603,
-0.0175185166,
-0.0343954302,
-0.0459414087,
-0.0373153314,
0.0219961032,
0.000672201,
0.0000148881,
0.0196971912,
2.718072169e-34,
0.0204658322,
0.0461574607,
-0.0041556638,
0.052088093,
0.0477382615,
0.037296202,
0.0357084684,
0.0381164774,
0.0221929532,
-0.0347733982,
-0.004252797
] |
http://arxiv.org/abs/2405.15329v2
|
http://arxiv.org/pdf/2405.15329v2
|
Decompose and Aggregate: A Step-by-Step Interpretable Evaluation Framework
| 2024-05-24T08:12:30 | 2024-06-14T08:32:19 |
Minzhi Li, Zhengyuan Liu, Shumin Deng, Shafiq Joty, Nancy F. Chen, Min-Yen Kan
|
cs.CL
|
cs.CL
|
The acceleration of Large Language Models (LLMs) research has opened up new
possibilities for evaluating generated texts. They serve as scalable and
economical evaluators, but the question of how reliable these evaluators are
has emerged as a crucial research question. Prior research efforts in the
meta-evaluation of LLMs as judges limit the prompting of an LLM to a single use
to obtain a final evaluation decision. They then compute the agreement between
LLMs' outputs and human labels. This lacks interpretability in understanding
the evaluation capability of LLMs. In light of this challenge, we propose
Decompose and Aggregate, which breaks down the evaluation process into
different stages based on pedagogical practices. Our experiments illustrate
that it not only provides a more interpretable window for how well LLMs
evaluate, but also leads to improvements up to 39.6% for different LLMs on a
variety of meta-evaluation benchmarks.
|
Decompose and Aggregate: A Step-by-Step Interpretable Evaluation Framework: The acceleration of Large Language Models (LLMs) research has opened up new
possibilities for evaluating generated texts. They serve as scalable and
economical evaluators, but the question of how reliable these evaluators are
has emerged as a crucial research question. Prior research efforts in the
meta-evaluation of LLMs as judges limit the prompting of an LLM to a single use
to obtain a final evaluation decision. They then compute the agreement between
LLMs' outputs and human labels. This lacks interpretability in understanding
the evaluation capability of LLMs. In light of this challenge, we propose
Decompose and Aggregate, which breaks down the evaluation process into
different stages based on pedagogical practices. Our experiments illustrate
that it not only provides a more interpretable window for how well LLMs
evaluate, but also leads to improvements up to 39.6% for different LLMs on a
variety of meta-evaluation benchmarks.
|
[
0.0361446738,
0.0424007699,
-0.0153678833,
0.0330277458,
-0.0383668244,
0.0255032424,
0.037309505,
-0.0196585115,
-0.0349667892,
-0.065839991,
-0.0499699675,
-0.0224416256,
0.0316088907,
0.0207422394,
0.0150809186,
-0.0490562841,
0.0123861646,
-0.035881184,
0.0062026302,
0.01233296,
0.0131807756,
0.0140998606,
0.0222785398,
0.0591246597,
0.0562513508,
-0.0170563869,
0.0140116801,
-0.0094962362,
0.0114957364,
-0.0156067694,
-0.0434166528,
0.0517937951,
0.0031918648,
0.0904053375,
0.0000021967,
-0.0689820647,
-0.0078329016,
0.0278528091,
-0.0212249272,
0.0884759352,
0.0791668668,
0.0009290866,
-0.0359519087,
0.0093824044,
-0.0031393664,
-0.0531338304,
0.0172741357,
-0.030523397,
0.0592803657,
0.031077832,
0.0074863583,
-0.041455172,
0.0246819425,
-0.0158323925,
-0.0215809084,
-0.0013110838,
0.0150518091,
0.0314314477,
0.0139128361,
0.0287556313,
-0.0173928682,
0.0305871293,
0.0392555371,
-0.053936217,
0.0193941984,
0.0103073614,
-0.0411944203,
-0.0206216071,
0.0025194385,
0.0530429445,
0.0333798192,
-0.0218796339,
-0.0133668585,
-0.0193275232,
0.0017397477,
0.031808503,
-0.0576720163,
0.0441933163,
0.0225091558,
0.0342978984,
0.0035146198,
0.0336095318,
0.0418978706,
0.0259377547,
-0.040561229,
0.0565303974,
0.021452697,
-0.0247612502,
0.0121887419,
0.0508815572,
-0.0308293533,
-0.0349783935,
0.0169844292,
0.005756184,
0.0056564501,
0.0326166898,
-0.0170647427,
-0.0464420728,
0.0369136743,
-0.0414256193,
0.0564641133,
-0.0027252287,
0.0026833692,
0.0347748436,
-0.0015187901,
0.0625538528,
0.0084362188,
0.0056250021,
-0.0537974276,
0.0274744779,
-0.0658961609,
-0.0055979821,
-0.0103862137,
0.0005547923,
-0.027432004,
0.0150120072,
-0.0542063117,
-0.0078395531,
-0.0057394174,
-0.0108372187,
0.0002215318,
-0.0422576442,
0.010486464,
-0.0147682726,
-0.0162595455,
0.0493805893,
-0.0072167753,
0.0309491958,
-0.0108202193,
-0.0395424068,
0.0270089786,
0.0305101685,
-0.0064363154,
-0.0128119038,
-0.0053585242,
0.0361753777,
-0.0145647433,
-0.0134933619,
-0.0137242712,
-0.024148019,
-0.0399971679,
-0.0698809922,
-0.0139570618,
-0.0875165835,
0.0129759749,
0.0267432332,
0.0108204931,
-0.0180079192,
-0.0047427537,
0.0013175581,
-0.0083865449,
-0.0037714639,
-0.0220670272,
0.0364357568,
0.0056048683,
0.0233232304,
0.0104980385,
0.0612676814,
0.0635355264,
0.0442294404,
0.0035559295,
-0.0198415034,
0.0118710399,
0.0462514795,
0.0166616049,
0.034776628,
-0.0121823922,
-0.0160940532,
-0.0226219147,
0.0319208801,
-0.0006496613,
0.083514221,
-0.0359390266,
0.0570830628,
0.0592092089,
0.039196521,
0.0142609468,
0.009121961,
0.0704945102,
0.0219850261,
0.0575478077,
0.0028227214,
-0.0270250626,
-0.0188293047,
-0.0287775882,
0.0202118959,
0.0307821557,
0.0478956886,
0.0059282663,
-0.0716082826,
-0.0097411452,
0.0203388445,
-0.0057752295,
0.0036953087,
0.0733075738,
0.0029043483,
-0.017928943,
0.0009298989,
0.0187082775,
-0.0707452372,
0.0522916131,
-0.0555995964,
0.0304871537,
0.0748150796,
0.0935489088,
0.0258324929,
0.0188439358,
-0.0043032072,
0.0093164258,
0.0550999567,
0.0113961501,
-0.0138312988,
0.0048353891,
-0.056305103,
-0.0230775736,
-0.0078067733,
0.0156682376,
0.0147741698,
-0.0129820872,
-0.0330147184,
0.0244580694,
-0.0744186714,
-0.0907465219,
0.029917486,
-0.0373795219,
-0.0203255564,
-0.0403049476,
-0.0187874548,
0.0191479158,
0.0331183597,
-0.0357178226,
0.0266141631,
-0.000870074,
-0.0252657812,
0.0066572023,
0.0120802429,
-0.0225347616,
0.0013875391,
0.0162369888,
-0.0571254492,
-0.0240620002,
-0.0081779929,
-0.0192358922,
-0.0441405699,
0.0232299995,
-0.0461718552,
0.0373006649,
-0.0205397829,
0.0349013805,
-0.0500651784,
-0.0229433682,
-0.0479173474,
0.0476792492,
-0.0064248354,
0.0111444695,
0.0308283959,
-0.0210552569,
0.0341435857,
0.0205766708,
-0.0546003655,
0.0188087728,
0.0033377758,
0.0211256649,
0.0238479543,
0.0224404093,
-0.0382319354,
-0.0753752962,
-0.0620027073,
-0.0095655974,
-0.0156209385,
0.039293617,
0.0324832648,
-0.0403480865,
0.0042191232,
0.060182564,
-0.0061616208,
-0.0443720296,
0.0174224619,
0.0044893296,
-0.0030162374,
0.0051416773,
-0.0120416461,
-0.0513821431,
-0.0318949819,
-0.0038461471,
0.0984223261,
0.0625321865,
0.0369683765,
-0.030641593,
0.0391198844,
-0.0162206553,
-0.0628620163,
0.0184793565,
-0.0494856834,
0.024286529,
-0.0062754867,
0.0541664734,
0.0104637817,
-0.0144761121,
0.0328185856,
0.0719219446,
-0.0122395679,
-0.0717647001,
-0.0567618161,
0.0151136871,
-0.0312444381,
-0.0584909096,
-0.0902778506,
-0.0148955481,
0.0667495057,
0.0167492293,
-0.0067826058,
-0.037443202,
0.0223647058,
-0.0817252323,
-0.0630470514,
-0.051703684,
-0.0315607451,
0.0058620088,
0.0411303379,
-0.0253984276,
-0.0709951892,
0.0605790056,
-0.0461384952,
-0.0045068543,
-0.0458143614,
-0.0034919197,
-0.0546655059,
0.0648459792,
-0.0060427031,
0.0010965219,
-0.0267921705,
-0.0047322791,
-0.0283716582,
-0.0012456384,
0.0154471584,
-0.0560044795,
0.0784432292,
-0.0684880912,
-0.0091883959,
-0.0040692235,
0.0193459,
-0.0081295781,
-0.0139282634,
-0.0186921488,
-0.0062143616,
0.0595564544,
0.017160099,
-0.0454897508,
0.00856754,
0.0492923893,
-0.0135092679,
-0.0257736836,
-0.0169967413,
-0.0112047214,
0.0003678064,
0.0178040527,
-0.0133660035,
0.0380210057,
-0.0764912516,
0.0431645699,
0.0486969538,
0.017524574,
0.0138418917,
-0.0554852597,
-0.0094836904,
0.0332825445,
-0.0007420038,
-0.0087970328,
0.0218701456,
0.0365667343,
-0.0712269396,
0.0074585825,
0.1213937253,
0.0337267593,
-0.0381115414,
0.0019199569,
-0.021946298,
0.0241754223,
-0.0145199578,
0.0067756278,
-0.0400900096,
0.0002163179,
-0.003779029,
0.0335376821,
-0.0113539184,
0.0003162446,
-0.0024110493,
-0.1068894714,
0.0408950895,
0.030226713,
-0.0192117039,
0.0013465422,
-0.0350205638,
-0.0221393295,
-0.0167378038,
0.0169371311,
0.0477796011,
-0.0344752781,
0.0315950215,
0.0361322127,
-0.0255938694,
-0.0542107597,
0.0062643872,
-0.0177113917,
0.0526428409,
0.0411046594,
0.0171223879,
-0.0356998779,
-0.0117948046,
0.0295918696,
-0.0256221779,
-0.0197836291,
-0.0019650958,
0.0354766436,
0.0383048207,
0.0150543349,
-0.0603991039,
-0.0037255681,
-0.0157193374,
-0.0365193561,
0.0187509824,
0.0779650062,
0.009886804,
-0.0182882044,
-0.0304973964,
0.020405801,
-0.0286040399,
0.0802851468,
0.0608223379,
-0.0499208309,
-0.0180568509,
-0.0221181642,
-0.0216532741,
-0.0416015498,
-0.0662128255,
0.0501283668,
-0.0092918379,
-0.0149043957,
-0.0384560563,
-0.0518906303,
0.0083904127,
0.0164900068,
0.0237810947,
0.0131972637,
0.0072869887,
-0.0176469702,
-0.0088053886,
0.0465797819,
-0.0190008152,
0.0081932517,
-0.0801179335,
-0.0034430518,
0.0172389355,
-0.0331914797,
-0.0150016211,
0.0334634595,
-0.0662704557,
-0.0606988929,
0.0004759959,
0.0151802022,
0.0292213,
-0.0211645309,
0.0059627169,
-0.0752593428,
0.0408584923,
-0.0082195699,
-0.0148987817,
0.0456111096,
0.0210948251,
0.0059829094,
-0.004067116,
-0.0009572456,
-0.0269064792,
0.0455377214,
0.0222245827,
0.0066646328,
-0.0093107354,
-0.0371276364,
0.0141377226,
0.0469470285,
0.0077621136,
-0.0499058068,
0.0032767465,
-0.0351322033,
-0.0210719276,
-0.0575434305,
-0.0047936961,
-0.0601112172,
-0.008530559,
0.0482890718,
-0.0084839649,
-0.0178974848,
0.0368001387,
-0.0063615297,
0.0189221166,
-0.0373766758,
0.0668682083,
-0.0273386575,
0.0158436652,
0.0073037497,
-0.0091241188,
-0.0237550084,
-0.0269489698,
-0.0550767928,
0.0195556413,
0.0066961413,
0.0318137817,
-0.010837811,
-0.0006471102,
0.0294137206,
0.0488353446,
-0.0578330308,
-0.0079137068,
-0.0104502086,
0.0032654938,
0.0209238585,
0.0435217544,
0.1104583666,
-0.0196774565,
0.0208505113,
-0.0484621674,
0.0382850096,
0.0083778463,
-0.0216622669,
0.0039002209,
0.0126338899,
-0.0183094107,
-0.0194088351,
0.0671941787,
0.0211953335,
-0.0132542988,
0.0044018365,
0.0022501424,
-0.013675658,
-0.0735734329,
-0.0448549539,
0.0002458572,
-0.0213115718,
-0.0256765801,
-0.0901353881,
-0.0375550799,
-0.0270397123,
0.0714321136,
0.0189220067,
-0.0299910009,
-0.0115772812,
0.0600107796,
-0.0231194254,
-0.0069227819,
-0.0518102832,
0.0212291051,
-0.0032565999,
-0.011491986,
0.0025850669,
-6.470665142e-33,
-0.0271409191,
-0.0261435658,
0.0273586139,
-0.0356710628,
-0.0304583497,
0.0214217566,
-0.0513571911,
0.0098628215,
0.0174949616,
0.0313378274,
0.0188512485,
-0.0101732081,
0.0051624863,
0.0341166332,
0.0232604351,
0.0105739683,
0.0379371718,
0.0143547533,
-0.0239648279,
0.0062846062,
0.0234488733,
-0.0170901157,
0.0361662097,
-0.0602749772,
0.0516080298,
-0.0119815255,
-0.0342227146,
-0.0508722439,
-0.0254270323,
0.0797315016,
-0.0121378312,
0.039174024,
0.0325008146,
-0.0453811362,
-0.0364904106,
-0.0049719885,
-0.0332081243,
0.0020356192,
0.0658032969,
0.0032973047,
-0.0380213931,
-0.0345716998,
0.0634359568,
-0.001907437,
-0.0419922657,
-0.0140159642,
0.0358825736,
-0.0518007651,
-0.0338018611,
-0.0151862521,
-0.0629485622,
0.0548026636,
-0.0201319419,
0.0056268312,
0.0277438927,
0.0316876508,
-0.0287443474,
0.0472056568,
-0.0312563181,
0.0332053043,
0.0442929342,
0.0419865362,
0.0669243485,
0.0131692989,
-0.0096501913,
0.0589225963,
-0.0000913569,
0.0000246865,
-0.0206747204,
-0.0090092896,
0.0668602511,
0.0147838444,
0.0532574356,
0.0878171846,
0.0026404981,
-0.0019899125,
-0.0635295436,
0.0350799561,
-0.0104199899,
-0.0069095977,
0.0378107056,
-0.0111290487,
-0.044724375,
-0.0440948084,
-0.0714698806,
-0.0125101442,
-0.0174593143,
-0.0103691481,
0.0383118279,
-0.0277658999,
-0.0471013151,
-0.0370830819,
0.0386065468,
-0.0478290878,
-0.0166148208,
0.0282106884,
-0.0225217473,
0.0855306163,
-0.0118876891,
0.0272682048,
-0.0559447557,
0.0011846031,
0.0260999817,
0.0232745782,
0.0219658986,
-0.0053442842,
0.0127742495,
0.0007707178,
-0.0374991819,
0.0173053555,
-0.0433229804,
-0.0113638798,
0.0109816734,
-0.0427556597,
0.0228961408,
0.0005722425,
-0.0116384896,
0.013335566,
0.0150022348,
0.016235942,
-0.0130346762,
0.0054387827,
0.0051860427,
-0.0045763515,
-0.0302004293,
0.0131215099,
-0.0188103877,
-0.0068389196,
0.0725715011,
-0.0208208729,
-0.0054761679,
0.0069508282,
2.798e-7,
0.0741097927,
0.0557711795,
0.006504734,
-0.006116116,
0.0239500981,
0.0222801734,
-0.0299166813,
-0.013739177,
0.0463213585,
0.0026856048,
0.0257214084,
0.0249155629,
0.0829108283,
0.0293043386,
-0.0129399002,
0.0218875688,
-0.0638451427,
-0.0005835538,
-0.0519778505,
0.0150031867,
0.0487703606,
0.0630953535,
0.0044366843,
0.0085289469,
-0.0266900603,
-0.0358822607,
0.0531766973,
-0.0557972938,
-0.0295039341,
0.0542673655,
0.0269316435,
0.0051559675,
0.0084222835,
-0.0211587101,
-0.005554169,
0.0089099277,
0.0135220615,
0.0914119631,
0.0141794374,
0.037694227,
0.0266696718,
-0.0503306165,
-0.0020825092,
-0.0536927879,
0.0526405275,
-0.0337190256,
-0.0372515433,
-0.0223875921,
-0.0485357791,
-0.0167928338,
0.0053045726,
-0.0091037713,
0.0057793618,
0.0343274139,
-0.0013253973,
-0.0385595821,
-0.0178834163,
-0.0275159162,
-0.0369057059,
-0.0007849283,
-0.0228034314,
-0.0797010362,
-0.0413650274,
-0.0159803741,
0.0046349508,
-0.0899349302,
0.0050271191,
2.546571447e-34,
0.0590041168,
-0.0348448455,
0.0417274162,
-0.0244746879,
0.0330435,
0.0202475041,
-0.0334445722,
0.00699049,
-0.0050615072,
-0.0930573046,
0.0136278775
] |
http://arxiv.org/abs/2405.15320v1
|
http://arxiv.org/pdf/2405.15320v1
|
Organic Data-Driven Approach for Turkish Grammatical Error Correction and LLMs
| 2024-05-24T08:00:24 | 2024-05-24T08:00:24 |
Asım Ersoy, Olcay Taner Yıldız
|
cs.CL
|
cs.CL, cs.AI
|
Grammatical Error Correction has seen significant progress with the recent
advancements in deep learning. As those methods require huge amounts of data,
synthetic datasets are being built to fill this gap. Unfortunately, synthetic
datasets are not organic enough in some cases and even require clean data to
start with. Furthermore, most of the work that has been done is focused mostly
on English. In this work, we introduce a new organic data-driven approach,
clean insertions, to build parallel Turkish Grammatical Error Correction
datasets from any organic data, and to clean the data used for training Large
Language Models. We achieve state-of-the-art results on two Turkish Grammatical
Error Correction test sets out of the three publicly available ones. We also
show the effectiveness of our method on the training losses of training
language models.
|
Organic Data-Driven Approach for Turkish Grammatical Error Correction and LLMs: Grammatical Error Correction has seen significant progress with the recent
advancements in deep learning. As those methods require huge amounts of data,
synthetic datasets are being built to fill this gap. Unfortunately, synthetic
datasets are not organic enough in some cases and even require clean data to
start with. Furthermore, most of the work that has been done is focused mostly
on English. In this work, we introduce a new organic data-driven approach,
clean insertions, to build parallel Turkish Grammatical Error Correction
datasets from any organic data, and to clean the data used for training Large
Language Models. We achieve state-of-the-art results on two Turkish Grammatical
Error Correction test sets out of the three publicly available ones. We also
show the effectiveness of our method on the training losses of training
language models.
|
[
0.0422527269,
0.0904575735,
0.0014300791,
0.0235752743,
-0.0319162272,
0.0245744865,
0.0183904171,
-0.0496049672,
-0.018518582,
-0.0523800664,
-0.0339154303,
-0.040051084,
0.0308440737,
-0.0016464399,
0.0380965024,
-0.0015786826,
0.0703003854,
0.0110876048,
-0.0353510827,
0.0006722112,
-0.0199398603,
0.0127828838,
-0.0288634896,
0.0003152909,
-0.0468461476,
-0.0374540761,
-0.025700558,
0.0177710764,
0.0133599788,
0.0260036178,
0.0575976744,
0.0437452532,
-0.0236126445,
0.089753449,
0.000001638,
-0.0232551321,
-0.0116136158,
-0.0008401285,
0.0043000686,
0.0304470845,
0.0372055024,
-0.0275460444,
-0.0878818408,
0.0154580623,
-0.0183943808,
-0.0222733244,
0.0495592803,
0.0159080774,
0.0670346692,
0.108177945,
-0.0091732154,
-0.016958693,
0.0726659968,
0.0020065422,
-0.0132037634,
0.036749728,
0.0342498943,
-0.0332523026,
-0.0903700888,
-0.0283651426,
0.002313131,
0.0486529656,
0.0077613089,
-0.0381570458,
-0.0420191921,
-0.0157860294,
-0.045267649,
-0.0348052271,
-0.0142083205,
0.0388952196,
-0.020842284,
-0.0194822457,
-0.0058925957,
0.0144578638,
0.023520533,
-0.0195364822,
-0.0277156979,
0.0246223714,
0.0053906688,
0.0361004323,
0.0669620782,
-0.0832626969,
0.037094947,
0.0031471031,
-0.0000059933,
0.0398181826,
0.051904168,
-0.0682892352,
-0.0111302715,
0.0370878652,
0.0202973671,
-0.0084033711,
0.0241532139,
0.02199292,
-0.0015200622,
-0.0057785287,
-0.0579378642,
-0.07145302,
0.0193619169,
-0.0896896422,
0.0784185454,
-0.0038651708,
-0.0404904895,
0.0082507012,
-0.049426809,
0.003714839,
0.0485665649,
0.0300572794,
0.0081915101,
0.003918292,
-0.028135445,
0.0123269139,
-0.0386613384,
0.0191344228,
-0.0441080816,
0.0560505278,
-0.0448134132,
0.00271106,
-0.041659683,
0.0383244455,
0.0371986069,
0.0015829309,
-0.0626047552,
-0.0034653309,
-0.0074559171,
0.0120220743,
0.0114194062,
0.0489652194,
-0.0086040469,
-0.0845553502,
0.0166186001,
-0.0024259784,
0.0507843718,
0.0108409496,
0.0129817268,
0.0142583903,
-0.0207881685,
-0.0335793197,
-0.0614278726,
0.0161367916,
-0.0107930871,
-0.0368286669,
-0.0374402702,
-0.0415984429,
0.0087419376,
-0.0516261384,
-0.003282713,
-0.0321152881,
0.0015218862,
0.0412505008,
-0.001571612,
0.0887387395,
-0.0769537166,
0.0298384391,
0.0045250976,
-0.0475799926,
0.0611649267,
0.0347371958,
0.0341258869,
-0.0405765101,
0.0352782905,
-0.0105131762,
0.0285303202,
0.0151826702,
-0.0251219459,
-0.0252956506,
-0.0065494445,
-0.0181304831,
-0.0324638784,
0.0596996471,
0.0272388812,
0.0726870373,
0.0048435633,
0.0431774445,
0.0669649392,
0.0080751013,
0.0069047459,
0.0416620448,
0.0271192044,
-0.0347624868,
0.016151797,
-0.0043807621,
-0.0585378259,
-0.0285489466,
0.00020526,
0.0106892437,
-0.0021694917,
0.036633648,
-0.0150594274,
-0.0616791323,
0.0020200473,
0.0049045533,
-0.004551128,
-0.039426513,
0.0551803261,
0.0399132445,
-0.0344704501,
0.0325759724,
0.0276106223,
-0.0794930831,
0.0024912003,
0.0041973707,
0.0120391548,
0.0483133607,
0.0364687108,
-0.0241818167,
-0.0514866784,
-0.0303625036,
0.0076379944,
0.0279016662,
0.0618917979,
-0.0374562182,
0.0092276484,
-0.0567896515,
0.0202802718,
0.0087317778,
0.0408623964,
-0.0071097463,
-0.0036959383,
0.0115713198,
0.0240289737,
-0.0020853083,
-0.0431385189,
0.0471694097,
-0.0158230495,
0.0054207882,
0.0035373524,
0.0212418586,
0.0036878809,
0.0308796968,
-0.015092697,
-0.0197243392,
0.019185042,
-0.0300098453,
-0.0147757567,
-0.0188871492,
-0.0098851863,
-0.0484772474,
0.0561891943,
-0.0423611924,
-0.0653640702,
0.0794048831,
-0.0583728962,
-0.042151019,
0.0090916464,
-0.0240755975,
0.0020135744,
-0.0557211228,
0.0479561165,
0.0226568729,
0.0045191268,
-0.0257093534,
0.0020992255,
0.0002553407,
0.0183152035,
-0.011704566,
-0.0013249816,
0.0120341238,
-0.0063710767,
-0.0101363594,
0.0072228806,
-0.0425890088,
0.0385486446,
-0.0153703764,
0.0155490898,
-0.0353184715,
-0.0179551728,
-0.072446391,
-0.0112470603,
-0.0438972674,
0.040911682,
0.035300184,
0.006998667,
-0.0128493933,
-0.0168445949,
-0.0119472751,
0.0187929291,
0.0204157736,
-0.0013372488,
0.0774451122,
-0.0472868755,
-0.0393153504,
-0.0468542017,
-0.0131680528,
-0.0281544905,
-0.0213392507,
-0.0109521793,
-0.0293049533,
-0.0404249392,
0.0188216362,
-0.009035971,
0.0518250912,
0.0380686708,
-0.0794330537,
-0.0119355749,
0.0108267758,
0.0524734482,
0.024029864,
-0.0356283002,
-0.0061087469,
0.0320225954,
-0.025037637,
-0.0100882789,
-0.0022453263,
0.0074312063,
-0.0156217711,
-0.0155803906,
-0.011344675,
0.0029268565,
0.0464832,
-0.0333241671,
-0.044057373,
0.0016955875,
-0.0005805954,
-0.0203615967,
-0.000851234,
0.0061076232,
0.0339509808,
-0.0285363439,
0.0219619907,
-0.0344442278,
-0.0101324636,
0.0276167057,
0.0019271208,
0.0229998473,
-0.0339934789,
-0.0150775295,
-0.0167500991,
-0.0353418328,
-0.0018261805,
-0.0005551409,
-0.0265245307,
-0.0385043435,
-0.06843853,
0.0204127003,
0.0397130065,
-0.0133591257,
0.0777647346,
-0.0774683431,
0.0284199901,
-0.0127425725,
-0.0531624816,
-0.033630155,
-0.0472315624,
0.0049996544,
-0.0278187543,
0.0007345089,
-0.0100606643,
-0.0417483971,
-0.0637958348,
-0.0300528239,
0.0869187638,
0.0022894274,
0.0033246956,
-0.0305317733,
0.0071388357,
0.0022252509,
0.0019929216,
0.0004394676,
0.0192288961,
0.0310542993,
0.0689756796,
0.0000763237,
0.0132523067,
-0.0393776521,
0.0108486656,
0.0195146184,
-0.0434256047,
-0.0244188495,
-0.0366331153,
0.0206929371,
-0.0442101434,
-0.0403046943,
0.0654839799,
0.028151162,
-0.0136892935,
-0.0045326557,
-0.0013299556,
-0.0290685836,
0.0386106111,
0.0223762188,
-0.0700668469,
-0.0182981864,
0.0536922924,
-0.0062579005,
-0.0154480971,
0.0099781761,
0.0228391364,
-0.1423342079,
0.0206052493,
0.024972178,
-0.0496743657,
-0.0742100477,
-0.0054427339,
0.0003255538,
-0.0183725338,
0.0291819237,
0.0387949236,
-0.0303993076,
-0.0052782507,
0.039064128,
-0.0627119169,
-0.1087144911,
-0.0694451332,
-0.0297206137,
0.0671532974,
0.028884897,
0.0280975886,
-0.0116931247,
-0.0028463746,
0.0667778701,
0.0124684041,
-0.0386104845,
0.0275113955,
0.0485897325,
0.0192846432,
-0.0338319167,
-0.0038544224,
0.0375204608,
-0.0971488878,
-0.0019601134,
-0.0109369066,
0.0743682235,
0.0077219722,
0.0187113546,
0.0217281356,
-0.0019332517,
0.0136721563,
0.02602553,
0.0212059561,
-0.0163231473,
-0.0120561486,
-0.0368957818,
-0.0431065708,
-0.0696362108,
0.029284237,
0.0526851453,
-0.0508124158,
-0.0294167232,
-0.0915675163,
-0.0835412145,
0.0416295156,
-0.0086752642,
0.0384600759,
-0.0191905405,
-0.0354868583,
-0.0042183376,
-0.0211779643,
0.0128553333,
-0.0035099192,
0.022126466,
-0.0563196018,
0.0097530959,
0.0090580033,
0.0431965403,
0.006943367,
0.01845867,
-0.0646837726,
-0.0470934361,
0.0634455681,
0.0277097244,
0.0078948895,
0.0255796909,
0.0018123674,
-0.0435946994,
0.0081872586,
0.0450383835,
-0.0285742003,
0.0080737825,
0.0444673151,
-0.0205811374,
-0.0005003797,
0.0372695066,
0.0187628511,
0.0019398684,
0.0237858668,
0.0135695944,
-0.0262688994,
0.0265424307,
0.048487626,
0.0217049588,
0.0588102005,
0.013064689,
0.010220346,
-0.0731053054,
0.0549801551,
-0.0118606398,
-0.0228645951,
-0.0143473269,
0.0374149829,
0.0203157272,
0.0033296961,
-0.0345474072,
-0.0093047256,
-0.0208223183,
0.006288101,
-0.0401436761,
-0.0436847024,
0.022015186,
0.0492172055,
0.0205339305,
0.0284911152,
-0.022537265,
0.0542338341,
0.0056846258,
0.0207682159,
-0.0138871334,
-0.0289855283,
-0.0002477988,
0.0416907184,
-0.0073785558,
0.0163281467,
-0.0204667728,
0.0202180315,
-0.0485628881,
0.0427678861,
0.1015583053,
0.0014555785,
0.0522934496,
-0.0274018589,
-0.0313941762,
-0.0165871792,
0.0356480479,
-0.0126686841,
0.0061826799,
0.01312315,
-0.0246077441,
-0.0001020811,
-0.0420743898,
0.0360421911,
-0.0007005064,
0.0639181286,
-0.0220165551,
0.0183405988,
-0.0237960089,
-0.0010922004,
0.0148574682,
-0.0069743101,
-0.0211532712,
-0.0523802564,
-0.0038472898,
-0.1096449196,
-0.0204530954,
0.066495575,
0.0242049173,
-0.0041424874,
-0.0002986594,
0.0293118823,
0.0096573057,
0.0242928807,
-0.0323178098,
0.0057949494,
0.0193701144,
-0.0425818451,
-0.0413872264,
-5.148967212e-33,
-0.0025301126,
0.0662330091,
-0.0088492082,
-0.027663786,
-0.0162049346,
-0.0316729471,
-0.0094330087,
-0.0089440141,
0.0273647308,
0.069271937,
-0.0339362323,
0.0439357162,
0.0253859777,
0.0254471153,
0.0342363119,
-0.0065885526,
0.0639336705,
-0.0198920686,
0.0280893929,
-0.0309128799,
0.037063498,
0.0563564636,
0.1054179892,
-0.0110629015,
0.0066013895,
0.0166762527,
-0.0203105081,
0.0386948287,
-0.022312697,
0.0274460465,
0.0345546827,
0.0669968501,
0.0114158858,
0.029023014,
-0.0433005318,
-0.0067633823,
-0.082974799,
-0.0223772787,
-0.0155765973,
0.074989818,
-0.0376859419,
-0.0480585918,
0.0476541482,
-0.0297673959,
0.0084588025,
0.0045772004,
0.0156118413,
-0.0531332232,
-0.0007807962,
-0.0178814847,
-0.0471127033,
0.0221236032,
-0.0164104998,
0.0029051115,
0.0015666846,
0.0327909961,
0.0119180921,
-0.05879163,
-0.0345423892,
0.0108668907,
-0.0087788338,
0.0546601675,
0.0354615785,
-0.0530679412,
-0.0604113713,
0.0342691019,
0.021282563,
0.0264594164,
0.0122500155,
0.0037198206,
0.0048875911,
0.0310649704,
0.028290676,
0.0808906257,
0.0601585954,
-0.0572085381,
-0.0579340607,
0.0408230387,
-0.0121963052,
0.0177448355,
-0.0483807474,
-0.0115194172,
-0.0010164896,
-0.0196783319,
-0.0158908553,
-0.0069229743,
-0.0086647496,
0.0141259395,
0.0334831476,
-0.0008552917,
-0.0178138595,
0.0136385588,
-0.0325826667,
-0.0429464281,
-0.0102650402,
-0.0321043544,
0.010517275,
0.038671758,
-0.0341271646,
-0.0016485406,
-0.0169841535,
-0.0651187077,
-0.0161718316,
0.0408536494,
0.0216766261,
0.0069062966,
0.058088772,
0.0330960602,
-0.0902101547,
0.011138415,
-0.0197317861,
0.0374595821,
0.0350795984,
0.0420502312,
0.0153601021,
-0.0137757696,
0.0181899797,
0.0189979505,
0.0261438955,
0.0052682073,
0.0015268472,
-0.0310967155,
-0.0128522236,
-0.0020676509,
0.0084121907,
-0.0033164537,
-0.0471460521,
-0.0355380066,
0.0483570285,
-0.0673786998,
0.018883964,
-0.0253089927,
2.379e-7,
0.002797859,
0.0384747684,
0.0255411789,
-0.0075034639,
-0.0071247621,
-0.0072270827,
-0.0212538708,
0.0385695398,
-0.0208433215,
0.0400097035,
0.0497228689,
-0.0230751745,
0.0815430507,
0.0027419603,
-0.0739354193,
-0.0299684219,
-0.055697564,
0.0288099796,
-0.036948774,
-0.0136669399,
-0.0116039813,
0.0821457803,
0.0188692454,
-0.0450217947,
-0.0226701982,
0.0089089787,
0.0283948202,
-0.0251851194,
-0.0389462076,
0.0323339701,
0.0757483318,
0.0079786191,
0.0236168541,
0.0144205829,
-0.0584846251,
-0.0031949354,
0.0191351622,
0.0416304022,
0.0174751058,
0.0542169921,
0.0049405568,
-0.0326117985,
-0.0262643024,
-0.011888626,
0.0492069125,
-0.036223866,
0.0111691831,
-0.0211182367,
-0.0419921763,
-0.0039713448,
0.0152229341,
-0.0226341691,
-0.006326823,
0.0393731706,
0.0604123808,
0.0032967716,
-0.0341533013,
-0.037591584,
0.0033030277,
0.0055032372,
0.0154791772,
-0.0147258602,
0.0186709426,
0.0045087961,
0.025462918,
-0.0236295387,
-0.01122561,
1.927033807e-34,
0.0370907821,
0.015552355,
0.0556778535,
-0.006212492,
0.0338752083,
-0.0183842517,
0.0390681922,
0.019667834,
-0.016054811,
-0.0288195424,
-0.0071065058
] |
http://arxiv.org/abs/2405.15319v1
|
http://arxiv.org/pdf/2405.15319v1
|
Stacking Your Transformers: A Closer Look at Model Growth for Efficient LLM Pre-Training
| 2024-05-24T08:00:00 | 2024-05-24T08:00:00 |
Wenyu Du, Tongxu Luo, Zihan Qiu, Zeyu Huang, Yikang Shen, Reynold Cheng, Yike Guo, Jie Fu
|
cs.CL
|
cs.CL, cs.AI
|
LLMs are computationally expensive to pre-train due to their large scale.
Model growth emerges as a promising approach by leveraging smaller models to
accelerate the training of larger ones. However, the viability of these model
growth methods in efficient LLM pre-training remains underexplored. This work
identifies three critical $\underline{\textit{O}}$bstacles: ($\textit{O}$1)
lack of comprehensive evaluation, ($\textit{O}$2) untested viability for
scaling, and ($\textit{O}$3) lack of empirical guidelines. To tackle
$\textit{O}$1, we summarize existing approaches into four atomic growth
operators and systematically evaluate them in a standardized LLM pre-training
setting. Our findings reveal that a depthwise stacking operator, called
$G_{\text{stack}}$, exhibits remarkable acceleration in training, leading to
decreased loss and improved overall performance on eight standard NLP
benchmarks compared to strong baselines. Motivated by these promising results,
we conduct extensive experiments to delve deeper into $G_{\text{stack}}$ to
address $\textit{O}$2 and $\textit{O}$3. For $\textit{O}$2 (untested
scalability), our study shows that $G_{\text{stack}}$ is scalable and
consistently performs well, with experiments up to 7B LLMs after growth and
pre-training LLMs with 750B tokens. For example, compared to a conventionally
trained 7B model using 300B tokens, our $G_{\text{stack}}$ model converges to
the same loss with 194B tokens, resulting in a 54.6\% speedup. We further
address $\textit{O}$3 (lack of empirical guidelines) by formalizing guidelines
to determine growth timing and growth factor for $G_{\text{stack}}$, making it
practical in general LLM pre-training. We also provide in-depth discussions and
comprehensive ablation studies of $G_{\text{stack}}$. Our code and pre-trained
model are available at
$\href{https://llm-stacking.github.io/}{https://llm-stacking.github.io/}$.
|
Stacking Your Transformers: A Closer Look at Model Growth for Efficient LLM Pre-Training: LLMs are computationally expensive to pre-train due to their large scale.
Model growth emerges as a promising approach by leveraging smaller models to
accelerate the training of larger ones. However, the viability of these model
growth methods in efficient LLM pre-training remains underexplored. This work
identifies three critical $\underline{\textit{O}}$bstacles: ($\textit{O}$1)
lack of comprehensive evaluation, ($\textit{O}$2) untested viability for
scaling, and ($\textit{O}$3) lack of empirical guidelines. To tackle
$\textit{O}$1, we summarize existing approaches into four atomic growth
operators and systematically evaluate them in a standardized LLM pre-training
setting. Our findings reveal that a depthwise stacking operator, called
$G_{\text{stack}}$, exhibits remarkable acceleration in training, leading to
decreased loss and improved overall performance on eight standard NLP
benchmarks compared to strong baselines. Motivated by these promising results,
we conduct extensive experiments to delve deeper into $G_{\text{stack}}$ to
address $\textit{O}$2 and $\textit{O}$3. For $\textit{O}$2 (untested
scalability), our study shows that $G_{\text{stack}}$ is scalable and
consistently performs well, with experiments up to 7B LLMs after growth and
pre-training LLMs with 750B tokens. For example, compared to a conventionally
trained 7B model using 300B tokens, our $G_{\text{stack}}$ model converges to
the same loss with 194B tokens, resulting in a 54.6\% speedup. We further
address $\textit{O}$3 (lack of empirical guidelines) by formalizing guidelines
to determine growth timing and growth factor for $G_{\text{stack}}$, making it
practical in general LLM pre-training. We also provide in-depth discussions and
comprehensive ablation studies of $G_{\text{stack}}$. Our code and pre-trained
model are available at
$\href{https://llm-stacking.github.io/}{https://llm-stacking.github.io/}$.
|
[
0.0260741152,
0.0773781836,
0.0181766804,
0.0109776091,
-0.0089866985,
-0.0007298127,
0.001111508,
0.0258945748,
-0.0267116223,
-0.0465927906,
-0.0606511198,
-0.0285144057,
-0.0452122279,
0.0140371565,
0.0205210578,
-0.0131406039,
0.0408733152,
-0.0343381874,
-0.0591626167,
-0.0246664863,
-0.0287782364,
-0.0094293766,
-0.0302760564,
0.0173696429,
-0.0195882265,
0.0095485598,
-0.0018488317,
0.0195108596,
-0.0019445079,
-0.0428309962,
-0.0177559145,
0.0262645483,
0.0087614907,
0.0425747074,
0.0000020747,
-0.062608704,
-0.0052937572,
0.0163227934,
-0.0519582815,
0.0335203074,
0.004343925,
-0.0393985286,
-0.0182442851,
0.0377714373,
-0.0207764469,
-0.0167587567,
0.0326832831,
0.0804769024,
0.0375727713,
0.0730555728,
-0.0054007098,
-0.0588702932,
0.0756770596,
-0.0440194607,
-0.0125521524,
-0.0130910398,
0.02749197,
-0.0318257511,
-0.0347916745,
0.0362655371,
0.0029647856,
0.0102619622,
0.0191035382,
0.0036264872,
0.0267133005,
0.0746479854,
-0.0238625277,
-0.0283903629,
-0.0458588302,
0.0332206674,
-0.0284870062,
-0.0219065994,
-0.0150388554,
-0.0126186786,
-0.0237135086,
-0.0105391238,
-0.0543801151,
0.038325157,
0.0256826542,
0.021929333,
0.0076038796,
-0.0013438251,
0.0531982519,
0.0192165691,
-0.0270590559,
0.0597861409,
0.0256288256,
-0.0365917757,
0.0353405252,
0.0131241791,
-0.0350034125,
-0.0432945713,
-0.0031952499,
0.0470174551,
0.0069314805,
0.042357754,
-0.0019567332,
-0.0043440894,
0.0529560372,
-0.0442866199,
0.0629581362,
0.0100068282,
-0.0235216971,
0.0344435759,
-0.0639067143,
0.0097300438,
-0.0473865941,
0.0513020791,
-0.0290208794,
0.0123565644,
-0.0213119704,
-0.0097615449,
-0.0664423779,
0.0667973012,
-0.0208923984,
-0.023225816,
-0.0766945854,
0.0010545504,
0.0110173468,
-0.0070355022,
0.0308670402,
-0.0015657669,
0.011409618,
0.0384720117,
-0.0282731913,
0.0328700468,
-0.0321768038,
0.0054742265,
0.0088984333,
-0.0569127947,
-0.0425133631,
0.0104779545,
-0.0192344021,
-0.0154762501,
0.0496604778,
0.0808915943,
0.0077407425,
-0.0449082814,
-0.0641993582,
0.0043910663,
-0.0141460244,
-0.0604087114,
0.00940107,
-0.0494399518,
0.0491267741,
0.000905768,
0.012713884,
0.0004123435,
0.0061534015,
0.0279606804,
-0.0163602363,
0.0608792976,
-0.0306918863,
-0.008718797,
-0.0086979093,
0.008625567,
-0.0140915355,
0.0836797357,
0.0211868156,
0.0305284448,
0.0346865468,
-0.029593708,
-0.0202577896,
0.0360179059,
0.0407848209,
0.03134517,
0.0002949596,
-0.0218055844,
-0.008613823,
0.0118667725,
0.0055387565,
0.0524261594,
-0.0485100895,
0.0042422777,
0.0119070858,
0.0524486639,
0.0438164808,
0.0625698343,
0.022770755,
-0.0275993217,
0.0189877879,
-0.0018875684,
-0.0528827943,
0.001024538,
-0.0303912275,
0.0024352507,
0.0070965611,
0.0375624411,
-0.0342888758,
-0.0458686799,
-0.0524168983,
-0.0002393936,
-0.0442647971,
-0.0259963386,
0.0401715003,
0.0339941792,
-0.0358579755,
0.142772615,
0.0465491526,
-0.082745105,
0.0008969488,
-0.0586460605,
-0.0412522741,
-0.0088732922,
0.0153025072,
-0.0249884073,
-0.0569703355,
0.0159734152,
-0.0415524468,
0.0313315652,
0.0179096106,
0.0056149522,
0.0275698081,
-0.0433747582,
-0.0061280923,
-0.0013536072,
0.0132918116,
-0.0154499663,
0.0409494154,
0.0035042746,
0.0138781518,
-0.0477898382,
0.0263275076,
0.0491557978,
-0.0561252385,
-0.0094100814,
0.0066089761,
-0.0079363994,
0.0471565351,
0.0244385581,
-0.0057721972,
0.0116642518,
0.0700288638,
-0.0435527824,
-0.0652372092,
0.0152801275,
-0.0651074275,
-0.0012004095,
0.0484685935,
0.0416474491,
0.0311459545,
0.0143714612,
-0.0470426828,
-0.0413972773,
-0.0035042749,
-0.0254274607,
0.0804711059,
-0.0460705385,
0.037858434,
-0.0084651951,
0.0239634737,
-0.0500661768,
0.0822246,
-0.0416302755,
0.0015824677,
-0.0124530131,
0.0223562773,
0.0321762972,
0.0016144696,
-0.0233812742,
0.0137908794,
-0.0134905074,
-0.0029913331,
-0.0114115952,
-0.0080217822,
-0.0592776276,
-0.0448325761,
-0.082191281,
0.0195126794,
-0.028726982,
0.0451280586,
0.0140500469,
0.0291772261,
0.0176345594,
0.0097002098,
-0.0018490868,
-0.0265989024,
0.002205702,
0.0233813673,
0.0162743535,
0.011684875,
-0.0457957685,
-0.0073654,
-0.0430410951,
-0.0164269209,
0.0371049941,
-0.0083215544,
-0.0086137131,
-0.0399292558,
0.0305740498,
-0.0226076208,
0.0208978243,
-0.0305082146,
-0.0334921107,
-0.0072942609,
-0.0100281322,
0.0305565745,
0.0091535188,
0.0216710921,
0.0333573706,
0.0247547533,
0.0408598334,
-0.0706648976,
0.0153941838,
-0.0020806957,
-0.0302056801,
-0.0524067655,
0.0226408746,
0.0305144992,
0.0918957517,
-0.0473817922,
-0.0025689469,
-0.0338554308,
-0.0298897773,
-0.0101638269,
-0.034879975,
0.0209495202,
-0.0086090807,
-0.0313148089,
0.0168054681,
0.0168953314,
-0.04943946,
0.0170172825,
-0.0298828986,
-0.002440274,
-0.0217154212,
-0.0216522701,
-0.0888094157,
0.0618153326,
0.0144226104,
-0.0052662254,
-0.001578952,
-0.0413047038,
-0.0144334668,
-0.0020940891,
-0.0161185861,
-0.00950683,
0.0878815576,
-0.016287405,
0.0429018438,
-0.0460916571,
-0.0656010062,
-0.0624616258,
-0.0198808126,
-0.0118594468,
0.0317748077,
0.0389323756,
-0.0184310861,
-0.0673688203,
-0.0182526391,
0.0005290933,
0.0176733956,
0.0025211058,
-0.0108336601,
-0.006909519,
0.0354279093,
0.0160521735,
0.0076750359,
0.0772839338,
-0.0422427803,
0.0355219282,
0.0589758344,
0.0450221039,
-0.0076310472,
-0.047794722,
-0.0012874618,
0.0622848012,
-0.0616357736,
-0.0201719664,
-0.0166563243,
0.0541151203,
-0.0216303635,
0.0075757243,
0.1024910063,
0.0034462733,
-0.0289981402,
0.0356118158,
0.0156535711,
0.0492916182,
0.0344584249,
0.0241693482,
-0.0752936304,
-0.0091153448,
0.0073182932,
-0.0050938758,
-0.0055955681,
-0.0450961404,
-0.0192989018,
-0.1210898757,
0.0200917181,
-0.0053520463,
-0.0031480726,
0.0634507239,
-0.052671358,
0.0025270751,
-0.0162752643,
0.0118960794,
0.009102622,
-0.0748378113,
-0.0410226472,
0.0442856364,
-0.0146400556,
-0.0057914308,
-0.0018272579,
-0.0497321449,
-0.0006929507,
-0.0007317665,
-0.0110153416,
-0.027230965,
-0.0537336692,
0.0887734368,
-0.0002411735,
0.0319745019,
-0.0512319505,
0.0425601788,
-0.0068867025,
0.0004902337,
0.0121236937,
0.0117361816,
-0.0196267311,
-0.017239131,
-0.006356563,
0.0475009568,
0.0085284924,
-0.0246307421,
-0.0174958929,
0.0011211461,
-0.0339523815,
0.0529130101,
0.0032313084,
-0.0158710126,
-0.0426891036,
0.0066791703,
-0.001153998,
-0.0364292189,
-0.0129014738,
0.0386033729,
0.0082234349,
0.0105363494,
-0.0374323875,
0.0200480726,
-0.0148153296,
0.050036896,
-0.0044176723,
0.0064324546,
-0.0246175341,
0.0207536183,
-0.0182245746,
0.0343404971,
-0.011199032,
-0.0030318284,
-0.0994569138,
0.0023743685,
0.0843172148,
0.0114016598,
-0.0185635276,
-0.0410823077,
-0.1281495839,
-0.0648218989,
0.0084180748,
0.0375165865,
0.0225073323,
0.031538669,
0.0371168889,
-0.0407140441,
0.0784387961,
0.052102305,
-0.0348165259,
0.0236545876,
0.0383188017,
-0.0163003653,
0.004117962,
0.012210221,
-0.0181334876,
0.053527873,
0.0184200704,
0.0551037863,
-0.0176808443,
-0.0115987044,
0.0007438961,
0.0463210233,
0.0401287079,
-0.0336898677,
0.0243962873,
0.0018082866,
0.0523673967,
-0.0157852694,
0.0277263634,
-0.0401382074,
-0.0058849938,
0.0015937799,
-0.0132332556,
-0.0108819623,
-0.0066541857,
0.0236589946,
0.0512656011,
-0.0154996039,
0.0467351973,
-0.0328602083,
-0.0618579313,
0.0494406968,
0.0208332613,
-0.0299829207,
-0.0100300619,
0.0295206197,
0.0272542555,
-0.0209202878,
0.017200917,
0.0140640335,
0.0283075869,
-0.0253926702,
0.0496742465,
-0.0590207763,
-0.0581595413,
-0.0086609647,
0.002393936,
0.0524481349,
0.0023104094,
0.0363411382,
-0.0270432439,
-0.0114848511,
0.0198088959,
0.0072513153,
-0.0287611224,
-0.0171551742,
0.0346539915,
0.0195817202,
-0.0023395668,
0.0027027577,
0.0584417097,
0.0110845817,
0.0349516198,
0.0099186413,
0.0572151169,
-0.0215444583,
-0.0367486887,
-0.0145491734,
0.0197070222,
0.0177494064,
0.0012009772,
-0.0660264194,
-0.0543072522,
0.0236332845,
-0.0143480273,
0.0140317502,
-0.0414840877,
0.0494476557,
0.0485549755,
0.0105519891,
-0.037274316,
-0.0015265782,
-0.0252595879,
0.0221936107,
0.0022187119,
0.0150816869,
-6.282209887e-33,
-0.0118348384,
-0.0113914851,
0.0398036875,
0.0235725213,
-0.0603936538,
0.021594787,
0.0127660958,
-0.0090729725,
-0.039655067,
-0.0202360693,
-0.0516020544,
0.0013613673,
0.0154983681,
0.0345350616,
-0.005515316,
-0.0159184281,
0.0844263211,
-0.0059455298,
-0.0065665408,
0.0060337167,
-0.0140928775,
0.02372385,
0.0611673705,
-0.0370776951,
-0.0344479606,
0.0399070606,
-0.0216380786,
0.0357946716,
-0.0414941534,
0.0512203351,
-0.0058796708,
0.0439138934,
0.0291846283,
-0.0676855817,
0.0081333444,
-0.0160172135,
-0.0922262892,
-0.019037392,
0.0089239459,
0.0182234421,
-0.0389540866,
-0.0006355058,
0.0357374288,
0.0186224226,
-0.0441624187,
-0.0283740237,
-0.0064552925,
0.0003499443,
-0.0498573035,
-0.0385628715,
-0.0188712087,
0.028057998,
0.0563160777,
0.0242303964,
-0.0079128873,
-0.0203268789,
0.0027107089,
0.0292195324,
-0.0356733538,
-0.009273936,
0.033828713,
0.0194840524,
0.032866668,
0.0350541957,
0.0031811055,
0.0626125261,
0.0224322025,
0.0190044567,
-0.0149392802,
-0.0579128899,
0.0242870674,
0.0871327668,
0.0251017772,
-0.0311059933,
0.0706111863,
0.0029253382,
-0.0868182853,
-0.0052797552,
0.0369828679,
0.0639816374,
-0.0161897372,
0.0076151951,
0.0028621473,
-0.0040974882,
-0.0335251056,
0.0329128653,
-0.0155307157,
-0.0311982092,
-0.0010278914,
0.044830326,
-0.0174944811,
0.000500174,
0.0058675567,
-0.0384355672,
-0.0424952246,
0.0362375937,
0.0418913178,
0.0235999152,
-0.0214029625,
-0.0087033967,
-0.0543677211,
-0.0415400006,
-0.024915399,
0.0090886988,
0.0229441859,
-0.0147514902,
0.0502843373,
0.0012899566,
-0.0931137055,
-0.0015189288,
-0.031967327,
0.0443436019,
0.0306416247,
0.0341218077,
0.0285315923,
0.0004246099,
0.0029346037,
0.0397413298,
0.0182870273,
-0.0036432275,
-0.0220315307,
0.0213896781,
0.0362249203,
0.0447361954,
0.002716565,
0.0070335707,
-0.0260525998,
-0.0053417021,
0.0576893426,
-0.0838182047,
-0.0380523615,
0.0179656036,
2.73e-7,
0.0328034572,
0.0395023562,
0.0425756015,
0.0308054388,
0.0221230723,
0.0189923458,
0.0000700736,
-0.002578022,
-0.0120084649,
0.0035537605,
-0.0082790134,
0.0132061141,
0.0309456643,
0.0160500444,
-0.0752483457,
0.0034664606,
-0.104226619,
0.0087925242,
-0.036037121,
-0.0140147796,
0.0037401952,
0.0825131088,
-0.0014921279,
-0.057633847,
-0.0289119594,
-0.0234558955,
0.075285919,
-0.0172504745,
0.0106697027,
0.0327175409,
0.0055048373,
0.026587164,
0.0155878775,
0.0209608395,
-0.0357244499,
0.0414569452,
-0.0415948592,
0.0464872494,
-0.0058626095,
-0.028404817,
0.0290199816,
-0.0154138906,
-0.0037729803,
-0.0360662639,
0.0699336752,
-0.0505028069,
-0.0147020984,
-0.0541976467,
-0.0370528065,
-0.0175567996,
-0.0150225079,
-0.014694578,
-0.0102810124,
0.0520579629,
0.0159516465,
-0.0017826437,
-0.0240773447,
-0.0983335376,
-0.0236942954,
0.0094285421,
-0.0299188718,
-0.0451708771,
-0.029503474,
-0.0199179333,
-0.0206685569,
-0.0486078374,
-0.0537273027,
2.830223284e-34,
0.0538319983,
-0.0345657393,
0.0528781787,
0.0481289737,
0.0008686327,
-0.0034787296,
0.021165261,
0.0043943156,
-0.0067806905,
-0.0720841885,
0.0048621129
] |
http://arxiv.org/abs/2405.15318v1
|
http://arxiv.org/pdf/2405.15318v1
|
Are Long-LLMs A Necessity For Long-Context Tasks?
| 2024-05-24T07:59:30 | 2024-05-24T07:59:30 |
Hongjin Qian, Zheng Liu, Peitian Zhang, Kelong Mao, Yujia Zhou, Xu Chen, Zhicheng Dou
|
cs.CL
|
cs.CL, cs.AI
|
The learning and deployment of long-LLMs remains a challenging problem
despite recent progresses. In this work, we argue that the long-LLMs are not a
necessity to solve long-context tasks, as common long-context tasks are
short-context solvable, i.e. they can be solved by purely working with oracle
short-contexts within the long-context tasks' inputs. On top of this argument,
we propose a framework called LC-Boost (Long-Context Bootstrapper), which
enables a short-LLM to address the long-context tasks in a bootstrapping
manner. In our framework, the short-LLM prompts itself to reason for two
critical decisions: 1) how to access to the appropriate part of context within
the input, 2) how to make effective use of the accessed context. By adaptively
accessing and utilizing the context based on the presented tasks, LC-Boost can
serve as a general framework to handle diversified long-context processing
problems. We comprehensively evaluate different types of tasks from popular
long-context benchmarks, where LC-Boost is able to achieve a substantially
improved performance with a much smaller consumption of resource.
|
Are Long-LLMs A Necessity For Long-Context Tasks?: The learning and deployment of long-LLMs remains a challenging problem
despite recent progresses. In this work, we argue that the long-LLMs are not a
necessity to solve long-context tasks, as common long-context tasks are
short-context solvable, i.e. they can be solved by purely working with oracle
short-contexts within the long-context tasks' inputs. On top of this argument,
we propose a framework called LC-Boost (Long-Context Bootstrapper), which
enables a short-LLM to address the long-context tasks in a bootstrapping
manner. In our framework, the short-LLM prompts itself to reason for two
critical decisions: 1) how to access to the appropriate part of context within
the input, 2) how to make effective use of the accessed context. By adaptively
accessing and utilizing the context based on the presented tasks, LC-Boost can
serve as a general framework to handle diversified long-context processing
problems. We comprehensively evaluate different types of tasks from popular
long-context benchmarks, where LC-Boost is able to achieve a substantially
improved performance with a much smaller consumption of resource.
|
[
0.0055061714,
0.0210705027,
-0.0062443381,
0.0527396388,
-0.0659764037,
0.0132776164,
-0.0020813576,
0.0040905531,
-0.0074048261,
-0.0449166484,
0.001911622,
-0.0495838784,
-0.0120071396,
0.0057945126,
0.01268184,
-0.0122444099,
0.0317008346,
-0.0158539992,
-0.031982027,
-0.0067348541,
-0.0578936487,
-0.0417219996,
0.0105526913,
-0.0086362287,
-0.0204052627,
-0.0304770265,
-0.0315152705,
0.0225327369,
0.0102867102,
-0.0184042137,
-0.0096609853,
-0.0323330723,
0.0260045007,
0.0020532375,
0.0000020191,
-0.0364114307,
-0.0144204153,
-0.0120339692,
-0.0393306129,
-0.0167912133,
0.0258303992,
0.0039902302,
-0.0303155929,
-0.0039676894,
-0.0284531638,
0.053529121,
0.0251389891,
0.0503650419,
0.0385436416,
0.0682675093,
-0.0138381924,
0.0088676373,
-0.0085503506,
-0.0162849221,
0.0531046353,
-0.0780094489,
0.0229519438,
-0.0388183445,
0.04258167,
-0.0228281207,
-0.0346541889,
0.0163430311,
0.0372964293,
0.0252631344,
-0.0058815749,
0.0079436218,
-0.0011752383,
0.0478201658,
-0.0183555502,
0.0330082737,
-0.0274254251,
0.0217722394,
0.0235314406,
0.0178026184,
-0.0146178911,
-0.046872817,
-0.0214565005,
0.0382672474,
-0.0135379685,
0.0287561156,
0.0386640616,
0.0108820954,
0.0169117209,
0.0049630073,
-0.0750267804,
0.1081920043,
-0.016354369,
-0.0081917346,
-0.0238880552,
0.0092096953,
0.0252341833,
-0.0323224738,
-0.0136520099,
0.0825737566,
-0.0557890013,
0.0212828778,
-0.0122618685,
0.0329123326,
-0.015198078,
-0.0153490147,
0.022387214,
0.0333189107,
-0.0633243918,
0.0839173421,
-0.0579677075,
-0.0197820719,
-0.0402343683,
-0.0195513256,
-0.006416535,
-0.0180260912,
-0.0261117034,
0.0409633853,
-0.0168386362,
0.0478942804,
-0.0886060521,
-0.026429547,
-0.046253752,
0.0303281527,
-0.0271892268,
0.0204897467,
-0.0348439962,
-0.015347532,
-0.0227643959,
-0.0001258148,
-0.0291883182,
-0.0143336831,
-0.0388769172,
0.0143619506,
0.0058063213,
-0.053525418,
-0.0180647764,
0.0088199815,
0.0040463009,
-0.0540944152,
0.0334403925,
0.1015877351,
-0.0018566718,
-0.0343094058,
-0.089342542,
-0.0309538897,
-0.0338706486,
-0.0822064281,
-0.0322117768,
-0.0773033351,
0.0041225241,
-0.0124735925,
0.0099337175,
0.0074418983,
-0.00006757,
0.0005744349,
-0.0069202171,
0.0965987295,
0.0333696492,
-0.0145258959,
0.0137753477,
0.0403990187,
0.0133668361,
0.1001021266,
-0.0023690334,
0.0359727144,
-0.0000012827,
0.0220458061,
0.0029560335,
0.0282269008,
0.004992581,
0.0018601669,
-0.0385161117,
0.0057213423,
0.0111440886,
-0.0035919917,
-0.0089214351,
0.0253817774,
-0.0049187196,
-0.0030791685,
0.0053239386,
0.0408635959,
0.0588533282,
-0.0080842488,
0.0265746973,
-0.0537064634,
-0.0540404879,
0.0096975211,
-0.0536759943,
0.043303825,
-0.0689501986,
0.0044253366,
-0.029921867,
0.0360601023,
-0.0205508191,
-0.0152387312,
-0.0137789845,
-0.0062606875,
-0.0316935852,
0.0259569865,
0.0488612428,
-0.0124279866,
-0.0190544501,
0.0088552833,
-0.0025004433,
-0.0331258625,
-0.0352865383,
-0.056121394,
0.0087418491,
0.0306108836,
-0.0042081731,
-0.0648650154,
-0.0467458218,
0.0055092541,
0.0183489658,
-0.0090468265,
-0.0304571129,
-0.0016778955,
0.0390788205,
0.0035825865,
-0.0513126552,
-0.0033365018,
-0.0134659484,
-0.0035338181,
0.0424174368,
0.0311994925,
0.0004335329,
-0.0190573055,
0.0471790694,
0.0224645566,
-0.0171586983,
-0.0324444175,
0.0267807171,
0.0641007498,
0.0606782287,
0.0092109786,
0.0235682111,
-0.004421365,
0.0467391126,
-0.0553371087,
0.0002286142,
-0.0048313797,
-0.0609335229,
-0.0409832224,
0.0530005172,
-0.0031696546,
-0.0074319174,
-0.0363814458,
-0.0302594956,
-0.0753629953,
-0.0227276012,
-0.0338317789,
0.0333821736,
0.0068853106,
0.0301928967,
-0.0099067595,
-0.0004232072,
-0.044668708,
0.0622869693,
0.0019106105,
-0.008791633,
0.0455882736,
0.0583171435,
0.0172617212,
-0.0483984426,
-0.025297273,
0.0018949114,
-0.0020984979,
-0.0341241471,
-0.0024180824,
-0.0105419206,
-0.0587168075,
0.0071629882,
-0.0623029359,
0.0384795405,
-0.0210134219,
0.0372661501,
-0.014586363,
0.0148074813,
0.0118688373,
-0.0165630877,
-0.0600988604,
0.0398956053,
-0.004503936,
0.0189675223,
0.0180584528,
0.0363236703,
-0.0488310717,
0.0052180439,
-0.0180791896,
0.0042697596,
0.0176291782,
-0.0326403603,
-0.0742020011,
-0.0540613048,
0.0331798755,
0.008053828,
0.0136249661,
-0.0484241284,
0.0336613655,
-0.0190855768,
-0.0377146825,
0.0518531762,
0.0534948073,
0.008326631,
0.0148714762,
0.009996755,
-0.0216505099,
-0.0177051071,
0.0115446802,
-0.0103964237,
-0.0281399097,
-0.0248680823,
0.0248770602,
-0.0140787587,
0.0477603115,
-0.0462755747,
0.0076908288,
0.0050500995,
-0.0395157635,
0.0188437197,
-0.0240619201,
-0.0525891148,
0.0002037844,
-0.0259572268,
0.0590635091,
-0.0147018749,
-0.0444075465,
0.0666003451,
-0.0010614914,
0.0112669682,
0.0061244909,
-0.0755101144,
-0.0780750662,
0.0622231588,
-0.0153181534,
0.0066668293,
-0.0555394962,
-0.0446606353,
-0.0156743228,
0.0092279622,
-0.0344113521,
-0.0137352524,
0.1049218625,
-0.0030561846,
0.0258239172,
-0.025167577,
-0.0631475002,
-0.0448123701,
0.0506530739,
-0.0020697983,
0.0075486754,
0.0059771179,
-0.0019646704,
0.0051043923,
0.0212565679,
0.0101352278,
0.0330687016,
0.0028511824,
-0.0005302228,
-0.0313478708,
-0.0055247745,
-0.0055985805,
0.0189897325,
0.0413522013,
-0.0206723697,
0.0393287539,
0.0407947861,
0.0852594897,
-0.0483039469,
-0.0381511264,
0.0300902482,
0.035938859,
-0.0166331045,
-0.0276253168,
0.041595418,
0.0486171842,
-0.0725496858,
-0.0415931232,
0.0456835777,
0.0032729344,
-0.0648592189,
-0.0054130214,
-0.0129581885,
-0.0262596495,
0.0328332335,
0.0188153964,
-0.1086373329,
-0.0295712203,
0.0367568843,
0.017197052,
-0.0063077821,
-0.0406465419,
-0.0102612833,
-0.0928875953,
0.0561194122,
0.0266282,
-0.0761565417,
-0.0109187802,
-0.0333171748,
0.0085779,
-0.0020361939,
0.0003609978,
0.0403149426,
-0.029918218,
-0.048361145,
-0.0266428925,
0.033801347,
0.0082902825,
0.0148458201,
-0.0352549776,
0.0215326697,
0.0102745509,
-0.0157979466,
0.0225546975,
-0.0025136296,
0.043944072,
-0.0254326295,
0.0273140147,
-0.0235999469,
-0.0093296021,
-0.0206903275,
-0.0199382156,
0.0233349781,
0.0324142128,
-0.0546592288,
-0.0128168259,
0.0328972451,
0.1067229956,
-0.0450208709,
0.0399404801,
-0.0172408875,
-0.037006896,
0.0289669596,
0.0730787516,
0.0055725952,
-0.0050126584,
-0.0770527422,
-0.0160218198,
-0.0649920404,
-0.0359638333,
0.0098075075,
0.015749108,
-0.048596181,
0.0138401315,
-0.0241447445,
0.0196788367,
0.0353403501,
0.1012938097,
-0.0145466877,
0.0480105095,
-0.0158806127,
-0.0054429793,
-0.1214961335,
-0.0179884788,
-0.0689820275,
0.0600605235,
-0.0403804854,
0.0496585369,
0.040933907,
-0.0039464352,
0.0289602373,
0.0043045767,
-0.0862092823,
-0.0595265366,
0.0010788585,
0.0152108837,
0.0305748861,
0.0511383228,
0.0005515812,
0.0105898743,
0.0586882196,
0.0571391508,
-0.0146621829,
0.0362874381,
0.0074471501,
0.0338253938,
0.0102513395,
0.017329216,
-0.0466537364,
0.024357887,
0.0338623039,
0.0284200609,
-0.0143453432,
-0.002681186,
-0.0018618491,
0.0414370187,
0.0190258492,
-0.0005705592,
-0.0162449684,
0.0221728198,
0.0429694392,
-0.023525279,
-0.0051618936,
-0.02308405,
0.0217272397,
-0.0194815286,
-0.0018813545,
0.0397976823,
-0.0060926359,
0.0312014204,
0.0050211367,
-0.0599116944,
0.0499430932,
0.0029821219,
-0.0251617655,
0.0438315123,
0.0171100218,
0.0076653399,
-0.0025186874,
0.059122134,
0.0323716514,
-0.0050288574,
0.0442916937,
0.0316578746,
0.0480584912,
-0.0877343938,
0.0164807178,
-0.0556389615,
-0.0254455097,
0.0200460199,
0.0241646804,
0.0628833696,
-0.0148201631,
-0.0159723498,
-0.0282090437,
-0.0167999119,
-0.0003799503,
0.0145551935,
0.0034166072,
-0.0154798068,
0.0231236611,
0.0572072119,
0.0042437776,
0.024683198,
0.0586097129,
-0.0160474442,
0.0429268479,
0.0253449492,
0.0013099626,
-0.0558141805,
-0.0364760906,
-0.0073085777,
0.0558355227,
0.01618254,
-0.019610988,
-0.0435201749,
-0.0551495478,
0.0042871349,
-0.0174066778,
0.0119325453,
-0.0337998755,
0.0412839875,
0.0507562198,
0.0723370239,
0.0012192103,
-0.0180367827,
0.0186211523,
0.0191125236,
-0.0159648266,
0.0454408526,
-5.887181337e-33,
-0.0182780828,
-0.0552975386,
0.0521669425,
0.0285696499,
-0.0413512513,
0.0292490739,
-0.0209105127,
0.0056411987,
0.0308481269,
-0.0186340585,
-0.0235453229,
0.017970629,
0.0287446231,
0.0007928325,
0.0212146025,
-0.0168454573,
0.013247747,
0.0273159072,
0.0089498674,
-0.0150602926,
0.0386496261,
0.0369890817,
0.0117634535,
-0.0609192066,
-0.010825689,
0.0220537465,
0.0104770353,
0.0385103114,
-0.0032678545,
0.0522312485,
-0.0079478053,
0.0252639968,
0.0434186943,
-0.0900487378,
-0.018889429,
0.0367251933,
-0.09026777,
0.0165300034,
-0.0081435898,
-0.0111987665,
-0.0789705664,
-0.0488496423,
0.0315882601,
0.0243572257,
-0.0103635006,
-0.0010643949,
-0.0103901094,
-0.0844250098,
-0.0198471453,
-0.0276434831,
-0.042002555,
0.0291613489,
0.0279766172,
0.0109319659,
0.0769318491,
-0.0367673449,
-0.0234225355,
0.0421796627,
-0.0329930075,
0.0144807277,
-0.009302441,
0.0236059409,
0.0615459047,
0.0411544256,
0.0225004051,
0.0238796789,
0.0248443503,
0.0032299056,
-0.0027951528,
-0.0246680994,
-0.0244249161,
0.0920341387,
-0.0027258215,
0.0200384613,
0.0713279769,
-0.019928297,
-0.0770547017,
0.0103981122,
0.0587060712,
-0.0007531641,
-0.0016588835,
0.0114866002,
-0.0195105765,
-0.0166511666,
0.0201383494,
0.0270253904,
-0.0383206122,
-0.0009295647,
0.0233361088,
0.0329541899,
0.00272635,
0.0385607928,
0.0435070097,
-0.0123506971,
0.0246330425,
0.0149979359,
0.0120683881,
-0.0040068696,
0.0341520756,
0.0223291777,
-0.0166174676,
-0.0492079183,
-0.0162148867,
-0.0023899651,
0.0181628261,
-0.0156278741,
0.0368569866,
0.0184547082,
-0.084397018,
-0.0236908738,
-0.0504092984,
0.0322447605,
0.0050269892,
0.0393465832,
0.0456720404,
-0.0279620998,
0.0257574022,
0.0962847024,
-0.0180920139,
-0.0362887457,
0.0280091558,
-0.0100588715,
-0.0149544813,
-0.0006987319,
-0.0231763981,
-0.0126451356,
-0.0061120051,
0.0188483279,
0.0260218531,
-0.047570087,
-0.0032805461,
0.0142126689,
2.765e-7,
0.000222825,
0.009944763,
0.0765972212,
0.0007473992,
0.0404714644,
-0.0017188459,
-0.0047168406,
0.0146384258,
-0.0472007319,
-0.0450556353,
0.0076803402,
0.0062803039,
0.0245580543,
0.0050448501,
-0.0278766472,
0.0129515324,
0.0110240029,
-0.0122751389,
-0.0475401357,
-0.0060082492,
0.0457104892,
0.0635801777,
-0.0329836123,
-0.0462533087,
0.0309218038,
-0.0267159045,
0.0320490673,
-0.0366035365,
0.0180247873,
-0.0089007504,
0.0069023557,
0.0040405402,
-0.0043378277,
0.0070287101,
0.0226084162,
0.0460398681,
-0.035436362,
0.074411355,
0.0395056568,
-0.0531612821,
0.0219122395,
-0.0016407373,
0.0057596653,
0.0155530097,
0.0785481259,
-0.0381282456,
-0.020233022,
-0.0545888096,
-0.0030808921,
-0.0405526049,
-0.0225501265,
0.0318041258,
0.040845789,
0.0179116838,
0.0008147224,
-0.0138561195,
-0.0263143051,
-0.1093987226,
-0.0095705036,
0.0457659625,
-0.0292700529,
0.0048645865,
-0.0201533549,
0.012916795,
0.0364695415,
-0.0627478063,
-0.0645286366,
2.838074301e-34,
0.0160477068,
-0.0357843526,
0.0139265638,
0.0516489223,
-0.0075494745,
-0.0188488234,
-0.1016655639,
0.0381380096,
0.0290246755,
-0.0284576286,
-0.022472553
] |
http://arxiv.org/abs/2405.15307v1
|
http://arxiv.org/pdf/2405.15307v1
|
Before Generation, Align it! A Novel and Effective Strategy for Mitigating Hallucinations in Text-to-SQL Generation
| 2024-05-24T07:51:08 | 2024-05-24T07:51:08 |
Ge Qu, Jinyang Li, Bowen Li, Bowen Qin, Nan Huo, Chenhao Ma, Reynold Cheng
|
cs.CL
|
cs.CL
|
Large Language Models (LLMs) driven by In-Context Learning (ICL) have
significantly improved the performance of text-to-SQL. Previous methods
generally employ a two-stage reasoning framework, namely 1) schema linking and
2) logical synthesis, making the framework not only effective but also
interpretable. Despite these advancements, the inherent bad nature of the
generalization of LLMs often results in hallucinations, which limits the full
potential of LLMs. In this work, we first identify and categorize the common
types of hallucinations at each stage in text-to-SQL. We then introduce a novel
strategy, Task Alignment (TA), designed to mitigate hallucinations at each
stage. TA encourages LLMs to take advantage of experiences from similar tasks
rather than starting the tasks from scratch. This can help LLMs reduce the
burden of generalization, thereby mitigating hallucinations effectively. We
further propose TA-SQL, a text-to-SQL framework based on this strategy. The
experimental results and comprehensive analysis demonstrate the effectiveness
and robustness of our framework. Specifically, it enhances the performance of
the GPT-4 baseline by 21.23% relatively on BIRD dev and it yields significant
improvements across six models and four mainstream, complex text-to-SQL
benchmarks.
|
Before Generation, Align it! A Novel and Effective Strategy for Mitigating Hallucinations in Text-to-SQL Generation: Large Language Models (LLMs) driven by In-Context Learning (ICL) have
significantly improved the performance of text-to-SQL. Previous methods
generally employ a two-stage reasoning framework, namely 1) schema linking and
2) logical synthesis, making the framework not only effective but also
interpretable. Despite these advancements, the inherent bad nature of the
generalization of LLMs often results in hallucinations, which limits the full
potential of LLMs. In this work, we first identify and categorize the common
types of hallucinations at each stage in text-to-SQL. We then introduce a novel
strategy, Task Alignment (TA), designed to mitigate hallucinations at each
stage. TA encourages LLMs to take advantage of experiences from similar tasks
rather than starting the tasks from scratch. This can help LLMs reduce the
burden of generalization, thereby mitigating hallucinations effectively. We
further propose TA-SQL, a text-to-SQL framework based on this strategy. The
experimental results and comprehensive analysis demonstrate the effectiveness
and robustness of our framework. Specifically, it enhances the performance of
the GPT-4 baseline by 21.23% relatively on BIRD dev and it yields significant
improvements across six models and four mainstream, complex text-to-SQL
benchmarks.
|
[
-0.00458166,
0.0806292519,
-0.0170109253,
0.0191792306,
-0.0365323387,
0.0241987538,
-0.045471672,
-0.012189799,
0.0158732347,
-0.0433153287,
-0.0027612241,
-0.0283708591,
-0.0099500101,
0.0187351089,
0.0409758762,
-0.0454215817,
0.0373522639,
-0.0142383501,
-0.0444196016,
0.0232721344,
-0.0144490693,
0.0380953066,
0.0173645373,
-0.0017209805,
-0.0108717568,
-0.0376255102,
-0.0336931273,
0.0463435426,
-0.0050499314,
-0.0114499154,
-0.0239322279,
0.0470412262,
0.0049451673,
0.0419655964,
0.0000022414,
-0.0530645251,
-0.0544338189,
0.0089022797,
-0.0010981687,
0.0036194755,
0.0306061134,
-0.049582459,
0.0074473238,
0.0504872426,
-0.0457515717,
0.0051566856,
0.0356173553,
0.032334704,
0.0381324589,
0.0937174633,
-0.0283691883,
-0.066970475,
0.0185409654,
0.0090305116,
0.0582988635,
-0.038891308,
-0.0130821755,
0.0501696393,
0.0353464484,
-0.0244566556,
0.0042794649,
0.0217937697,
0.0057984288,
0.0079426877,
0.0024909475,
0.0104459906,
-0.0420101061,
-0.0027598208,
-0.00302142,
0.0349282511,
-0.0337734334,
-0.046507217,
0.0048903828,
0.0162935592,
-0.0003419379,
-0.0271326806,
0.0051102065,
0.03255767,
-0.0129414788,
0.0221380591,
0.0241670683,
-0.0328368805,
0.0322769769,
-0.0162179545,
-0.0675382018,
0.0229750481,
0.0047563636,
-0.0336368568,
0.0293578077,
-0.0028071697,
-0.0203292463,
-0.035767585,
0.0017678181,
0.0464096367,
-0.1035126746,
0.0258612465,
-0.0687233508,
0.0073468173,
0.0148358038,
-0.074913986,
0.0288190413,
0.0602454245,
0.0081885988,
0.0294073671,
-0.0314969122,
0.0186067317,
-0.002406447,
0.0062482255,
-0.0562713891,
-0.0209207721,
-0.0896466747,
-0.0011981633,
-0.0599158294,
0.0199689697,
-0.071300216,
-0.0416269973,
-0.0200895742,
0.0527888276,
0.0023563972,
0.0499987416,
0.0039179241,
0.0059630801,
-0.0173497386,
-0.0169739835,
-0.0205307864,
-0.0190968253,
0.0109158875,
-0.0099776601,
0.002833806,
-0.0555477552,
-0.0089810109,
0.0102053471,
-0.0344957151,
-0.0261905026,
0.0050385832,
0.0522528365,
-0.0284076016,
0.033352606,
-0.1016061381,
-0.08412572,
-0.0051223785,
-0.0549613498,
-0.027783433,
-0.0568004772,
0.0392250195,
0.026349036,
-0.0380635336,
-0.0500068441,
0.0032100261,
0.0305764638,
-0.0097610345,
0.0647095442,
-0.0469921604,
0.0260800216,
0.018401226,
0.0073852483,
0.0019710481,
0.107863456,
0.0142084192,
0.0340677947,
-0.0221609063,
-0.0236929823,
-0.0005763457,
0.0211439729,
0.0176806394,
0.0160459671,
-0.073100999,
0.0415739119,
0.0368927047,
0.0428635776,
0.0073660351,
0.06169229,
0.0111539755,
0.0644465834,
0.0622892752,
-0.0252815541,
0.0398870669,
0.0537232831,
0.0442482457,
0.0202145763,
-0.0289508272,
-0.0064517884,
-0.0735923052,
0.0229536016,
-0.0335709937,
0.0473143347,
0.0277630165,
0.0220476557,
-0.014340627,
-0.042609334,
-0.0134247066,
0.0126839904,
-0.0085028037,
0.0112586142,
-0.0002011486,
0.0194672886,
-0.0534357987,
0.0422042124,
-0.0053767515,
-0.0335891172,
-0.0009547227,
0.0010698932,
-0.02931251,
0.0917234719,
0.0174390804,
-0.0335178077,
-0.0385661237,
-0.0115485443,
0.0247929189,
0.0521263815,
0.0192547925,
-0.0290103685,
-0.0253463443,
-0.0342469141,
-0.0232827589,
0.0251013525,
0.0163704846,
-0.0271106325,
0.0194629114,
-0.0087086745,
0.023703482,
-0.0342671573,
-0.047811497,
0.0348047316,
-0.042649202,
-0.0371507183,
-0.0123402756,
-0.0047531025,
0.0219781846,
0.0032453726,
-0.0080267731,
0.0508983545,
0.0200466476,
-0.009983426,
-0.0313550122,
0.0134714199,
-0.0213828366,
-0.0145130157,
0.025103895,
0.0117512774,
-0.0084025469,
-0.0538338497,
-0.0549097881,
-0.0339850001,
0.0103096534,
-0.0327590033,
0.0465783253,
-0.0286734644,
0.0467510559,
0.0099659087,
-0.0353611447,
-0.0333473645,
0.0911877379,
0.0134010874,
-0.0156548508,
0.0081203366,
-0.0254316293,
0.0136861559,
-0.0373726375,
0.0604657009,
0.0024472962,
0.0052016629,
0.0325983725,
-0.0156493243,
0.0132351071,
-0.0725798681,
-0.0353992246,
-0.0796596706,
0.0244073309,
-0.0153120002,
0.0455484539,
0.0436927229,
0.0021136622,
0.0083479229,
0.0133715356,
-0.0583116934,
-0.0087085543,
-0.006064543,
0.0145964818,
0.0258130431,
-0.0052391076,
-0.036133308,
0.0068745809,
0.0108860349,
-0.0677958876,
0.0487014428,
-0.0261746161,
-0.028860135,
-0.0642083436,
0.0535488017,
-0.0108598312,
0.0227162745,
0.0124692917,
0.0420078002,
-0.0026238118,
-0.0227875281,
0.0620924123,
0.0388192721,
0.0179616939,
-0.0288106184,
0.0130051039,
-0.0209449604,
-0.0411054082,
0.0073335348,
0.0035962851,
-0.00603793,
-0.0216992237,
-0.0534467064,
0.0287199914,
0.0458183438,
-0.0196940266,
-0.020987818,
-0.0687515065,
0.0234217122,
0.0065340898,
0.003085678,
-0.0196610261,
-0.0161713939,
0.0172616504,
0.0077206977,
-0.03234265,
-0.0243537091,
0.0411198176,
-0.0241115168,
0.0398715995,
-0.0858997479,
0.025495097,
-0.0412430875,
0.0688058361,
-0.0010243326,
0.0368204638,
-0.0121651795,
-0.0350268595,
-0.0609335862,
-0.0228815377,
-0.0084149893,
-0.0210841428,
0.091419287,
-0.0542327873,
0.0306156687,
-0.0249003544,
-0.0071196421,
-0.0406590588,
0.025689248,
-0.0160388667,
0.0397831462,
0.0544676334,
0.0130776297,
-0.0082944259,
-0.0021010682,
-0.0136966147,
0.0195734575,
0.0142218517,
-0.0487626977,
-0.0056650355,
-0.042536173,
0.0301376544,
-0.0359136313,
0.0087084714,
-0.0318136401,
0.023804985,
0.0350852832,
0.048445221,
-0.0297931787,
-0.0185407382,
-0.0042936332,
-0.0047314144,
-0.0261484552,
-0.0248797685,
0.0492305644,
-0.0092976373,
-0.0501871742,
-0.0351763852,
0.1028529033,
0.0296312254,
0.0016584059,
0.0220815707,
-0.0026694904,
0.0548970103,
0.0254558697,
0.0355133265,
-0.1056904942,
-0.0378475711,
-0.0135340095,
0.0046075592,
0.0383716151,
-0.0241499878,
-0.0235129073,
-0.1112009361,
0.0233728699,
0.0248587877,
-0.0654714331,
-0.0115513587,
0.012587795,
-0.0371762812,
-0.0039902497,
0.0227574557,
0.0691440329,
-0.0241021719,
0.0053979703,
0.0121178348,
0.023920849,
-0.0362550281,
-0.0044705449,
-0.0448096059,
0.0179159753,
0.0623038262,
0.0136027187,
0.0511745363,
-0.0350392349,
0.0233546235,
0.0039843027,
0.0052054203,
-0.01812171,
0.054145202,
0.0115731172,
0.0237888452,
0.0313594639,
0.0166849382,
-0.0696253255,
-0.0139526492,
0.0216742903,
0.1078509018,
-0.0252319816,
-0.0128994472,
-0.0155259324,
-0.0151020652,
0.0161781609,
0.0706850141,
-0.0169588178,
-0.0175464991,
-0.1031480506,
-0.0175660197,
-0.0097017493,
-0.0610979758,
0.0121873133,
0.0534510687,
-0.0082175517,
-0.0412706733,
-0.0426065922,
-0.0383142382,
0.0393334813,
0.0586518534,
0.0287935399,
0.051546406,
0.0019524078,
-0.0354672,
-0.0560120083,
0.0005970809,
-0.0468874052,
0.0133972112,
-0.0222183708,
-0.0076411255,
-0.0065221214,
-0.0467042625,
0.0104970075,
0.0082927505,
-0.0822306871,
-0.0460814089,
0.0108050564,
-0.0054864776,
0.0267915484,
0.0333008505,
0.0147893652,
-0.0361951478,
0.0065340209,
0.0389635526,
-0.023901483,
-0.0230242927,
-0.0029602423,
0.0078194151,
0.0051753819,
0.0334059037,
-0.0406620055,
0.0086260447,
-0.0155301141,
0.0333896428,
-0.0001635921,
0.0388737433,
0.0576641634,
0.026605431,
-0.0347830988,
-0.0081616817,
-0.0259387977,
0.0014951298,
0.0276726317,
-0.0599233694,
-0.0436760299,
-0.0289915428,
0.0094614094,
0.0360244252,
0.0087206904,
0.0421709083,
-0.00950904,
0.007601371,
0.0266047884,
0.0046078963,
-0.0281164609,
0.0141451601,
-0.0076275589,
0.0638370961,
-0.0419206768,
-0.0231638402,
0.0145540815,
0.0206191782,
0.0638899282,
-0.0332573168,
0.0654545501,
0.0444398411,
0.0250961781,
-0.0325074382,
-0.0008446532,
-0.0147082247,
-0.0461530685,
-0.0332238935,
-0.0155216027,
0.0376855582,
0.0033821187,
0.0682939067,
-0.0385476761,
-0.0009511131,
-0.0165848862,
0.0202697255,
0.0395317599,
-0.0296874903,
-0.0081819156,
0.0071905567,
-0.0406010002,
-0.0626011714,
0.0771628022,
-0.0083352998,
0.0223044381,
-0.0217431877,
-0.048335433,
-0.0317792073,
-0.0529767349,
0.0144550717,
-0.0170468278,
-0.0263082311,
-0.0313615687,
-0.0586567856,
-0.0265050586,
0.0111903818,
0.0276016612,
-0.0102740079,
-0.020550862,
0.043336045,
0.08666832,
0.0343135744,
-0.0348319598,
-0.0052108834,
0.0043966663,
-0.0138113452,
-0.052467335,
-0.0304222181,
-6.325946359e-33,
-0.026468128,
-0.0204817615,
0.0672350526,
-0.0281827692,
-0.007119183,
-0.0484941229,
-0.0304646175,
-0.0001188792,
-0.0044914898,
-0.0064614592,
-0.0107192723,
0.0236559883,
0.0178045053,
0.0433165431,
0.0449331887,
-0.0209891051,
0.0354894474,
-0.0000427261,
-0.0074544628,
0.0327238478,
0.0279293843,
0.0140010035,
0.0244773198,
-0.0418994427,
-0.002564884,
0.0044594933,
-0.0098029291,
0.0131577393,
0.0106615759,
0.0053699804,
-0.0055200337,
0.0076176045,
0.020190388,
-0.0210585445,
-0.0243017729,
-0.0234460682,
-0.0260673072,
-0.0114456173,
0.0189679917,
-0.0064167008,
-0.0309137926,
-0.0400960371,
0.0554249883,
-0.0229432583,
-0.0290597267,
-0.0448841378,
0.0058253747,
-0.0545733236,
-0.0293631181,
-0.0315102562,
-0.1238070875,
0.0146069936,
0.0374312662,
0.0093973624,
0.0136184786,
0.0443507954,
-0.0169160478,
0.01720828,
-0.0427993983,
0.0219704136,
-0.0001518248,
0.0549453497,
0.079477638,
0.0376770683,
0.0020699801,
-0.0002536762,
0.0433853157,
0.0324049443,
-0.035802383,
0.0545679554,
-0.004558566,
0.0753660947,
0.0246063247,
0.0324339718,
0.0320173912,
0.0010195563,
-0.0860264078,
0.0427856743,
-0.0554348044,
-0.0051280675,
0.0474527739,
0.0190443005,
-0.0650176853,
-0.0471077338,
-0.0178351086,
0.0452837609,
-0.0262844563,
-0.0282941777,
0.0521749668,
-0.0080625778,
-0.0025856118,
0.0009044645,
0.0238298643,
-0.0249523986,
-0.0315683372,
0.0184858739,
-0.0486080013,
-0.030354755,
-0.019205736,
0.0360846594,
-0.0311131533,
-0.0675592944,
0.0181639325,
0.0432743393,
0.042950511,
0.0504222363,
0.0344297402,
0.02397817,
-0.0742917284,
0.0314664952,
-0.0471719801,
0.0512419827,
-0.0271342043,
0.0151562886,
0.0433041975,
-0.0147238681,
0.0281419959,
0.0165093653,
0.0338460431,
-0.0255200267,
-0.0004572826,
0.0168175008,
0.0013697892,
0.0174849275,
0.041436296,
0.0306748506,
-0.0020594972,
-0.0384615511,
0.0497071482,
-0.0502916649,
-0.0223421417,
0.0468153954,
2.95e-7,
0.0470304601,
0.0366435982,
0.0293812212,
0.059549246,
0.0005004185,
0.0355030969,
0.0133204591,
0.0082766442,
0.0136353485,
-0.0020746042,
0.0227503441,
0.0036268784,
0.0564674661,
0.0516808182,
-0.0590530708,
0.0120300408,
-0.0218236521,
0.0198657848,
-0.0631333366,
-0.0056818686,
0.0561487377,
0.0640294328,
-0.0502537452,
-0.0166607015,
0.0436365977,
-0.0434546098,
0.0352459103,
-0.0131361429,
-0.008616291,
0.0681475475,
-0.0002088527,
-0.0127900317,
-0.0155646559,
0.0494039319,
-0.0152925998,
-0.0124991918,
-0.0236610807,
0.0456213169,
-0.0253575966,
0.005838275,
0.0050427997,
-0.0723238289,
0.0220288821,
-0.0638410151,
0.0561421663,
-0.0583848543,
-0.0187616087,
0.0646166801,
-0.027462516,
-0.0084981751,
-0.0166399796,
0.0353559442,
0.0243190546,
-0.0219944343,
0.0371561721,
0.0438945107,
-0.0166025553,
-0.0423794948,
-0.0394257307,
0.0688302293,
-0.0229608528,
-0.0070242425,
-0.0070558097,
0.0520179346,
-0.0132164583,
-0.0479373075,
-0.011683153,
3.302457247e-34,
-0.0028781437,
0.0195371769,
0.0581211597,
0.0048981421,
0.0412630849,
0.0029280405,
0.0078464206,
0.0308882166,
0.0063099414,
-0.0511902161,
-0.0253641624
] |
http://arxiv.org/abs/2405.15306v2
|
http://arxiv.org/pdf/2405.15306v2
|
DeTikZify: Synthesizing Graphics Programs for Scientific Figures and Sketches with TikZ
| 2024-05-24T07:48:35 | 2024-05-28T06:48:58 |
Jonas Belouadi, Simone Paolo Ponzetto, Steffen Eger
|
cs.CL
|
cs.CL, cs.CV
|
Creating high-quality scientific figures can be time-consuming and
challenging, even though sketching ideas on paper is relatively easy.
Furthermore, recreating existing figures that are not stored in formats
preserving semantic information is equally complex. To tackle this problem, we
introduce DeTikZify, a novel multimodal language model that automatically
synthesizes scientific figures as semantics-preserving TikZ graphics programs
based on sketches and existing figures. To achieve this, we create three new
datasets: DaTikZv2, the largest TikZ dataset to date, containing over 360k
human-created TikZ graphics; SketchFig, a dataset that pairs hand-drawn
sketches with their corresponding scientific figures; and SciCap++, a
collection of diverse scientific figures and associated metadata. We train
DeTikZify on SciCap++ and DaTikZv2, along with synthetically generated sketches
learned from SketchFig. We also introduce an MCTS-based inference algorithm
that enables DeTikZify to iteratively refine its outputs without the need for
additional training. Through both automatic and human evaluation, we
demonstrate that DeTikZify outperforms commercial Claude 3 and GPT-4V in
synthesizing TikZ programs, with the MCTS algorithm effectively boosting its
performance. We make our code, models, and datasets publicly available.
|
DeTikZify: Synthesizing Graphics Programs for Scientific Figures and Sketches with TikZ: Creating high-quality scientific figures can be time-consuming and
challenging, even though sketching ideas on paper is relatively easy.
Furthermore, recreating existing figures that are not stored in formats
preserving semantic information is equally complex. To tackle this problem, we
introduce DeTikZify, a novel multimodal language model that automatically
synthesizes scientific figures as semantics-preserving TikZ graphics programs
based on sketches and existing figures. To achieve this, we create three new
datasets: DaTikZv2, the largest TikZ dataset to date, containing over 360k
human-created TikZ graphics; SketchFig, a dataset that pairs hand-drawn
sketches with their corresponding scientific figures; and SciCap++, a
collection of diverse scientific figures and associated metadata. We train
DeTikZify on SciCap++ and DaTikZv2, along with synthetically generated sketches
learned from SketchFig. We also introduce an MCTS-based inference algorithm
that enables DeTikZify to iteratively refine its outputs without the need for
additional training. Through both automatic and human evaluation, we
demonstrate that DeTikZify outperforms commercial Claude 3 and GPT-4V in
synthesizing TikZ programs, with the MCTS algorithm effectively boosting its
performance. We make our code, models, and datasets publicly available.
|
[
0.0019466809,
0.1039390936,
0.0011782998,
0.0165219661,
-0.0100236917,
-0.0016088438,
0.0135251451,
0.0386281274,
-0.0510187633,
0.0002840809,
-0.0399657935,
0.0197462682,
0.0059628612,
0.0652312338,
0.0128439721,
-0.1023330092,
-0.014739993,
-0.0299440715,
-0.030368384,
-0.014591855,
-0.011926502,
0.0078896023,
-0.0105639938,
0.0394152552,
-0.0379336141,
-0.0259034261,
-0.0406777486,
0.0191843342,
0.028226655,
-0.0438008681,
-0.0667996928,
0.067332536,
0.0310215224,
0.0402244888,
0.0000021058,
-0.0187479444,
0.0018865822,
0.0149131408,
0.0571290292,
0.0342552103,
0.0923298821,
-0.077301912,
-0.0581124164,
-0.1100039184,
0.0043281624,
-0.0288933944,
0.08515466,
0.0413304269,
0.0993476734,
-0.011601585,
0.0031454482,
-0.056870468,
0.0223316476,
0.0085492041,
0.0350493528,
0.0349853896,
-0.0084717842,
0.0420913994,
0.0294794235,
0.0197900143,
-0.0138912303,
0.0686549321,
0.0265503228,
-0.0164757296,
0.0210496653,
0.049511902,
0.0022351777,
-0.0298810154,
0.0218396001,
0.075240761,
-0.0343852341,
-0.0108749811,
0.0229454022,
-0.0006825265,
-0.0189821199,
0.0346197821,
-0.0998202413,
0.0232055113,
-0.0069862264,
-0.0526748635,
-0.0372021496,
-0.029746376,
-0.0059052976,
-0.0159511231,
0.0208845008,
-0.004633049,
0.0416029803,
0.0183276925,
-0.0325820595,
-0.0137284892,
-0.0523313954,
-0.0783793628,
0.029842969,
-0.0245094281,
0.037402384,
-0.0202993266,
-0.0268328954,
-0.0280517805,
0.0003382859,
0.0397968516,
0.0655223578,
0.033977367,
-0.0248292852,
0.0327920616,
-0.0161469597,
-0.01071613,
-0.0462516025,
0.0435302667,
-0.0147507973,
0.0016790982,
-0.0773854181,
-0.0246364772,
-0.0331932865,
-0.0315764584,
-0.0211367011,
0.0080685876,
-0.0330774523,
0.0204644781,
-0.0192953702,
-0.0013908446,
-0.013158218,
0.0488722622,
-0.0545836575,
0.0040555215,
-0.0498887561,
0.0050709289,
-0.0307643171,
0.0143308453,
0.0114515508,
0.028964676,
-0.0121165179,
0.0072692195,
0.015449821,
-0.0090283463,
-0.0110705402,
0.0437684618,
0.0210575797,
0.0081101563,
0.0068906648,
-0.0376747884,
0.0528825298,
-0.0346570462,
0.0586528927,
-0.0682986826,
0.0168057699,
-0.0393962264,
0.0156691056,
-0.0347866714,
-0.0126150856,
0.0085404776,
-0.0083192224,
0.081311658,
-0.0666947216,
-0.0292864367,
0.0254699159,
0.0344675742,
-0.0225365218,
-0.0043394607,
0.039162688,
0.0323213898,
-0.0391561799,
-0.0693405271,
-0.0058130082,
-0.0120123504,
0.0351342261,
0.0221650023,
-0.0305229034,
0.0021555743,
-0.0270194169,
0.0967653617,
0.0335848741,
-0.0010588028,
-0.0796210915,
-0.0028910595,
0.023132639,
0.0735920817,
0.0205950122,
0.0280821547,
0.0674807802,
-0.0079390062,
0.0365996249,
-0.0386437327,
-0.0419436209,
0.0809747353,
0.0542557426,
0.0406758152,
-0.0273983758,
0.0376056135,
-0.0085720895,
-0.0374760404,
0.0279617682,
-0.0238713101,
0.0236801561,
-0.0877100974,
0.0364596359,
0.0602726601,
-0.0096857371,
0.0128604462,
-0.0151456762,
-0.053956762,
-0.0348272435,
-0.0220126025,
0.0159203187,
0.0631068647,
0.023663586,
-0.0175090563,
-0.0015024357,
-0.0193804782,
0.0134249451,
0.0621851198,
0.0001674916,
-0.005077268,
0.0416234583,
-0.0190790035,
-0.0005625336,
-0.0543395914,
0.0061366903,
0.0545931011,
0.0053278836,
-0.0139439059,
0.0052739196,
-0.0483272225,
0.0014063492,
0.0383562148,
-0.0068623996,
-0.0140527897,
-0.0128765162,
0.0051287492,
0.0038256405,
0.0080094058,
-0.0127703696,
0.0062720133,
0.0316866226,
-0.0204931349,
0.0059614652,
0.010886156,
-0.0032955853,
-0.0450735018,
0.0192759074,
-0.0905465484,
-0.0440404043,
0.0050707734,
-0.0085377377,
-0.017778283,
0.0799488276,
0.0123424716,
0.0028731171,
0.0074649509,
-0.0354289301,
0.016638469,
-0.0216556415,
-0.0064295372,
0.0238073412,
-0.0135322828,
0.0015953153,
0.0591883995,
0.0183966309,
0.0022637742,
-0.0018770066,
0.0063568735,
-0.0014876302,
0.023730807,
0.0052010464,
-0.0198698137,
0.0194716007,
-0.04626384,
0.0196033996,
-0.0653507486,
-0.0243752152,
0.0423336476,
-0.0101777883,
0.0085448958,
-0.0363010652,
0.0026476986,
0.0149696292,
0.0201176088,
-0.006067811,
-0.0219201297,
-0.0162910372,
0.0519536287,
-0.01315339,
-0.0142191015,
-0.0135014933,
-0.0247134641,
-0.03065788,
0.0187226851,
0.0639589205,
-0.007186715,
-0.0242440607,
0.0460251682,
-0.0640759319,
-0.0042244247,
0.01346015,
0.0009004201,
0.0534040891,
-0.0546903238,
-0.0405932106,
0.0417671613,
-0.0289204214,
-0.0216106474,
0.0300458521,
-0.0398340523,
-0.0395749025,
0.0073102936,
0.0145977698,
0.0581564941,
0.0579074174,
-0.0169504415,
0.0175731871,
0.0423930213,
-0.0127693396,
0.0364581719,
-0.0038969454,
0.0766680837,
0.0395458229,
0.0106000006,
0.024248248,
0.0044214521,
0.0245098136,
0.0150358146,
-0.0162196867,
-0.0448900759,
0.0679111332,
-0.0337408632,
-0.0074816733,
0.0369159505,
0.0235276185,
0.0003483709,
0.0674853325,
0.0113618225,
0.0276243649,
0.0324429534,
0.0249811076,
-0.0237462632,
0.0330616385,
0.0402812585,
-0.0269299224,
-0.0288610216,
-0.043175593,
0.0097310012,
-0.0348722339,
-0.0173615851,
-0.0043436922,
0.0368602388,
-0.011089229,
-0.0113956695,
-0.0203902312,
-0.0784192234,
-0.0020692213,
0.0169124138,
0.0133042103,
0.0015075785,
0.0009503823,
-0.0324191526,
-0.0102595743,
0.0144561967,
0.024236951,
0.0216774531,
-0.0722147673,
-0.0236423612,
-0.0345179997,
0.0448196344,
-0.0152159035,
0.0217788089,
-0.0840406865,
0.0112943929,
-0.0102367196,
0.0746214613,
0.0029595224,
-0.0168636534,
0.0548400767,
-0.0381489992,
0.0093331309,
0.0294127073,
-0.0009680632,
0.0010770033,
-0.0160934627,
0.0179102737,
0.0106441919,
0.0078125708,
0.0119347479,
-0.0970897675,
-0.0470715463,
-0.0172567479,
0.0051676217,
0.014619031,
-0.0332232155,
-0.0158498753,
-0.0600857027,
0.0579251722,
0.04274242,
-0.1022292525,
-0.0347462781,
-0.0125162592,
0.0161747783,
0.0166468117,
-0.073270604,
0.0058252639,
0.0034574706,
0.0118427938,
0.0024524403,
0.0188069586,
-0.0399225131,
0.0048693134,
0.0082074599,
0.0100087021,
0.0454829037,
0.026555948,
0.0044565587,
-0.0111852158,
0.0414007716,
0.0585533641,
-0.0075515057,
-0.0236856546,
0.0272881426,
0.065199919,
0.0035377066,
0.008479448,
0.0299028736,
-0.0252480041,
-0.0181727353,
-0.0062989336,
0.0429176502,
-0.0011149331,
-0.0227385089,
-0.0687217638,
-0.041086372,
-0.0097217783,
0.0223565139,
0.0438269228,
0.0038752016,
-0.0699395239,
-0.0852190405,
0.0521890968,
-0.0573712885,
-0.0499585606,
-0.0276854243,
-0.017239213,
-0.0605234988,
-0.0346855447,
0.0254714806,
0.0357774571,
0.0460700244,
0.0094203856,
-0.0547037534,
-0.0326677524,
-0.0164962448,
-0.0718663931,
0.06522277,
0.0057829982,
0.0147872129,
0.0067898687,
-0.0453095622,
-0.0051678121,
-0.0207840446,
-0.0032218772,
0.0767442957,
-0.0843386948,
-0.0219831001,
-0.028687261,
0.0978040844,
0.0188362468,
-0.0300037246,
0.0014922812,
-0.0447057299,
0.0158698596,
0.1024636254,
0.0243180189,
-0.023902338,
-0.0050270357,
0.0364992544,
0.0122926533,
-0.0314050168,
-0.0094420174,
0.0374531336,
-0.0134362094,
0.0109246261,
-0.0052017062,
0.0026869064,
0.0031510605,
0.0471138209,
0.0166331399,
-0.0205345154,
0.0091874581,
-0.0221485682,
-0.0396519266,
0.0308325123,
-0.0222670082,
-0.038070675,
0.0453203507,
0.015079814,
-0.0068060546,
0.0406306908,
-0.0332380086,
0.0228654202,
0.0008978739,
-0.0240304843,
-0.0115340035,
0.0634108856,
0.0176371336,
0.0449844152,
0.0018686192,
-0.0048303073,
0.0176502708,
-0.0504517555,
0.0494472422,
-0.0218485743,
-0.016837135,
0.0212481804,
-0.0240170229,
-0.0302126743,
0.0528562739,
-0.056647785,
-0.0548444539,
-0.053377036,
0.0472799689,
0.0590274855,
0.0683657825,
0.0401113629,
-0.0505355708,
-0.0078271311,
-0.0116394432,
0.0584664606,
0.0589427277,
0.0006648069,
0.0403775424,
-0.00090176,
0.0116895949,
-0.0932871103,
0.0357895158,
0.0059190984,
0.0265428424,
-0.0333252251,
-0.034545254,
-0.0370330065,
0.0188567135,
-0.0842876956,
0.01188483,
0.0203595757,
0.0061494079,
-0.0193303153,
-0.071705915,
0.0299249254,
0.0412907898,
-0.0243676305,
0.0088031599,
0.0635621399,
0.0281353779,
-0.0188245382,
-0.0332913138,
0.0550276041,
0.0525457077,
0.0061322195,
-0.0702360943,
-0.0034606834,
-5.184733097e-33,
-0.0115040494,
-0.0149874464,
0.0042415382,
-0.0114429211,
-0.0005314038,
-0.033460509,
-0.0609689057,
0.0012863927,
0.0386130586,
0.013517173,
0.0120621612,
-0.0530756079,
-0.0213963296,
-0.0178137776,
0.0339234732,
0.0293805413,
0.0203853585,
0.0333774351,
-0.0766739622,
0.0176152941,
0.0199502632,
0.0366619602,
0.0648967102,
-0.0678034648,
-0.0439272597,
0.0072481548,
0.0224667694,
-0.0435084999,
-0.009706255,
0.0494258441,
-0.0318063945,
0.0392930545,
-0.0082190586,
-0.0121180201,
-0.0245506912,
-0.0061251461,
-0.0832589865,
-0.0215143934,
-0.0062587676,
-0.0118489554,
-0.0668982342,
-0.0354177319,
0.0436971933,
0.0010301681,
-0.0280557498,
-0.0215942618,
0.0327571593,
-0.0249939691,
-0.00119045,
-0.0078722397,
-0.0177589245,
0.0155119263,
-0.0487237237,
-0.0339938477,
0.0507528149,
-0.0237029381,
0.0187958851,
0.0276630558,
-0.0237455405,
0.0757469758,
0.0027685638,
-0.0020268939,
0.0501760952,
-0.0033611562,
0.0596234314,
0.0615099855,
0.0431724377,
-0.0042900005,
-0.0088582393,
-0.0121578714,
-0.0113819437,
0.0417428352,
0.0102641694,
0.0331724919,
0.0660285577,
0.0132149709,
-0.0213385038,
0.0375371501,
-0.0087621519,
0.0267578512,
-0.0257241037,
0.0481390134,
0.0455212183,
-0.0248838291,
-0.0031552627,
-0.0561493672,
-0.0264631007,
0.0078637674,
0.0087917577,
-0.0128181474,
-0.0330800712,
0.0146244662,
-0.011706288,
-0.0159059018,
-0.0250698086,
-0.0394834094,
-0.0607662089,
0.0394099355,
-0.0062811919,
-0.0392084718,
0.0022877906,
-0.034245722,
0.0171238855,
0.0040225121,
0.0580464825,
0.0087230029,
0.0057175299,
0.0084881093,
0.0001853628,
0.0001081858,
-0.042172879,
-0.0341482125,
0.0152139999,
-0.0294572283,
0.0082649682,
-0.0170002114,
-0.0256234575,
-0.0186456852,
-0.0082512591,
0.042886246,
-0.0365018584,
-0.0066994657,
-0.0481196903,
0.0008710386,
-0.0226911213,
-0.0087612867,
-0.0289326981,
-0.0301403385,
0.0186582506,
-0.0666994229,
-0.0395232178,
0.0206006207,
2.702e-7,
0.0460950322,
0.0424664579,
0.0279982518,
0.0005298481,
0.0072815455,
0.0307311583,
-0.0685960501,
-0.0121235186,
-0.0176116414,
-0.0101914555,
-0.0230593365,
-0.0063242437,
0.0461618789,
0.0666595623,
-0.042121049,
-0.0201966856,
0.0009504567,
-0.0031233458,
-0.0540034026,
0.0267480966,
0.0214823093,
0.0220841654,
-0.001240155,
-0.0550085679,
0.0577673353,
-0.0361141898,
0.0476189926,
-0.035825368,
0.0539547987,
0.0024004378,
-0.0269300211,
0.0174569208,
-0.0045849187,
0.0157435872,
0.0092754336,
-0.0541806109,
0.0501690246,
0.0296613146,
-0.0236911792,
0.0821417347,
0.0498189926,
0.0005059102,
0.023561161,
-0.0138483858,
0.0009146459,
-0.0244245958,
-0.003994436,
-0.0014509059,
0.0281885937,
0.0035729511,
0.0402203016,
-0.0322692469,
-0.0266703814,
-0.0118238274,
0.0049415589,
0.0129251843,
0.0176264383,
-0.0219055805,
-0.0372692794,
-0.0157021508,
-0.0348039754,
-0.0447331555,
-0.0484902412,
-0.015569672,
0.0238835942,
-0.0090268366,
-0.0348618962,
2.474679406e-34,
0.0361452624,
0.0169506688,
0.042779509,
0.0188329276,
0.0238181707,
0.0011729045,
0.0681641921,
-0.0223823339,
0.0076240087,
-0.0181251056,
-0.0446435362
] |
http://arxiv.org/abs/2405.15302v1
|
http://arxiv.org/pdf/2405.15302v1
|
Towards Understanding How Transformer Perform Multi-step Reasoning with Matching Operation
| 2024-05-24T07:41:26 | 2024-05-24T07:41:26 |
Zhiwei Wang, Yunji Wang, Zhongwang Zhang, Zhangchen Zhou, Hui Jin, Tianyang Hu, Jiacheng Sun, Zhenguo Li, Yaoyu Zhang, Zhi-Qin John Xu
|
cs.AI
|
cs.AI, cs.CL, cs.LG
|
Large language models have consistently struggled with complex reasoning
tasks, such as mathematical problem-solving. Investigating the internal
reasoning mechanisms of these models can help us design better model
architectures and training strategies, ultimately enhancing their reasoning
capabilities. In this study, we examine the matching mechanism employed by
Transformer for multi-step reasoning on a constructed dataset. We investigate
factors that influence the model's matching mechanism and discover that small
initialization and post-LayerNorm can facilitate the formation of the matching
mechanism, thereby enhancing the model's reasoning ability. Moreover, we
propose a method to improve the model's reasoning capability by adding
orthogonal noise. Finally, we investigate the parallel reasoning mechanism of
Transformers and propose a conjecture on the upper bound of the model's
reasoning ability based on this phenomenon. These insights contribute to a
deeper understanding of the reasoning processes in large language models and
guide designing more effective reasoning architectures and training strategies.
|
Towards Understanding How Transformer Perform Multi-step Reasoning with Matching Operation: Large language models have consistently struggled with complex reasoning
tasks, such as mathematical problem-solving. Investigating the internal
reasoning mechanisms of these models can help us design better model
architectures and training strategies, ultimately enhancing their reasoning
capabilities. In this study, we examine the matching mechanism employed by
Transformer for multi-step reasoning on a constructed dataset. We investigate
factors that influence the model's matching mechanism and discover that small
initialization and post-LayerNorm can facilitate the formation of the matching
mechanism, thereby enhancing the model's reasoning ability. Moreover, we
propose a method to improve the model's reasoning capability by adding
orthogonal noise. Finally, we investigate the parallel reasoning mechanism of
Transformers and propose a conjecture on the upper bound of the model's
reasoning ability based on this phenomenon. These insights contribute to a
deeper understanding of the reasoning processes in large language models and
guide designing more effective reasoning architectures and training strategies.
|
[
0.0168587957,
0.0585303679,
-0.002151303,
0.0512735583,
-0.0484278202,
0.0359885618,
-0.0051594474,
-0.0331803784,
-0.0703491867,
-0.0402606614,
-0.0352026522,
-0.0425760336,
0.036744535,
-0.0071740937,
0.0129267257,
-0.0110940626,
0.0290744342,
-0.0326555222,
-0.0363365449,
-0.0368261822,
-0.0257086288,
-0.0007315752,
-0.0227162018,
0.0447618328,
0.019328244,
-0.0193283074,
-0.0090783108,
-0.0295596607,
0.0020136295,
-0.0061946893,
-0.0187369883,
0.0294395443,
-0.0007147632,
0.0627874285,
0.0000018097,
-0.0556895435,
-0.0258866325,
0.0037650124,
-0.0332714468,
-0.0424118042,
-0.0205524284,
-0.0332016759,
-0.0240900796,
0.0234102532,
-0.034882158,
-0.0169339254,
0.0635285079,
0.0875410587,
0.0634020418,
0.0752661601,
-0.010844579,
-0.0832109302,
0.0234217495,
-0.0081214309,
-0.033640828,
-0.0341341645,
0.0271338541,
-0.007345384,
-0.0098396335,
-0.0253208186,
-0.0097763129,
0.0032231782,
0.033298701,
0.0275010727,
-0.0427672863,
0.042910967,
-0.0028762938,
-0.0365751199,
-0.0046044835,
0.0430100076,
-0.0561072417,
-0.0110469311,
-0.0132854003,
0.0030425163,
-0.0159954876,
0.0243564043,
-0.0374519266,
0.0495776348,
-0.0054975227,
0.0341687426,
0.0033680445,
-0.0031873754,
0.037000563,
-0.0286227949,
-0.0857276767,
0.1152870506,
0.0204117466,
-0.0254154652,
-0.0022400643,
0.0026614845,
-0.0386774652,
-0.0553501658,
0.0160754789,
0.0690957904,
-0.0196833368,
0.0263616443,
-0.012483702,
-0.0291198529,
-0.0393204428,
-0.0408775695,
0.0394470654,
0.071741879,
0.003498099,
0.0330482498,
-0.0360340625,
0.0483342707,
-0.0673922673,
0.0558840707,
-0.034163706,
-0.0154794985,
-0.0098927366,
-0.0245755333,
-0.049621664,
0.036003273,
0.0057673822,
-0.0086057531,
-0.0371683203,
0.0105080269,
-0.0029424345,
0.0200705063,
0.018471146,
-0.0246156082,
-0.0156872962,
-0.0326972492,
-0.0423004888,
0.029891571,
-0.0093403393,
0.0233382005,
-0.0006984678,
-0.0293569583,
-0.0215374473,
-0.0122891944,
0.0174167138,
-0.0538095906,
0.0050024539,
0.0535896197,
-0.0222851895,
0.002293593,
-0.0268670544,
-0.0163872279,
0.0227366686,
-0.0804019198,
0.0507956892,
-0.0531857908,
0.000332841,
0.0366269238,
-0.0001546968,
-0.032300815,
0.0078146085,
0.0214390215,
-0.0289534684,
0.0869092047,
0.0038327491,
0.0009569133,
0.0083462261,
0.014015479,
-0.0001233378,
0.0585907884,
0.0095542483,
-0.0389319696,
-0.0110203018,
-0.0299475845,
0.0176475979,
0.027748581,
0.0183098074,
-0.0548477024,
-0.0420782678,
0.0037296209,
-0.0112373987,
0.022453865,
-0.004966693,
0.0513387695,
-0.0490056984,
-0.0209577121,
0.0163350198,
0.0633316636,
0.0609189942,
0.072260052,
0.0118853766,
-0.0074254223,
-0.0028843789,
0.0620275214,
-0.0567676835,
0.0467844494,
-0.0342436619,
0.0018119223,
-0.0304202866,
0.0511222146,
-0.0415884107,
-0.0520677567,
-0.0423620418,
0.0072265631,
0.0023046646,
-0.0160602257,
0.062413767,
0.0398134403,
-0.0325225592,
0.1155162081,
0.0136802942,
-0.0381738544,
-0.0357573852,
-0.090627037,
0.0002218247,
-0.0012888891,
0.0223853625,
-0.0005100716,
-0.0313263386,
0.0032605743,
-0.0123472968,
0.0286188424,
0.045174703,
0.0172027685,
-0.0103519168,
-0.0543158799,
-0.0553413108,
0.0151215522,
0.0062774583,
-0.0074411267,
0.0043586176,
-0.0335852429,
0.0138397189,
0.0049908119,
0.0213273745,
0.0247518402,
-0.0345148705,
-0.006885644,
-0.0021380009,
0.0361856446,
0.0063070683,
-0.009147252,
0.002537671,
0.0185587425,
0.0391907282,
-0.010136297,
-0.0439961627,
0.0531881638,
-0.06034052,
-0.0099759363,
0.0311728157,
-0.0088963043,
-0.0247986652,
0.0572170466,
-0.0074926894,
-0.0492345057,
0.0407231934,
-0.0385560133,
0.0290111545,
-0.0246800091,
0.0412785895,
0.0062355502,
0.0250867717,
0.0001785068,
0.0830520466,
0.0013355725,
0.0208253898,
0.0016307565,
0.0262919385,
0.0238771308,
-0.0190285947,
-0.0467028469,
0.0171079691,
-0.0537162386,
0.0190034863,
-0.0186016764,
-0.0353493281,
-0.0120403646,
-0.0623664334,
0.0050893263,
-0.0108441049,
-0.0156017365,
0.0155214667,
0.0132626826,
0.0017095928,
0.0252772924,
-0.0033556398,
-0.03299243,
0.0069820466,
0.0332979411,
0.0254534073,
0.0365915932,
0.0014303451,
-0.0558507144,
0.0065149367,
-0.0257229861,
-0.0357888564,
0.0673249215,
0.0124694575,
0.024824718,
-0.0809746683,
0.0220305603,
-0.0185774937,
0.0424625948,
-0.0278795268,
-0.0270497818,
0.0292392485,
-0.0394602679,
0.0145185189,
0.0473446883,
-0.0050449073,
0.0315657146,
0.0626925826,
0.002279826,
-0.0143951941,
0.0232930575,
-0.0552260801,
-0.0088228537,
-0.0524716638,
0.0031532114,
-0.0405343026,
0.1283169091,
-0.0388364829,
-0.013321341,
-0.0508243218,
0.0323282182,
0.0066559426,
0.008621132,
0.0087176301,
-0.0124652041,
-0.0059837699,
0.0739827082,
-0.0230148491,
-0.0059438627,
0.0630030409,
-0.0090689221,
-0.009635672,
-0.0751691312,
0.0356189162,
-0.0334419757,
0.0579673722,
-0.0102092884,
0.0144272428,
0.0110256439,
-0.0203273911,
-0.0525167957,
-0.0297332928,
0.0245030392,
-0.0284741689,
0.0606872179,
-0.0053414181,
0.0052315132,
-0.0306833014,
-0.0427111462,
-0.067412965,
0.0021764229,
-0.0357332602,
0.0202678274,
0.0061835852,
-0.0338192321,
-0.0672618449,
0.0214671623,
0.0059347325,
0.0410905443,
-0.0148304347,
-0.0250574574,
-0.0329380594,
0.0185691379,
0.0271746535,
0.0790054053,
0.064649187,
-0.0239745919,
0.0624711737,
0.0255596414,
0.0516784079,
0.0069878362,
-0.0490298681,
0.0135062002,
0.0514915437,
-0.0213432573,
-0.0208929516,
0.0140592279,
0.0261122696,
-0.0390792489,
-0.0146938758,
0.104794912,
0.0153894639,
-0.0279751197,
0.0141831515,
-0.0496256053,
-0.0088390838,
0.0366231091,
0.0410234965,
-0.0599639006,
-0.0152357658,
0.0214779358,
-0.0166445691,
0.0135829942,
-0.0078881206,
-0.0260710586,
-0.1113434285,
0.0341070294,
-0.00664652,
-0.0308043398,
-0.0416575149,
-0.0006932028,
0.0002865278,
-0.0172361452,
0.0351545662,
0.0162149929,
-0.0125688398,
-0.0351420566,
0.0289301705,
0.0348379426,
-0.0502969734,
0.0282323435,
-0.0437471643,
0.0269085634,
0.0426914915,
0.0240576193,
-0.0089873262,
-0.0681771114,
0.0267034192,
0.0292080287,
0.0047895084,
0.0050294455,
0.0105568599,
0.0412454531,
-0.0215832945,
0.0068327552,
0.0186002273,
-0.0422041677,
-0.0255866088,
-0.0443065651,
0.0740545988,
-0.0172840673,
-0.0220898911,
0.0279615559,
0.0239371099,
-0.0425461233,
0.0640947521,
0.0088018849,
-0.0067690015,
-0.045498345,
-0.0369498618,
-0.0059392028,
-0.0272457916,
-0.0420834832,
0.0398561992,
0.060756173,
-0.0442207307,
-0.0370303579,
0.0019837832,
0.0004325357,
0.0425365716,
0.015564369,
-0.035049919,
0.0427303128,
0.0149151133,
-0.0394841246,
0.0320672467,
-0.021650834,
0.0471837297,
-0.0412516072,
0.0406030342,
0.0145347668,
-0.0505245291,
-0.049134586,
-0.0278811436,
-0.0888176188,
-0.0422789194,
0.036679469,
0.0005294762,
0.0353496522,
0.0461770408,
-0.0227580313,
-0.0499826372,
-0.0031258818,
-0.0053916783,
-0.0242915768,
0.0029738292,
0.0768653825,
-0.0407439806,
0.0039771916,
0.0412918031,
-0.0025248455,
0.0326825902,
0.0372110531,
0.023892168,
-0.0267295707,
-0.0263753291,
0.0573262125,
0.004433495,
0.0561557785,
0.0460879914,
0.0371756963,
-0.0117242215,
0.0603292398,
0.0045373081,
-0.0088642798,
-0.0567355268,
-0.0136275413,
0.0250028204,
0.0206015967,
0.0172700528,
-0.01742642,
0.0196381696,
0.0175653081,
-0.001645627,
-0.0067093344,
-0.0061712218,
-0.0109309815,
0.033115264,
0.0117008025,
0.0106085604,
0.0146795651,
0.0391285382,
0.0350444689,
-0.0070756092,
0.0710847452,
0.0338536128,
0.007355283,
-0.0593079366,
-0.0275687166,
-0.0995487943,
-0.0665492937,
0.0021279943,
-0.0379466526,
0.0411719419,
-0.0125113353,
0.0155396862,
-0.0279012229,
0.006335204,
0.0182083827,
0.0205706917,
-0.0024015608,
-0.0171980318,
0.0648853257,
0.0189241581,
0.0066987271,
-0.047070533,
0.0673002675,
-0.0231572464,
0.0487300754,
-0.0261874925,
0.0095477067,
-0.0261442736,
-0.0512994044,
-0.0050049634,
0.0106419045,
-0.017879799,
0.0056417109,
-0.0410566889,
-0.0827249736,
-0.0003869928,
-0.0037034941,
0.0145812016,
-0.0401969627,
0.0131668923,
0.078910619,
-0.0099192588,
-0.0382135622,
0.0027587223,
-0.0294739883,
0.0140421009,
0.0270936694,
-0.0106562134,
-5.829913457e-33,
-0.0175911915,
-0.0104774609,
0.0516504049,
0.0186019372,
0.0263515469,
-0.0259701367,
-0.0220000055,
0.0014000414,
-0.0437990837,
0.0023968783,
-0.0226776861,
0.0265029483,
0.0234659128,
0.0341665931,
-0.0037276552,
-0.0217903629,
0.0497375317,
-0.0089376885,
-0.0187217016,
-0.0214021839,
0.0148738949,
0.0084280437,
0.0638037175,
-0.0160370972,
-0.0118156113,
0.0000798274,
-0.0352317281,
0.0444261096,
-0.0663511977,
0.0800264552,
-0.0136985946,
-0.0035413788,
0.0208882261,
-0.0652724355,
-0.0288612023,
0.0064875288,
-0.0666097701,
-0.0250566881,
0.0229660925,
0.0483165011,
-0.015522616,
-0.0135812117,
0.045207914,
0.0010779276,
-0.0347799137,
-0.0551554151,
0.022628719,
-0.0195457619,
0.0085530197,
0.0014362619,
-0.0487313345,
0.0210249368,
0.0008024792,
-0.0392789543,
0.0406223536,
0.0187293179,
-0.0272083506,
0.0042412998,
-0.0589664094,
0.0311915074,
0.0215557124,
0.0370016247,
0.0458453409,
0.0382712595,
-0.0268864185,
0.0310044736,
0.0506829992,
-0.006422691,
-0.0138870981,
-0.0266214628,
0.0190917384,
0.0556224808,
0.0722088739,
-0.0208081268,
0.048569113,
0.0223506019,
-0.0461994894,
-0.0113334171,
-0.0026124264,
0.0558754504,
0.044334922,
-0.0061459313,
-0.0061819339,
-0.0240290053,
-0.0514520295,
-0.0080520762,
-0.0689655766,
0.0085166031,
0.0071173818,
0.0124990772,
-0.0189158432,
0.0366105102,
0.0114373481,
-0.066796653,
0.0003333154,
0.0086602233,
-0.0090040946,
-0.0004475986,
-0.0131481169,
0.0283572171,
-0.0351748466,
-0.0336434469,
-0.0085959807,
0.0009906043,
-0.0002880329,
0.0251062047,
0.0309403352,
0.0486652665,
-0.0570630096,
-0.0243328325,
-0.0642694086,
-0.0009677604,
0.0143605592,
0.0460203998,
0.0459654108,
0.0269024614,
-0.0096997628,
0.0127478763,
-0.0008525263,
0.0079668313,
-0.0475550741,
0.0552619211,
0.0078665046,
0.0477241799,
-0.0114673628,
-0.045790717,
-0.0616584457,
-0.0066115335,
0.0495051369,
-0.0601544231,
-0.0264045205,
0.029315168,
2.619e-7,
0.0418000109,
0.0157105774,
0.0075774253,
0.0284786746,
0.0290823523,
0.0332120918,
-0.0441379212,
-0.0072199046,
0.0425690785,
0.0130125992,
-0.0012692609,
0.0245780777,
0.0396094918,
0.0291268621,
-0.0347566158,
-0.0157164931,
-0.0565348901,
0.0135234119,
-0.0025286097,
0.031294886,
0.0320849493,
0.1009341106,
0.0046212603,
-0.0219250172,
-0.0198271796,
-0.0419802368,
0.0573543273,
0.0056766723,
0.0461661965,
0.0271022357,
0.0151928393,
-0.0135671673,
0.0065893964,
0.0347000137,
-0.0058143651,
0.066451706,
0.001676186,
0.0643698871,
0.0017751564,
-0.0090531977,
-0.0111667542,
0.0005775737,
-0.0194226298,
-0.0484115668,
0.0721363798,
-0.1059735119,
-0.0621529855,
-0.0364279523,
0.0118962536,
-0.0147808297,
-0.0084475558,
-0.0006831891,
-0.0012536421,
-0.0057828571,
0.0296257269,
-0.0126725873,
0.0095355455,
-0.1095477492,
-0.03592382,
0.0000322013,
-0.0480545275,
-0.0330136381,
-0.022368636,
-0.0491395518,
0.0731288418,
0.0016493341,
-0.0274343435,
2.61013286e-34,
0.0177050363,
-0.00242855,
0.0004109503,
0.0248274375,
0.0239695031,
-0.0344006158,
0.0138886534,
0.0439587384,
0.00436247,
-0.0917054266,
-0.0239404105
] |
http://arxiv.org/abs/2405.20770v1
|
http://arxiv.org/pdf/2405.20770v1
|
Large Language Model Sentinel: Advancing Adversarial Robustness by LLM Agent
| 2024-05-24T07:23:56 | 2024-05-24T07:23:56 |
Guang Lin, Qibin Zhao
|
cs.CL
|
cs.CL, cs.AI, cs.CR
|
Over the past two years, the use of large language models (LLMs) has advanced
rapidly. While these LLMs offer considerable convenience, they also raise
security concerns, as LLMs are vulnerable to adversarial attacks by some
well-designed textual perturbations. In this paper, we introduce a novel
defense technique named Large LAnguage MOdel Sentinel (LLAMOS), which is
designed to enhance the adversarial robustness of LLMs by purifying the
adversarial textual examples before feeding them into the target LLM. Our
method comprises two main components: a) Agent instruction, which can simulate
a new agent for adversarial defense, altering minimal characters to maintain
the original meaning of the sentence while defending against attacks; b)
Defense guidance, which provides strategies for modifying clean or adversarial
examples to ensure effective defense and accurate outputs from the target LLMs.
Remarkably, the defense agent demonstrates robust defensive capabilities even
without learning from adversarial examples. Additionally, we conduct an
intriguing adversarial experiment where we develop two agents, one for defense
and one for defense, and engage them in mutual confrontation. During the
adversarial interactions, neither agent completely beat the other. Extensive
experiments on both open-source and closed-source LLMs demonstrate that our
method effectively defends against adversarial attacks, thereby enhancing
adversarial robustness.
|
Large Language Model Sentinel: Advancing Adversarial Robustness by LLM Agent: Over the past two years, the use of large language models (LLMs) has advanced
rapidly. While these LLMs offer considerable convenience, they also raise
security concerns, as LLMs are vulnerable to adversarial attacks by some
well-designed textual perturbations. In this paper, we introduce a novel
defense technique named Large LAnguage MOdel Sentinel (LLAMOS), which is
designed to enhance the adversarial robustness of LLMs by purifying the
adversarial textual examples before feeding them into the target LLM. Our
method comprises two main components: a) Agent instruction, which can simulate
a new agent for adversarial defense, altering minimal characters to maintain
the original meaning of the sentence while defending against attacks; b)
Defense guidance, which provides strategies for modifying clean or adversarial
examples to ensure effective defense and accurate outputs from the target LLMs.
Remarkably, the defense agent demonstrates robust defensive capabilities even
without learning from adversarial examples. Additionally, we conduct an
intriguing adversarial experiment where we develop two agents, one for defense
and one for defense, and engage them in mutual confrontation. During the
adversarial interactions, neither agent completely beat the other. Extensive
experiments on both open-source and closed-source LLMs demonstrate that our
method effectively defends against adversarial attacks, thereby enhancing
adversarial robustness.
|
[
0.0451412201,
0.05467223,
-0.0136162182,
0.0542376712,
-0.0343482718,
-0.0261646677,
0.0203964971,
0.0021637061,
0.0255906563,
-0.0558735207,
-0.0123771802,
-0.0203189533,
-0.0023560307,
0.0041945074,
0.0749408156,
0.0016067021,
0.0454344377,
-0.0226841439,
-0.021182375,
-0.0163299702,
-0.0034966534,
-0.0325214751,
0.0102066705,
0.0065285596,
0.0242791865,
-0.023649266,
-0.0178860724,
0.0132000772,
0.0266389791,
-0.0264848564,
0.026930619,
0.0046219607,
0.000218392,
0.085475646,
0.0000018956,
-0.040899381,
-0.0202564299,
0.0353472233,
-0.0036435837,
0.0569424927,
0.0530999005,
-0.0406839885,
-0.0425520986,
0.0293485411,
-0.0314849503,
0.017180359,
0.0197983161,
0.0565685704,
0.0854480043,
0.1261345595,
0.0051929839,
-0.026593864,
0.0463868678,
-0.0377546735,
-0.0546550564,
-0.0587931015,
0.0322093368,
-0.0336948186,
-0.0225118697,
0.034142524,
0.0012612934,
0.0228076708,
0.0196503159,
-0.0028661797,
-0.0579491444,
0.0198137853,
-0.0233386811,
0.0262163002,
0.0061954712,
0.0564398989,
-0.0390082635,
-0.0196960755,
-0.0005921079,
0.0040307003,
0.0114505552,
-0.0080204755,
-0.0164587777,
0.0100450851,
0.0307084508,
0.0102139162,
-0.0053182165,
-0.0216337908,
0.0611566976,
0.0230647977,
-0.0123497592,
0.0329991281,
0.0327984281,
-0.0164485499,
-0.0243692379,
-0.0454075038,
-0.0125457225,
-0.0457529575,
-0.002594192,
0.026327895,
0.0015847348,
0.0139473928,
-0.0962510332,
-0.0551745817,
0.0382869951,
-0.0580947176,
0.0232682843,
-0.0046292706,
-0.0242488757,
0.0299030617,
-0.0452937782,
0.0543693341,
-0.0252040718,
0.0611086786,
0.0061687096,
0.0027213681,
-0.0036895752,
-0.0105195744,
0.0105607212,
0.035214562,
-0.0075186859,
0.0065842397,
-0.0313234068,
0.0341118164,
-0.0291102603,
0.0612043142,
-0.0556522273,
0.0092227086,
-0.0082572233,
-0.0136222309,
0.0116817793,
-0.026686335,
-0.016925117,
-0.0091605298,
-0.0347654298,
-0.0786770061,
-0.0417896844,
0.0296767019,
-0.0583528206,
-0.0425387211,
0.0019796381,
0.1478607804,
0.0152020482,
-0.0330946036,
-0.0410295725,
-0.0276756324,
-0.0052855993,
-0.0606160127,
0.0115613127,
-0.0361676328,
-0.0261447504,
-0.0250799358,
-0.0146109089,
-0.0990626216,
0.0220893994,
0.0067853215,
-0.0263648536,
0.0780685693,
0.0321001858,
-0.0387015156,
0.0193710551,
0.0144408597,
-0.0019756292,
0.0626682788,
0.0244749263,
0.0490364879,
0.0411281474,
0.0324037634,
0.0148955164,
0.0395139568,
0.022867104,
0.043084003,
-0.0464840457,
0.0104407137,
0.0016997011,
0.0041875811,
-0.0153068844,
0.0699037239,
-0.00335822,
-0.0801495463,
0.1192547157,
0.0271899011,
-0.0587428994,
0.0445948839,
-0.0043306504,
0.0107865129,
0.0110348109,
0.0120609952,
-0.06995143,
0.0015353682,
-0.0891338438,
-0.0360652842,
-0.039350275,
0.044478409,
-0.0279506799,
-0.0235548802,
-0.0119323963,
-0.0242002755,
0.0244290121,
-0.0112389605,
0.0355191752,
0.0315368809,
-0.039737694,
0.0180311147,
0.0217184369,
-0.0252361912,
-0.0150138447,
-0.0838800147,
0.0053116437,
0.021586515,
0.0741876513,
-0.0495689102,
0.0190366767,
-0.0312288851,
0.0120151956,
0.0473570339,
0.0212717038,
0.0224275906,
0.0115919886,
-0.0421684757,
0.0038263195,
-0.0826032609,
0.0072494284,
0.0263847746,
0.0348609351,
0.0189677384,
0.0106513873,
0.0273897201,
-0.0368752033,
0.0577133894,
0.0110155931,
-0.0750143975,
0.0197043326,
0.0060136649,
0.0330936164,
0.0047775945,
-0.0461902842,
0.0263635889,
0.0613631979,
-0.009319433,
-0.0268439334,
-0.017360704,
0.0148046473,
-0.0266158916,
0.0359111615,
-0.0089850072,
-0.0039306325,
0.0711325258,
-0.048700545,
0.0197573025,
0.0492556579,
-0.0335597284,
-0.0006375156,
-0.0112066381,
0.0183592197,
-0.0323610492,
0.046115078,
-0.0433108509,
0.0348850861,
-0.0000982418,
-0.0438654497,
-0.0202161297,
0.0436659195,
0.0082102809,
-0.0122382166,
-0.0563884005,
0.0185804516,
-0.0231560227,
0.0232964996,
0.0032219915,
-0.0068762922,
0.0023136989,
-0.0420736559,
-0.0074432427,
0.0017648775,
-0.0197842587,
-0.0065954546,
0.0365953408,
-0.0268592592,
-0.0032500548,
-0.0000826457,
-0.0045685372,
0.0155026577,
-0.0129316244,
0.011847469,
0.0249160174,
-0.013793081,
-0.0277208667,
-0.0249551218,
0.0076066144,
-0.0085808709,
0.0827524215,
-0.0228034854,
0.0001373452,
-0.0339118913,
0.0328033045,
-0.0220699087,
0.059840031,
-0.0268647969,
-0.0717398375,
0.0303728841,
0.0063157706,
-0.0071267742,
0.0869183317,
0.0359275378,
0.0505229123,
0.0845026895,
0.0220415704,
-0.0203352869,
0.0012131486,
0.0130929286,
-0.0298569333,
-0.0649116039,
-0.0058801943,
-0.0020312429,
0.0441742539,
0.029619595,
-0.0112848924,
-0.0213125125,
0.0539210029,
0.0082163224,
-0.0456828177,
-0.0055503608,
-0.0250206869,
-0.0372617505,
-0.0474774912,
-0.0326056294,
-0.0355621316,
0.0901864767,
-0.0483427718,
0.024853304,
-0.0146635985,
-0.0374133997,
0.0139068076,
0.0706255585,
-0.016837921,
0.0079924632,
-0.0000724076,
0.0205600429,
-0.0564596355,
-0.0439102836,
0.0383110382,
-0.0120145772,
0.0222540013,
-0.0326405205,
-0.0102080526,
0.0004160593,
-0.0089806635,
-0.0653368905,
-0.0566168614,
-0.032536801,
0.0288719684,
-0.0056819511,
-0.0155945085,
-0.052332256,
-0.0218870118,
0.0332500972,
-0.0102521479,
-0.019863179,
0.0110291224,
-0.0323197618,
0.0083295964,
0.039276965,
-0.0427337326,
0.0135353087,
-0.0244621132,
0.0690724924,
0.0870327279,
-0.0110576963,
-0.00704794,
-0.0362071358,
-0.032778345,
0.0011054281,
-0.0112436051,
0.0106149688,
0.0028438261,
0.0405792221,
-0.0468453504,
0.0031442198,
0.0499698035,
0.0057243798,
0.0138242543,
0.0308611616,
-0.0090237148,
0.01996563,
0.0128200464,
0.0094676344,
-0.0454313606,
-0.0120240767,
0.015191637,
0.012909486,
-0.0080290996,
-0.0015023535,
-0.0424868986,
-0.0940104872,
-0.0028686931,
0.0213637687,
-0.0186151396,
-0.0139333559,
-0.0128834303,
-0.022144258,
0.006124421,
0.0314768031,
0.0060727317,
0.0216490235,
-0.0315733775,
-0.0115112746,
0.010755511,
-0.0171569157,
0.0002326767,
-0.0535869338,
0.0394861028,
0.0890472084,
0.0378616117,
-0.0187249742,
-0.01153237,
0.0367912762,
-0.0078227753,
0.0091885403,
0.0347545482,
0.0498198904,
-0.0037427265,
-0.0201975014,
0.0321894288,
0.0398909338,
0.0187760461,
-0.0182570815,
0.0519808643,
0.0069566271,
-0.0130154956,
-0.0147342524,
-0.0126444008,
0.0015075206,
0.0099542998,
0.0713449493,
0.0065253042,
-0.0537414812,
-0.050994169,
-0.0726436973,
-0.032085754,
-0.0406557471,
-0.0638034046,
0.0149334557,
0.002210163,
-0.0111293141,
-0.0668400526,
-0.0307672452,
-0.0226680357,
0.010606125,
0.0049245786,
0.0034923258,
-0.0372698158,
-0.0005603685,
0.0001109575,
-0.0361265279,
-0.0290844552,
0.0606137291,
0.0050024367,
0.0196284354,
0.0255915802,
-0.0240383185,
-0.0153174959,
0.0023523096,
-0.0767394006,
-0.115632847,
0.01046375,
0.0500768013,
0.0153029719,
0.0218294617,
-0.0427765064,
-0.0709065124,
0.0905934423,
0.0315039866,
-0.0119049558,
0.0311994869,
0.0723660812,
-0.0214819219,
0.0346985087,
0.0551962741,
-0.0361231267,
-0.0132017732,
0.0730720311,
-0.0022934848,
-0.0296161324,
0.0018985304,
-0.0011162335,
0.033546526,
0.0333270729,
-0.0086387275,
-0.0024982919,
0.0373964012,
0.0463319756,
-0.0509487838,
-0.0189454369,
0.0036786087,
0.0288844686,
0.0095589776,
-0.0106621478,
0.0062533454,
0.0224196631,
0.0054712486,
0.0390052944,
0.0037770416,
0.0120088989,
-0.0001660182,
0.0518346243,
0.0345521979,
0.0295451134,
-0.0114145847,
0.0467027836,
0.0002376944,
0.0066468143,
0.0010993413,
0.0285718106,
0.0113121746,
-0.0061047184,
-0.0309077855,
0.0195845515,
-0.0851782933,
-0.0512644649,
-0.0012489399,
0.0616197363,
0.0432865471,
0.0354993194,
0.0202576257,
-0.0317069776,
-0.0122874128,
-0.0166393314,
-0.015148419,
-0.0049943039,
-0.0416722298,
0.0403553098,
0.0010561366,
-0.0068787634,
0.0324865244,
0.0738813132,
0.0295412093,
0.0745369643,
0.0154950051,
-0.0003796082,
-0.0387713648,
-0.0424282812,
0.0367286429,
-0.0021758103,
-0.0282413308,
-0.0153248599,
-0.0319580287,
-0.0792870671,
-0.0006916371,
0.0222173464,
0.0234596413,
-0.0160495844,
-0.0101957172,
0.0368293822,
0.0090298885,
0.0407595299,
-0.059971869,
0.0355508476,
0.0282488707,
-0.0698765665,
0.0109087694,
-5.796482397e-33,
0.0039542983,
-0.0207354221,
-0.0146445595,
-0.0432934873,
0.0137392413,
-0.0048359595,
-0.0252813864,
-0.0238123294,
-0.0287896916,
-0.0083022602,
0.0016094099,
0.0073629464,
-0.0091049364,
0.0152926287,
-0.0573930964,
-0.0335459635,
0.041763939,
-0.022986874,
-0.0164715741,
-0.0068543046,
0.0318077728,
0.0041088527,
-0.0012916706,
-0.0619407855,
0.0451641493,
-0.0199708305,
-0.0633703992,
-0.0016506803,
-0.045166377,
0.0575978011,
-0.0021160888,
0.0188124441,
-0.0102941506,
0.0322865285,
-0.0026993558,
0.0116620557,
0.0126552274,
0.000661078,
-0.0020853104,
0.000705486,
-0.0237549786,
-0.0492638424,
0.03849959,
-0.0269645751,
-0.0063958359,
0.0068403101,
0.0014208822,
-0.0244142599,
-0.0007004305,
-0.0229521152,
-0.0339529626,
0.0174796693,
-0.0183965806,
-0.0067147259,
-0.0060379971,
-0.025612995,
0.036202684,
0.0220571533,
-0.0447446704,
0.0170329977,
0.0032570593,
0.0168210138,
0.0371443219,
0.0039159004,
0.0100320764,
0.069069773,
0.0000688823,
0.013987503,
-0.0145509634,
-0.03446595,
0.0401708111,
0.0384816416,
0.0788631737,
0.1116599366,
0.0156961773,
-0.0596432947,
-0.0981688127,
0.0063076359,
0.0383094586,
0.0313346982,
0.029775653,
0.0329742171,
-0.0175088011,
0.0177693516,
-0.0249951165,
-0.0461069979,
-0.0411743447,
-0.0457979999,
0.0063477587,
0.0181852803,
-0.05540042,
-0.0305350516,
-0.0018651393,
-0.0642831177,
-0.005767162,
0.0361421742,
-0.0368724391,
-0.0323597416,
-0.0001433754,
0.0021739975,
-0.0110303396,
-0.0146819055,
0.0129708275,
-0.0618852563,
0.0194881428,
-0.0057965727,
0.0505315326,
0.0303551778,
-0.055183433,
-0.0039738952,
-0.0141697126,
0.0334808603,
0.0341310129,
-0.0111028478,
-0.0004836858,
-0.0047718575,
-0.004137591,
-0.0080771111,
0.0243746135,
-0.0376075804,
-0.0211568475,
0.0307534933,
-0.0084417388,
0.0222618729,
-0.0300492831,
0.0006744094,
-0.0094158575,
0.063594915,
0.04307485,
-0.0816192031,
-0.0091444049,
-0.0502238534,
2.658e-7,
-0.0043762643,
0.0423274189,
0.0079995673,
0.0266192947,
0.0363593623,
0.0628558248,
-0.0265475623,
0.0864021629,
-0.0476046391,
-0.0311765485,
-0.0193961933,
0.026337171,
0.0449900664,
0.0281269103,
-0.0531081185,
0.0312605575,
-0.0859256312,
-0.0172185451,
-0.0537293702,
-0.0224457756,
0.0473269708,
0.1079995036,
0.0506204702,
-0.002952294,
-0.0196753107,
-0.0962815955,
0.0353313796,
-0.0048348359,
-0.0008748224,
-0.004882968,
0.0217125304,
0.0185312834,
0.0044257706,
-0.000874242,
-0.0307139587,
-0.0026421163,
-0.0202470515,
0.0738393068,
-0.0128858676,
-0.0322749913,
-0.0071104737,
-0.0226762015,
-0.0148188667,
-0.0214421526,
0.0693321005,
0.0110295368,
-0.0179745462,
-0.0241585486,
-0.0142873628,
-0.0229396615,
0.0234580003,
-0.0537560508,
-0.0271342341,
0.0072665145,
-0.0040395898,
0.0112463143,
0.0179011319,
-0.0699424073,
-0.0519894138,
-0.0229265895,
-0.0241685137,
-0.0387133174,
-0.0023817667,
0.0288144629,
0.0449889451,
-0.0106342137,
0.0051777475,
2.313833667e-34,
0.0228112768,
-0.0397959724,
-0.0242194254,
0.0237691421,
0.0423298366,
0.0049220608,
0.0168365184,
0.0161388814,
0.0172863863,
-0.0250586793,
-0.0366975665
] |
http://arxiv.org/abs/2406.00024v1
|
http://arxiv.org/pdf/2406.00024v1
|
Embedding-Aligned Language Models
| 2024-05-24T06:11:17 | 2024-05-24T06:11:17 |
Guy Tennenholtz, Yinlam Chow, Chih-Wei Hsu, Lior Shani, Ethan Liang, Craig Boutilier
|
cs.CL
|
cs.CL, cs.AI, cs.ET, cs.LG
|
We propose a novel approach for training large language models (LLMs) to
adhere to objectives defined within a latent embedding space. Our method
leverages reinforcement learning (RL), treating a pre-trained LLM as an
environment. Our embedding-aligned guided language (EAGLE) agent is trained to
iteratively steer the LLM's generation towards optimal regions of the latent
embedding space, w.r.t. some predefined criterion. We demonstrate the
effectiveness of the EAGLE agent using the MovieLens 25M dataset to surface
content gaps that satisfy latent user demand. We also demonstrate the benefit
of using an optimal design of a state-dependent action set to improve EAGLE's
efficiency. Our work paves the way for controlled and grounded text generation
using LLMs, ensuring consistency with domain-specific knowledge and data
representations.
|
Embedding-Aligned Language Models: We propose a novel approach for training large language models (LLMs) to
adhere to objectives defined within a latent embedding space. Our method
leverages reinforcement learning (RL), treating a pre-trained LLM as an
environment. Our embedding-aligned guided language (EAGLE) agent is trained to
iteratively steer the LLM's generation towards optimal regions of the latent
embedding space, w.r.t. some predefined criterion. We demonstrate the
effectiveness of the EAGLE agent using the MovieLens 25M dataset to surface
content gaps that satisfy latent user demand. We also demonstrate the benefit
of using an optimal design of a state-dependent action set to improve EAGLE's
efficiency. Our work paves the way for controlled and grounded text generation
using LLMs, ensuring consistency with domain-specific knowledge and data
representations.
|
[
0.0337877795,
0.0462898985,
0.0107082091,
0.0025825261,
-0.0021165169,
0.0009698456,
-0.0035785628,
-0.0099602249,
0.0004097707,
-0.0483787283,
-0.0058763544,
-0.0304099713,
-0.0313453823,
0.039468646,
0.0330797955,
-0.0472829901,
0.0064976462,
-0.0094364174,
0.012564729,
0.0081937304,
0.017428901,
0.0084989658,
0.0044391872,
0.038358923,
-0.0117364302,
-0.0155607844,
-0.0154214948,
-0.0116920797,
0.039876543,
-0.042267967,
0.031855654,
0.027896082,
0.0849612802,
0.0111758867,
0.0000020782,
-0.0016684448,
0.0187264271,
-0.0373582095,
0.0290714428,
0.0329855755,
0.0579081811,
-0.0130927749,
-0.005956097,
0.0209171753,
-0.0465620756,
-0.0138647901,
0.0575033389,
0.0327035412,
0.0503268242,
0.0683235601,
-0.016997423,
-0.0036183612,
-0.0153813679,
-0.03154926,
-0.0019752472,
0.0175461639,
-0.0056476961,
0.0350434482,
0.0210352801,
0.0112838745,
-0.0122497035,
0.0323692225,
0.0017527407,
-0.0566968359,
0.0305015892,
0.0691156015,
0.0161270779,
-0.0102760736,
-0.0178282596,
0.0833476186,
0.0045671803,
-0.0288328435,
0.0000567161,
-0.0446610786,
-0.0376258716,
0.0065857726,
-0.0250135548,
0.0320814401,
0.0321899056,
0.0334155522,
0.0124217635,
-0.0072698155,
0.0553544126,
0.0088211037,
0.0021526231,
0.0072761108,
0.031691514,
-0.0562572256,
-0.0372239947,
-0.0058482885,
-0.0259511638,
-0.039113041,
-0.0082713654,
0.0325005278,
-0.0264210682,
0.0381652899,
-0.0559567064,
-0.0191791374,
-0.0047304723,
-0.0222272594,
0.0385252312,
0.0420515984,
0.0271502789,
0.0369331092,
-0.0335191861,
0.0681375936,
-0.0193161406,
-0.0010723633,
-0.0287176557,
0.0713687167,
-0.0124091674,
-0.0301552154,
-0.0104267132,
0.0170045141,
-0.0222576074,
-0.0629760176,
-0.0275071263,
0.0529874489,
-0.0141813746,
0.0367994234,
0.0130911013,
0.040212173,
-0.0364583321,
0.0035717995,
-0.021968063,
0.0016442413,
-0.0343419574,
0.0144935595,
0.0263814274,
0.0218732823,
-0.0066805133,
0.0373137817,
0.0076937955,
0.0051435493,
-0.0184622072,
0.0649811849,
-0.0263043046,
-0.0311601479,
-0.0481826663,
-0.0159993805,
-0.0171781145,
-0.0377753973,
-0.0062461919,
-0.0474916473,
0.0482876226,
0.0122301364,
-0.0200917274,
-0.0394618921,
0.0142127527,
0.0071522458,
-0.0084501635,
0.0538218506,
0.0118298987,
0.0165509023,
0.0378062762,
-0.0024018115,
0.0173674915,
0.0674464926,
0.040488407,
0.0670220703,
0.0116614392,
0.0242166929,
-0.0189557113,
0.0261257123,
-0.0014946992,
0.0445518233,
-0.0390342884,
-0.0054903794,
-0.0148955174,
0.0748445466,
0.0283884164,
0.0260827243,
-0.0414025746,
0.0256860014,
0.0739957541,
0.0543415546,
0.0139541226,
-0.0112653542,
0.0801376328,
0.0454053357,
-0.0306310207,
0.0005399257,
-0.0031922879,
0.0440826491,
0.0375570692,
0.0110939797,
0.0449458808,
0.0005034306,
-0.0381301939,
-0.024200771,
0.005596383,
-0.0234830119,
0.0321930386,
-0.0170918033,
0.0150450645,
0.0049306052,
-0.0843808725,
0.0075010336,
0.0250166152,
-0.0460660197,
0.0561105162,
0.0017994683,
0.0438179225,
0.1067143381,
0.0655096918,
-0.0416985005,
-0.0381919332,
-0.032072939,
0.0276564211,
0.0868711695,
0.032231953,
-0.0429210551,
-0.0304414704,
-0.0834281743,
-0.0384914987,
-0.0520840362,
0.012173499,
0.0512950197,
0.0075223325,
0.0186898988,
0.0170902796,
-0.0109421201,
-0.028928047,
0.049782224,
-0.067312412,
-0.0576167032,
-0.0059826998,
0.0064188102,
0.0172454622,
0.0034607453,
-0.0278220028,
0.0117401965,
0.0169227496,
-0.0371548049,
-0.0421501659,
-0.0014064889,
-0.0554939732,
-0.0231483821,
0.0076326975,
-0.0387681685,
0.0203570444,
0.0133069633,
-0.0301050767,
-0.0553000085,
0.015009095,
-0.0376098864,
0.0592075624,
-0.0499937832,
0.0146124298,
0.0131637231,
-0.0162088666,
-0.0465086289,
0.0489603616,
-0.0213279296,
0.0147846108,
0.0413015634,
0.0444956236,
0.0080406861,
-0.0116745243,
-0.0032441902,
0.0010907357,
0.036074888,
0.0147382068,
0.0612695515,
0.0442755036,
-0.0021249691,
-0.0439557135,
-0.0457655303,
-0.0005265717,
-0.037847843,
0.0102676358,
0.0357759036,
0.0057900697,
-0.0109113185,
-0.0029586384,
-0.0331265852,
0.0028677853,
0.0172515195,
0.0068516387,
0.0339371189,
0.0175439529,
-0.0021110473,
-0.0355169736,
0.0212215465,
0.0095045315,
0.0126553709,
0.0310041402,
0.0136758704,
-0.0762185678,
0.0371677652,
-0.014091609,
0.0116759548,
-0.0328632668,
-0.0380286239,
0.0358598344,
-0.0070190644,
0.0557897985,
0.0334935896,
0.0173922107,
0.0643643737,
0.0588321388,
0.0384358354,
-0.0528874546,
0.0049912278,
-0.0549550913,
-0.0070973579,
-0.0425960347,
-0.0072868909,
0.0397249982,
0.0472434647,
0.0317057483,
0.0193961915,
-0.1117850542,
0.0215164348,
-0.0155784236,
-0.0214326493,
-0.0066598272,
0.0013174933,
-0.0324048959,
0.024993673,
-0.0328164995,
-0.0747081116,
0.0594515577,
-0.0193508472,
0.0084503973,
-0.0245879665,
-0.0103069739,
-0.0295525435,
0.0182920806,
0.0259249993,
0.0450243428,
-0.0396431051,
-0.0440374836,
-0.069867149,
0.0254840944,
0.0436675064,
-0.0077830488,
0.038562078,
-0.0196176954,
0.0127429543,
-0.001539799,
-0.0367968939,
-0.0675992519,
-0.0339817479,
0.0198969338,
-0.0248808507,
0.0238449462,
-0.039475657,
-0.0351329744,
0.0216666553,
0.0187179502,
0.0495152548,
-0.016797401,
-0.0159137007,
-0.0357134789,
0.0088968957,
0.0874864906,
-0.0202416722,
-0.0102505442,
-0.047249075,
0.0440738611,
0.0265999846,
0.0010955745,
0.0016647219,
-0.0274716504,
0.0119255763,
0.0401615351,
-0.0293218028,
-0.0004999338,
0.0326367505,
0.0322405472,
-0.0286618415,
-0.0056949747,
0.1241336688,
0.0424998961,
-0.0591623187,
0.0012803091,
-0.0169653222,
0.039462924,
0.02263592,
0.0494232439,
-0.0553175695,
0.0016822958,
-0.0325153507,
0.0320521519,
0.0368406959,
0.0009242509,
-0.0342450179,
-0.1030735821,
0.0547126122,
0.0323114917,
-0.0184048489,
-0.0240591429,
-0.0344541185,
-0.0089236675,
0.0144732287,
0.0031456689,
0.0117165186,
-0.0250728745,
0.0039914674,
0.0190389827,
0.0246828645,
-0.0800397769,
-0.0364744812,
-0.0276138652,
0.0620052889,
0.0577877946,
0.0516849719,
0.0090782847,
-0.0278132837,
-0.0278181415,
0.0156613551,
-0.011291611,
-0.0117563661,
0.0230335016,
-0.0441687405,
-0.0034041391,
0.0060812817,
0.0556567162,
-0.0009680244,
-0.0159088224,
-0.001953115,
0.0791794732,
-0.0255031344,
0.004681658,
-0.0100633362,
0.0095261196,
0.0439991467,
0.0549268499,
0.0038908764,
-0.0713812187,
-0.0478568375,
-0.0230632555,
0.0105582299,
-0.0340045653,
-0.0001325662,
-0.0183795225,
-0.0313766524,
0.0366476327,
-0.0910053179,
-0.0401560776,
0.0443958007,
0.031720601,
0.0377365947,
-0.0073186276,
-0.014664147,
-0.0433165319,
-0.0341638066,
-0.0144434525,
-0.0071030632,
-0.0004313671,
-0.0558532253,
-0.0340585187,
0.0206014719,
-0.0700558797,
-0.0058770273,
-0.0334566757,
-0.0422614887,
-0.0944869667,
-0.0263132509,
0.0506814905,
0.0079098893,
0.0719609931,
0.0188143197,
-0.0490948483,
0.0973647311,
0.0029206222,
-0.0401999205,
0.0097679505,
0.0106424447,
0.0177941341,
-0.0217731986,
0.0239855368,
-0.062717773,
-0.0046829199,
-0.0070534353,
0.0705926567,
0.0034629656,
-0.0055412287,
0.0116252368,
-0.040044345,
-0.0253593735,
-0.0691430867,
-0.0007992667,
-0.0865402743,
-0.0039422591,
-0.0292759258,
-0.0085656783,
-0.0473245569,
0.0080623645,
0.0227604527,
0.0001572959,
-0.0047280458,
0.0219807755,
-0.0204120651,
0.0431304201,
-0.0103242779,
0.0097950036,
-0.0181096606,
0.011074312,
0.0621995889,
0.0234656334,
-0.0086863069,
0.0407185182,
-0.0405790433,
0.0028928723,
-0.0129244737,
0.0755132437,
0.0353663377,
0.0034118027,
-0.0139660016,
0.0270966943,
-0.023492353,
-0.0255118497,
-0.0385707468,
0.0224380419,
0.0482000411,
-0.0356332995,
0.0555953346,
-0.0763180703,
-0.0057256944,
-0.0049875677,
0.0380940996,
-0.0026676555,
-0.0514741354,
0.0250393748,
0.0377415232,
-0.0325256102,
0.023305323,
0.0845244899,
0.0544845872,
0.0268637836,
-0.021392161,
0.0033453982,
-0.0408918373,
-0.008942049,
0.0546785481,
0.0419713072,
-0.0310389549,
0.0230359156,
-0.0192056429,
-0.0604473613,
-0.0335152186,
0.0227858815,
-0.0570648313,
0.0032240651,
0.0310703088,
0.0364778712,
-0.0260862764,
0.0199468508,
-0.0598831959,
-0.0102000227,
-0.0571952909,
-0.0491307825,
-0.0590812638,
-6.485790815e-33,
-0.0204873458,
-0.0542751886,
0.054024417,
0.0007057269,
-0.0464135744,
0.0119100576,
-0.0279385727,
-0.0109483292,
0.0644587353,
0.0211012531,
-0.0011574851,
-0.0151300523,
-0.0069877058,
0.0559410155,
-0.0227933116,
-0.0053504147,
0.0682587028,
-0.0193392504,
-0.0359094515,
0.0422703512,
0.0238527097,
0.0282371584,
-0.0135366498,
-0.0345272496,
-0.0140523212,
-0.0285033006,
-0.020497568,
0.0121548586,
0.034201134,
0.0534624867,
-0.044019945,
0.0126213348,
0.0021412878,
-0.0217275955,
-0.0184033364,
0.0278617181,
-0.0453054868,
-0.034188576,
0.0021237326,
-0.0158560574,
-0.0563114993,
-0.0160434712,
-0.0094218981,
-0.0285894293,
0.0058928123,
0.0233048201,
0.0378489196,
-0.0093010487,
-0.0203252099,
-0.0105429431,
-0.0810031742,
0.0098219989,
-0.00123399,
0.0013155134,
0.0161557104,
0.0083989929,
0.0148340855,
0.0252804607,
-0.0485659167,
0.0484811924,
-0.0135701681,
0.0268903468,
0.0653770864,
0.073012203,
0.0174422953,
0.0649184659,
-0.0447925515,
-0.0175415203,
-0.0261143949,
-0.0195453148,
0.0080702445,
0.0102796666,
-0.0182493404,
0.0124460515,
-0.0339239202,
-0.0268478431,
-0.0294458959,
0.0221277867,
0.0012171958,
0.0004108823,
0.0529542193,
0.0432813354,
-0.0288898461,
-0.0224354807,
-0.0543578304,
0.0082325395,
-0.0051341467,
-0.0845674276,
0.0543665998,
0.0491571985,
-0.0460076444,
-0.088997744,
0.0181064419,
-0.0239200797,
-0.0054370905,
0.0239991583,
-0.0492979512,
0.0210385229,
-0.0069275741,
0.0103340345,
-0.0699618235,
-0.0299258642,
0.0249393694,
-0.0381584615,
-0.0133078694,
0.0069558895,
-0.0018419005,
0.0653344765,
-0.0656722263,
0.0259074885,
-0.0184287932,
0.0031205781,
0.0051519303,
0.0207701642,
0.0222317316,
-0.0075264103,
-0.0296436083,
0.0435253903,
0.0114070866,
-0.0049623833,
0.0001270552,
-0.0564860851,
-0.021296097,
0.0433387123,
-0.0636822358,
0.0294746608,
-0.0430899113,
0.046153266,
0.0433819108,
-0.1012205482,
-0.0148671353,
0.0044060829,
2.767e-7,
0.0012594761,
0.0413777418,
0.0089399759,
0.0291137993,
0.0152751086,
0.0180624109,
0.0004839746,
0.0285406001,
-0.006480095,
-0.0167682748,
0.0423691496,
0.0349143818,
0.0249431711,
-0.0104777422,
-0.0753846318,
0.0652287304,
-0.0695804507,
-0.0086611714,
-0.0356950462,
-0.0434653647,
0.0267254319,
0.069950588,
0.0264061857,
0.0096297422,
-0.0071154414,
-0.0432366207,
0.0809473544,
-0.0056309998,
-0.0270306021,
0.007254635,
0.0187213607,
0.0201622564,
-0.0508596003,
-0.0361071005,
-0.0374162346,
0.0179443453,
0.0069190133,
0.0855874717,
0.0136082321,
-0.0032857182,
-0.0030707277,
0.0091146538,
-0.0137190251,
-0.0247407705,
0.0404914767,
-0.0509724915,
-0.0211110283,
-0.0295751654,
-0.0024527258,
-0.0058437623,
-0.0185782276,
-0.048955664,
-0.0362032726,
0.0111301783,
0.0090824515,
0.0132027688,
0.0073741763,
-0.0821765065,
0.0013687038,
0.0084691411,
-0.0549613833,
0.0037350028,
-0.0581178479,
0.0316408984,
0.0445053391,
-0.0375183783,
-0.0072417138,
2.935281255e-34,
0.0228306018,
-0.0263899788,
-0.0048081093,
-0.0201569945,
0.0051914211,
0.0135539724,
0.0040702084,
0.015346081,
0.0664209574,
-0.0830439702,
-0.028408723
] |
http://arxiv.org/abs/2405.15232v2
|
http://arxiv.org/pdf/2405.15232v2
|
DEEM: Diffusion Models Serve as the Eyes of Large Language Models for Image Perception
| 2024-05-24T05:46:04 | 2024-07-03T17:30:53 |
Run Luo, Yunshui Li, Longze Chen, Wanwei He, Ting-En Lin, Ziqiang Liu, Lei Zhang, Zikai Song, Xiaobo Xia, Tongliang Liu, Min Yang, Binyuan Hui
|
cs.CV
|
cs.CV, cs.CL
|
The development of large language models (LLMs) has significantly advanced
the emergence of large multimodal models (LMMs). While LMMs have achieved
tremendous success by promoting the synergy between multimodal comprehension
and creation, they often face challenges when confronted with
out-of-distribution data. This is primarily due to their reliance on image
encoders trained to encode images into task-relevant features, which may lead
them to disregard irrelevant details. Delving into the modeling capabilities of
diffusion models for images naturally prompts the question: Can diffusion
models serve as the eyes of large language models for image perception? In this
paper, we propose DEEM, a simple and effective approach that utilizes the
generative feedback of diffusion models to align the semantic distributions of
the image encoder. This addresses the drawbacks of previous methods that solely
relied on image encoders like ViT, thereby enhancing the model's resilience
against out-of-distribution samples and reducing visual hallucinations.
Importantly, this is achieved without requiring additional training modules and
with fewer training parameters. We extensively evaluated DEEM on both our newly
constructed RobustVQA benchmark and another well-known benchmark, POPE, for
object hallucination. Compared to the state-of-the-art interleaved content
generation models, DEEM exhibits enhanced robustness and a superior capacity to
alleviate model hallucinations while utilizing fewer trainable parameters, less
pre-training data (10%), and a smaller base model size.
|
DEEM: Diffusion Models Serve as the Eyes of Large Language Models for Image Perception: The development of large language models (LLMs) has significantly advanced
the emergence of large multimodal models (LMMs). While LMMs have achieved
tremendous success by promoting the synergy between multimodal comprehension
and creation, they often face challenges when confronted with
out-of-distribution data. This is primarily due to their reliance on image
encoders trained to encode images into task-relevant features, which may lead
them to disregard irrelevant details. Delving into the modeling capabilities of
diffusion models for images naturally prompts the question: Can diffusion
models serve as the eyes of large language models for image perception? In this
paper, we propose DEEM, a simple and effective approach that utilizes the
generative feedback of diffusion models to align the semantic distributions of
the image encoder. This addresses the drawbacks of previous methods that solely
relied on image encoders like ViT, thereby enhancing the model's resilience
against out-of-distribution samples and reducing visual hallucinations.
Importantly, this is achieved without requiring additional training modules and
with fewer training parameters. We extensively evaluated DEEM on both our newly
constructed RobustVQA benchmark and another well-known benchmark, POPE, for
object hallucination. Compared to the state-of-the-art interleaved content
generation models, DEEM exhibits enhanced robustness and a superior capacity to
alleviate model hallucinations while utilizing fewer trainable parameters, less
pre-training data (10%), and a smaller base model size.
|
[
-0.0025333147,
0.0533460677,
0.013098551,
0.0531568229,
-0.0171251725,
0.0146264443,
0.0071252971,
-0.0096599841,
-0.0408340916,
-0.0330913924,
-0.0439983085,
-0.082164146,
0.0009469991,
0.0124060521,
0.0412569344,
-0.073363021,
-0.021438783,
-0.0377142578,
-0.0765921846,
-0.0227229483,
0.0061204485,
-0.0026355702,
0.0183395371,
0.0142941661,
-0.0363849737,
-0.0409329832,
-0.0249749962,
0.0443041846,
0.0329575613,
-0.0277177449,
-0.0130391307,
0.0473348536,
0.0324817896,
-0.0040286537,
0.0000023349,
-0.0007429611,
0.0195125788,
-0.0076770051,
0.0444019251,
0.0206936039,
0.0434150025,
-0.0304369014,
-0.00069721,
-0.0289380364,
0.0219083466,
0.0025843459,
0.0954881385,
0.0487351231,
0.073058024,
0.0425535925,
-0.0326474309,
-0.0880190581,
0.014319472,
-0.009312015,
0.0431351028,
-0.0032599142,
0.0306583736,
0.0658832863,
-0.011788574,
-0.0081559988,
-0.079056859,
0.0203091912,
0.0406859294,
0.033512149,
0.049687352,
0.0467084348,
0.0232500453,
-0.0582008474,
-0.0145570859,
0.0659200549,
-0.0751922131,
-0.0095315361,
-0.0154971154,
-0.0091468804,
-0.0011819204,
0.0414873883,
-0.0227488633,
0.0084179016,
-0.0148889013,
0.0300383698,
-0.0655200034,
-0.0091724657,
0.0450020544,
-0.015973255,
-0.0307649523,
-0.0112614101,
0.0457196832,
-0.0157679431,
-0.0020714391,
0.0193890762,
-0.0593693145,
-0.0255955104,
0.0388046131,
-0.0048184674,
-0.014184528,
-0.0250949226,
-0.0430204198,
-0.0067469892,
-0.0085233962,
-0.065538235,
0.0404533483,
0.007931469,
-0.0579182543,
0.0545763448,
-0.0297109801,
0.0424498282,
-0.061400041,
0.0491727479,
-0.0493045971,
0.0055204695,
-0.013424308,
-0.0042357985,
-0.0200263169,
-0.0153945368,
-0.0252458472,
-0.0337235183,
-0.0315329358,
0.0009472019,
-0.0120339785,
-0.0005650399,
0.0053310241,
0.0285958312,
-0.0161763597,
-0.0420039035,
-0.0069713858,
0.0369291641,
0.0096176779,
0.0012395954,
0.0128022786,
-0.0343843661,
-0.0172173157,
0.0294368025,
0.0098952204,
-0.0154327983,
0.0024409262,
0.0790805817,
-0.0320493951,
-0.0030571059,
-0.0400388651,
-0.0000324493,
-0.0087672742,
-0.0428937562,
0.0765272155,
-0.0593345761,
0.0223149788,
-0.0193917081,
0.0300319754,
-0.0425441302,
0.0114509352,
-0.0021610269,
-0.0257153325,
0.0588550195,
-0.0268875528,
0.0194609892,
0.0302744024,
0.0152029134,
0.003945814,
0.0803267211,
0.0232720133,
0.0517942235,
-0.016246805,
-0.0119741168,
0.0323851295,
0.0291560423,
-0.0048597939,
0.0457360148,
-0.0456877276,
-0.0536527634,
0.0158067252,
0.0315108486,
0.0790379345,
0.0797612816,
-0.0541806705,
-0.0392479897,
0.080362618,
0.0005614454,
-0.0125255557,
-0.0215059482,
0.0646092594,
0.0085461158,
-0.0320852026,
0.0205051545,
0.0033195012,
0.0614027865,
0.0431885459,
0.0124065075,
0.0604465567,
0.0271516331,
-0.0601926073,
-0.0304898378,
-0.0024388093,
-0.0311622787,
0.0368361622,
0.0087851221,
0.0273431297,
0.0194579158,
-0.0606425554,
0.0126048736,
0.0114650382,
-0.0135687301,
0.0037694201,
-0.0427747183,
0.0272921752,
0.0807279944,
0.0371945389,
0.0388687104,
0.041985102,
-0.0092754737,
0.0369201601,
0.0737098679,
-0.0279865973,
0.0242237411,
0.0364052653,
-0.0538627878,
-0.0378369465,
-0.0176527612,
0.0266089588,
0.0222740471,
-0.0369193852,
0.0091150887,
0.007887234,
-0.028050866,
-0.0135565326,
0.0334263556,
-0.0015479983,
0.0140585313,
-0.023860408,
-0.0078870589,
0.0217646863,
-0.0150497528,
-0.0193130653,
0.0148670673,
-0.0130324103,
-0.0359397717,
0.0006672985,
0.010170579,
-0.0371083729,
-0.0253402255,
0.0151981628,
-0.0615195557,
0.020543633,
-0.0097601414,
-0.0282393135,
-0.0171646513,
0.024441354,
-0.0239886045,
0.008596669,
-0.044236958,
0.0248699151,
0.0321597718,
0.0446749106,
0.0288456939,
0.0728596747,
-0.0603181794,
0.0107540051,
0.0035386898,
0.05599989,
-0.0113207595,
0.0129394885,
-0.0154934106,
0.0040053888,
0.0209326446,
0.0038402562,
-0.0240518581,
-0.0121703316,
-0.061823573,
-0.0002382721,
-0.0531517826,
-0.0005897524,
-0.0429510586,
-0.0469065234,
-0.0025250923,
0.0066266786,
0.0128680393,
0.0045149578,
-0.0127344802,
0.0403118357,
-0.0109487921,
0.0365374759,
-0.0138619011,
-0.0063690902,
-0.0072442438,
-0.0284811407,
-0.0093446542,
0.0430095531,
0.0754551068,
-0.0251081735,
-0.0439403169,
-0.0531113148,
0.0268282406,
-0.0189722348,
0.0136321094,
-0.0095238267,
-0.0326028354,
-0.0035307065,
-0.0252710525,
0.0683475211,
0.0670260489,
0.0309896693,
0.0398282595,
0.02708463,
-0.0283318926,
-0.0112382462,
-0.016759254,
-0.0100604612,
-0.0126648517,
0.0002188094,
-0.0336629003,
-0.0346211642,
0.0576327182,
0.0076649128,
0.0268904064,
-0.0850050598,
0.0117457937,
0.0094444221,
-0.0628878623,
-0.0289308969,
-0.0018798521,
-0.0194104277,
-0.0008109045,
-0.0665995032,
-0.0494972467,
0.0330443233,
-0.040068619,
-0.0065058437,
0.0017133717,
0.0292667411,
0.033193782,
-0.000529529,
0.0366716273,
0.0529768132,
-0.017606115,
0.0122645963,
-0.0434558354,
0.016008867,
-0.0042554834,
-0.0151455617,
-0.0059214071,
-0.0285782069,
-0.0186531246,
-0.007179209,
0.0142595405,
-0.0260721054,
-0.0187143907,
0.0001216974,
0.0019933784,
-0.054351382,
0.0334525518,
-0.025005864,
0.0377004817,
-0.0029569827,
-0.002448638,
0.0020975387,
-0.0275053959,
-0.0048506437,
0.0357315354,
0.0598020554,
0.0029374145,
-0.0316572785,
-0.0976269022,
0.0216170326,
0.0596538,
-0.0185881704,
0.0039418107,
-0.0391532518,
0.0370002054,
-0.0061023524,
0.0654490292,
-0.0199673213,
0.0348845646,
0.1028235033,
-0.0350615233,
-0.0232726969,
0.0570699275,
0.0058783181,
-0.0681711733,
-0.0028919585,
0.0010270113,
0.0666473359,
0.0088795731,
-0.0360407196,
-0.0774407163,
-0.040073622,
-0.034213189,
-0.0321715996,
0.0291218013,
0.0255848058,
-0.0063100043,
-0.0597186014,
0.0931355059,
0.048233781,
-0.0000467753,
-0.029711917,
0.0057864403,
-0.0073082265,
0.0339438692,
0.0129536223,
-0.0034735634,
0.0203389674,
-0.0107421651,
0.0383389629,
0.0646598116,
-0.0701119304,
-0.0104880286,
-0.0340161622,
-0.0113304732,
0.0881725773,
0.001291893,
0.0117343571,
-0.0914596319,
0.0477182381,
-0.0133560505,
-0.0062091365,
0.0175485257,
0.0225252863,
-0.0120599186,
-0.021425439,
0.020873962,
0.0385229588,
-0.0227610152,
-0.0295640025,
-0.025363246,
0.0243709181,
-0.0304010287,
-0.0453066975,
-0.0635951087,
0.0091643743,
0.0079479059,
0.0523726977,
0.0424798205,
-0.0363915376,
-0.0282482505,
-0.0132004414,
0.0396738648,
0.0004608481,
-0.0404636972,
0.0424687751,
0.0430340171,
-0.0736211687,
-0.0694000125,
-0.0138545297,
-0.0164304245,
0.0303925052,
-0.0135580404,
0.0072773886,
-0.0248147976,
-0.0210851226,
-0.0423156545,
0.0520989597,
-0.008962051,
-0.0089164497,
-0.0280835014,
-0.011562746,
0.0272247531,
-0.0297887679,
-0.0059731463,
0.0529982299,
-0.1044781059,
-0.0458528474,
-0.0218277629,
0.0466644801,
0.0154838171,
0.0566898584,
-0.0052310904,
-0.030682113,
0.0579551011,
0.0093783764,
0.0369942673,
0.0309150722,
-0.0087390719,
-0.004110713,
-0.0050500012,
-0.0004359339,
-0.0559594557,
0.0304475036,
-0.014944924,
0.0171372686,
0.0108761005,
0.0231311936,
0.0289234016,
-0.0628690496,
0.0031957692,
-0.0460003503,
0.0418301299,
-0.0187646486,
-0.0091348477,
-0.0056385505,
-0.0122282729,
-0.0686111525,
-0.0044418601,
0.0499493107,
0.0077363155,
-0.0046547344,
-0.0040034698,
0.0191795137,
0.0199014172,
-0.0356950723,
-0.0240803789,
0.0057902923,
-0.0217073783,
0.0072788177,
-0.0002781614,
0.0053045452,
-0.0138635756,
-0.0285960026,
0.0220935009,
0.0229187831,
0.0629692078,
-0.0019666317,
-0.0272524022,
-0.0401444025,
0.0396217033,
0.0071619144,
-0.0474011116,
-0.0780790374,
0.0678365529,
0.0539778583,
0.0214004833,
0.0488372073,
-0.0264138114,
-0.0004473538,
-0.0432720631,
0.043727126,
0.0321236178,
-0.042789083,
0.0487583838,
0.0045788563,
0.0065787607,
-0.031521704,
0.0699904189,
-0.0091750175,
0.0241110828,
-0.0232277811,
-0.0121563338,
-0.0277316906,
-0.0303752031,
-0.0333427414,
0.0052487855,
0.0203686636,
0.0383773483,
0.0014372922,
-0.0539128743,
-0.0432792939,
0.0139729092,
-0.008644009,
0.0249518696,
0.0524913892,
0.0570360199,
-0.0297462996,
-0.0242790673,
-0.014846324,
-0.0312634669,
0.0174622443,
-0.0086532673,
0.0311972555,
-6.672402013e-33,
-0.0171348006,
0.0159465913,
0.0345132686,
-0.0603490584,
-0.0377318747,
-0.0431187786,
-0.0452355146,
-0.0059737251,
0.0013236118,
0.0274567883,
0.0351514108,
0.0345772095,
0.0072187432,
0.0365722999,
-0.0037808064,
0.0024855162,
0.0653551519,
-0.0295314472,
-0.0209662486,
0.0100726532,
-0.0070366836,
0.0030940191,
0.0402208343,
-0.0293378793,
-0.0211751815,
-0.0068504424,
-0.032338541,
-0.0098469611,
0.0422623642,
0.0503451526,
-0.0712445974,
0.0203071609,
0.0069841342,
-0.0063609136,
-0.0416890271,
-0.0113817053,
-0.00719987,
0.0208053775,
-0.0190151054,
0.0196794495,
-0.097980395,
-0.049959287,
0.0822070837,
0.0149874743,
-0.0072413161,
-0.0033271736,
0.0342180952,
-0.0495091155,
-0.0121265538,
-0.002841471,
-0.0397321545,
0.0116542559,
-0.0295525678,
-0.0260996819,
0.0368979238,
0.0181466602,
-0.0113701718,
0.0364843085,
-0.047239475,
-0.0000505642,
-0.0061920839,
0.0182777178,
0.0627104715,
0.054074727,
0.0478278026,
0.0621522516,
0.0414024629,
-0.0174787156,
0.0203174986,
-0.008017011,
0.0222154595,
0.0645951033,
-0.0100120883,
0.0251812637,
0.0465574898,
0.0136930309,
-0.0310005266,
0.0409337282,
-0.1137102619,
0.0278559588,
0.0732451081,
0.0320854411,
0.0013364301,
-0.016735794,
-0.0137002775,
-0.027943939,
-0.0353335291,
-0.0345540605,
0.0472764783,
0.0054576746,
0.0135856038,
-0.027633179,
0.0807223693,
-0.0142763741,
-0.0078651272,
-0.0029530788,
-0.0624458455,
0.0727338567,
0.0059530889,
-0.0086964481,
0.0185595341,
-0.0101822745,
0.0524052046,
0.0076089506,
-0.0149164069,
0.0272100233,
-0.0149014881,
0.0038874513,
-0.0155362133,
-0.0113278311,
-0.0590124875,
-0.0071132067,
0.0004864024,
0.0329245254,
0.0208275318,
-0.0010930422,
-0.0455349609,
0.017351361,
-0.0284170974,
0.0153475069,
0.000368032,
-0.0180522595,
-0.0221436806,
0.041004274,
-0.0339328609,
0.0185911935,
-0.0432370007,
-0.0260689426,
0.0614332594,
-0.0508355312,
-0.0433799922,
0.0400807075,
2.923e-7,
0.0157014411,
0.0076967096,
-0.0298585221,
0.0016272331,
0.0242410023,
0.0159578118,
-0.0078506498,
0.0374369584,
0.0093137017,
-0.1198167503,
0.0035191379,
0.0884624049,
0.0363692194,
-0.0275166761,
-0.0482589677,
0.0072688637,
-0.0128062349,
0.0219212789,
-0.0421084836,
-0.04416316,
0.0273885075,
0.088862136,
-0.0366325341,
-0.0283465516,
0.0701986402,
0.0082052229,
0.112474151,
-0.0242115706,
-0.0045369905,
-0.0023676639,
0.0060189795,
0.0219263509,
0.0111045185,
0.017878117,
-0.0016322411,
-0.0360580236,
0.0547652282,
0.0130896382,
-0.0336799808,
-0.0572741106,
-0.0062165498,
0.0414736532,
-0.0118578011,
-0.0397777259,
0.0558828115,
-0.0236872006,
-0.0132656423,
0.000091942,
-0.0793632343,
-0.0151040731,
0.0756507814,
-0.0229699053,
-0.0326004252,
-0.0061704805,
0.0154142734,
0.0178845525,
-0.0395795815,
-0.0491160154,
-0.0091589577,
0.0064458586,
-0.0543285944,
-0.0397724807,
-0.0352429934,
-0.0242550485,
0.0094822384,
-0.0102467285,
-0.0363376737,
2.298957005e-34,
0.0457569323,
-0.0123350499,
0.0380370319,
0.0197336208,
0.0065861312,
-0.0274784062,
-0.0155510297,
0.0057966639,
-0.0278838482,
-0.0528127663,
-0.0172510184
] |
http://arxiv.org/abs/2405.15216v1
|
http://arxiv.org/pdf/2405.15216v1
|
Denoising LM: Pushing the Limits of Error Correction Models for Speech Recognition
| 2024-05-24T05:05:12 | 2024-05-24T05:05:12 |
Zijin Gu, Tatiana Likhomanenko, He Bai, Erik McDermott, Ronan Collobert, Navdeep Jaitly
|
cs.LG
|
cs.LG, cs.CL, cs.SD, eess.AS
|
Language models (LMs) have long been used to improve results of automatic
speech recognition (ASR) systems, but they are unaware of the errors that ASR
systems make. Error correction models are designed to fix ASR errors, however,
they showed little improvement over traditional LMs mainly due to the lack of
supervised training data. In this paper, we present Denoising LM (DLM), which
is a $\textit{scaled}$ error correction model trained with vast amounts of
synthetic data, significantly exceeding prior attempts meanwhile achieving new
state-of-the-art ASR performance. We use text-to-speech (TTS) systems to
synthesize audio, which is fed into an ASR system to produce noisy hypotheses,
which are then paired with the original texts to train the DLM. DLM has several
$\textit{key ingredients}$: (i) up-scaled model and data; (ii) usage of
multi-speaker TTS systems; (iii) combination of multiple noise augmentation
strategies; and (iv) new decoding techniques. With a Transformer-CTC ASR, DLM
achieves 1.5% word error rate (WER) on $\textit{test-clean}$ and 3.3% WER on
$\textit{test-other}$ on Librispeech, which to our knowledge are the best
reported numbers in the setting where no external audio data are used and even
match self-supervised methods which use external audio data. Furthermore, a
single DLM is applicable to different ASRs, and greatly surpassing the
performance of conventional LM based beam-search rescoring. These results
indicate that properly investigated error correction models have the potential
to replace conventional LMs, holding the key to a new level of accuracy in ASR
systems.
|
Denoising LM: Pushing the Limits of Error Correction Models for Speech Recognition: Language models (LMs) have long been used to improve results of automatic
speech recognition (ASR) systems, but they are unaware of the errors that ASR
systems make. Error correction models are designed to fix ASR errors, however,
they showed little improvement over traditional LMs mainly due to the lack of
supervised training data. In this paper, we present Denoising LM (DLM), which
is a $\textit{scaled}$ error correction model trained with vast amounts of
synthetic data, significantly exceeding prior attempts meanwhile achieving new
state-of-the-art ASR performance. We use text-to-speech (TTS) systems to
synthesize audio, which is fed into an ASR system to produce noisy hypotheses,
which are then paired with the original texts to train the DLM. DLM has several
$\textit{key ingredients}$: (i) up-scaled model and data; (ii) usage of
multi-speaker TTS systems; (iii) combination of multiple noise augmentation
strategies; and (iv) new decoding techniques. With a Transformer-CTC ASR, DLM
achieves 1.5% word error rate (WER) on $\textit{test-clean}$ and 3.3% WER on
$\textit{test-other}$ on Librispeech, which to our knowledge are the best
reported numbers in the setting where no external audio data are used and even
match self-supervised methods which use external audio data. Furthermore, a
single DLM is applicable to different ASRs, and greatly surpassing the
performance of conventional LM based beam-search rescoring. These results
indicate that properly investigated error correction models have the potential
to replace conventional LMs, holding the key to a new level of accuracy in ASR
systems.
|
[
-0.0229860358,
0.0473565497,
0.0322479047,
0.0343724973,
0.0149305789,
-0.0216512736,
0.0074501764,
0.0147095239,
-0.0067012608,
-0.0048413058,
-0.0461241156,
-0.0223366283,
0.0395640694,
0.0091748303,
0.0022590198,
-0.0073187398,
-0.0009969549,
-0.070066765,
0.0372348875,
-0.0229274724,
0.0047497293,
0.0110170916,
-0.0209850855,
0.0223829709,
-0.0018521779,
0.0523635373,
0.013570183,
0.0583665185,
0.0126779675,
-0.0030040718,
-0.0160017796,
-0.0057026958,
0.0421172753,
0.0165014789,
0.0000023218,
-0.0151135027,
-0.0306285545,
-0.0009489287,
-0.0556154586,
0.0043543708,
0.0251218081,
-0.0069825742,
-0.0017906275,
0.0429941565,
0.0064109312,
-0.03816038,
0.0314230584,
0.0246035773,
0.0155560086,
0.0805141851,
-0.0010211667,
-0.0665701777,
0.0158284772,
-0.0030794509,
-0.0154172899,
-0.0045911511,
-0.0118312454,
0.0169958137,
-0.008238811,
-0.020096492,
0.003345686,
0.0606029406,
-0.0186988898,
-0.0346619003,
0.0027527232,
-0.0185865201,
0.009419906,
-0.0697316676,
-0.0243805237,
0.0353210978,
0.0026206379,
-0.052764263,
0.0113714458,
0.0114221498,
-0.0328918099,
0.0113996677,
0.0007231378,
0.0329209268,
0.0060381382,
0.0326682217,
0.0042903777,
-0.0052551716,
0.0486912169,
-0.0210784581,
-0.0098419655,
-0.0179278664,
-0.0051304251,
-0.0406885035,
-0.0133729186,
-0.0205031112,
0.0095938556,
-0.0259379912,
0.0834273472,
0.0202136431,
-0.0242281053,
-0.0101382853,
-0.0278075617,
-0.0841081962,
-0.0249224994,
-0.0123579493,
-0.0023864382,
0.0513959974,
-0.0266635437,
-0.0086580561,
-0.0224906076,
0.0488398448,
0.0295867696,
-0.0008822882,
-0.0515636913,
-0.0164086241,
0.0202567708,
0.0194390919,
-0.0245559495,
0.0307355169,
-0.0453281328,
0.0120081622,
-0.0450115353,
0.0024042286,
0.0017292858,
-0.00030514,
0.0319018997,
0.0026526211,
-0.0607816018,
-0.0576949976,
0.0378202498,
0.0378451645,
0.0882148221,
0.0505496114,
-0.0119055165,
-0.0747501254,
-0.0326209366,
0.0283454712,
0.0435356684,
0.0101279216,
0.0183229595,
0.0547661223,
0.0121265715,
-0.0670886263,
-0.0365041941,
0.0214976128,
-0.0159076322,
-0.0281266272,
-0.0192100499,
-0.0615031645,
-0.0056743207,
0.0090845525,
0.0123691252,
-0.0555583462,
0.0253865942,
-0.0223288741,
-0.0143068368,
0.0778025538,
-0.0314472429,
0.0309007261,
0.0716990158,
0.000665468,
0.0368030407,
0.0200067461,
0.0922587514,
-0.0357611589,
-0.0066986997,
-0.0333661586,
0.0458123498,
0.054505948,
0.0002354015,
0.0033997754,
-0.0335340537,
-0.0407078378,
-0.012916049,
0.0198505521,
0.0906585529,
0.0796089768,
-0.0386863612,
-0.000225075,
0.0775504485,
0.0161326565,
0.0604128093,
-0.0326234996,
-0.0022282465,
0.0173586868,
0.0137433494,
-0.0016640795,
0.0064436463,
-0.0325193517,
-0.0446577854,
0.0190987773,
0.0794618577,
0.0102660945,
-0.0205855798,
-0.0237794854,
-0.023385236,
-0.0395217687,
0.0143748801,
0.05473838,
0.0722286701,
0.0330911167,
-0.0513281673,
0.0573970601,
-0.0060666804,
-0.0861572102,
0.0273535196,
-0.0581974946,
0.0123593928,
0.0302443188,
0.0662949309,
-0.0181199573,
-0.0069636167,
-0.0367843024,
-0.0686222762,
0.0689756796,
0.005984338,
0.0361883081,
0.0302874483,
-0.0992584154,
-0.015191948,
0.0627057105,
-0.0113520734,
0.0020966018,
-0.0289003383,
0.007742269,
0.002088614,
-0.0240987595,
-0.0270154122,
0.002337222,
-0.0228844229,
-0.0211563818,
0.0157573633,
-0.0180435795,
0.0628266037,
0.0541632585,
-0.0504523516,
-0.0199763533,
0.0519406237,
-0.0189266205,
-0.0184691995,
0.0144623527,
-0.0659995601,
-0.0589984916,
0.060185995,
-0.0210331455,
0.0084026707,
0.0528148562,
0.0114260493,
-0.0763936266,
0.0353872404,
-0.0536206849,
0.0185276568,
-0.0517276786,
0.0718476027,
0.0333566591,
0.0018117164,
0.0027771201,
-0.0302263498,
-0.0211625583,
-0.0057136361,
0.0037000442,
0.0081690811,
-0.0217485428,
-0.0699993148,
-0.0415757298,
0.0061783828,
-0.0078177089,
0.0277045183,
-0.0023843008,
0.0141976876,
0.0021273496,
0.0188004933,
-0.0677073747,
-0.0121236285,
-0.0291819796,
-0.009169559,
-0.0253753476,
0.0283659287,
0.0077509684,
-0.0412635319,
0.0049081496,
-0.0257573724,
0.0303999614,
-0.0407287814,
-0.0298403129,
-0.0100479126,
-0.0500009917,
-0.0277290083,
0.0205808505,
0.0193998013,
-0.0113450624,
-0.0182102844,
-0.0442293957,
-0.0692564696,
0.0944971815,
0.0116345724,
0.0355779715,
0.0115775196,
-0.0404741615,
-0.0127649009,
-0.0282441955,
0.0254912842,
-0.0291006137,
0.00073217,
-0.01221037,
0.0204708911,
-0.0260927547,
-0.0519990064,
0.0049660723,
0.0103913089,
0.0061349049,
-0.0149226403,
-0.0168071371,
-0.0064946744,
0.0423654467,
0.0345089696,
0.0422713496,
-0.0163501091,
-0.0329234973,
-0.0008822713,
-0.0355666429,
-0.0013033013,
-0.0550677665,
-0.0393305607,
0.0125434976,
-0.0072472626,
-0.0161707159,
0.0391840413,
-0.0733977482,
-0.0329616852,
-0.0466963239,
-0.0607936829,
0.0072493572,
-0.0427420437,
0.0074810013,
0.0011931157,
-0.0203755908,
-0.0470499396,
-0.0062438548,
-0.0160277784,
0.0190495513,
-0.0877164751,
0.0513683818,
-0.0552773066,
0.0308189131,
-0.0157108866,
-0.0231577624,
-0.0078225089,
-0.0043861074,
-0.0021071406,
0.0080401562,
-0.0038606983,
-0.0370275714,
-0.0450017229,
0.0324807875,
0.0312203597,
0.0820602924,
0.0196875129,
0.0066232947,
-0.0294053871,
0.0856380016,
0.0059801592,
0.0361135192,
0.0174176432,
-0.0017033643,
0.0094572678,
0.0086607998,
0.0366914272,
0.0026409875,
-0.0646728873,
-0.023572959,
0.0600903928,
0.0276797134,
0.010005964,
0.0116820494,
0.0657800362,
-0.0389317311,
-0.0091314744,
0.1067735925,
-0.022886999,
-0.030845359,
-0.0505652428,
0.0058347676,
-0.0244574193,
-0.0047106803,
-0.0668538362,
-0.0784728751,
-0.0194725934,
0.0821629837,
0.0657737851,
0.0103339693,
0.0401841141,
-0.0052573862,
-0.0953855738,
0.0569371507,
0.041785948,
-0.0292539019,
-0.0664428994,
-0.0697056726,
-0.0097118607,
-0.0267495327,
0.0831580237,
0.0445771888,
0.0015123466,
-0.0366869234,
0.0459826216,
-0.0438001603,
-0.1065335944,
-0.0602344498,
-0.0416672193,
0.1290224493,
0.0372898206,
0.0065444219,
-0.0136557948,
-0.0332322977,
0.019204339,
0.0057764342,
0.0078824433,
-0.016817579,
0.0029224011,
0.0128482748,
0.000621099,
0.0250307228,
0.0481978618,
-0.0214471929,
-0.0091772163,
-0.0248318538,
0.0146558983,
-0.0395549573,
0.01992804,
0.0175374933,
0.0416127294,
0.0027858622,
0.0722280368,
0.0116955386,
-0.0182203278,
-0.0019039148,
-0.0030830267,
0.0277278684,
-0.0162571147,
-0.0137684159,
0.0041165128,
-0.0230517332,
-0.0351623707,
0.0016405467,
-0.0053961328,
-0.0046666791,
-0.0142598087,
0.0226023812,
0.0022961055,
-0.0385819189,
0.0178308077,
-0.0044093761,
0.0130549558,
-0.0432855673,
0.0024351079,
-0.0458538942,
0.0332702138,
-0.0339708924,
-0.0536641702,
-0.0078735724,
0.0520334356,
-0.0447678976,
-0.0504710488,
0.0355854966,
-0.0002242756,
0.0259353556,
0.0156442747,
-0.0075343656,
-0.0685946196,
0.06073815,
0.0485180393,
-0.0739651471,
0.0436847396,
0.0323367044,
0.0147288023,
-0.0052911537,
0.0150187099,
-0.0462680198,
0.0368869342,
-0.0021031005,
0.0394448638,
-0.044678703,
-0.0768858716,
0.0096059004,
0.0289507676,
0.0620647408,
-0.0031717019,
0.0294568148,
-0.0311068948,
0.0206769072,
-0.0044675115,
0.0348970145,
-0.0124965226,
0.0002131373,
0.0222289301,
0.016141776,
-0.0258115549,
0.0042650262,
-0.0753061995,
-0.0030362087,
-0.0251798388,
0.0548990928,
0.019479923,
-0.0015535173,
0.0210909173,
-0.0077620964,
-0.0092164651,
0.0333779603,
0.0442468151,
0.0019099424,
0.0261738747,
0.0234902911,
-0.0120906895,
0.0109543949,
0.0077674547,
0.0093407733,
-0.0479260199,
-0.0247879606,
-0.0285060331,
0.0025706531,
0.0382315107,
0.0335009843,
0.0961818993,
-0.0106804706,
0.0296701584,
-0.0587927662,
-0.0044332542,
0.0008922853,
0.0345858075,
0.0007701507,
-0.0065677278,
-0.0170693658,
0.045056574,
0.0445310175,
-0.0264129341,
0.0394829661,
-0.0505814739,
0.0354317315,
-0.0274270643,
-0.0301959198,
-0.018892765,
0.0059673302,
-0.0441609658,
-0.0510582849,
0.0202627108,
-0.0569657236,
0.0449886881,
-0.0011497743,
0.0424865559,
-0.059602242,
0.0079929233,
-0.0207416359,
-0.0253039971,
0.023259908,
0.0206753034,
-0.0276343171,
-0.0040651667,
-0.0303038917,
-0.0191044621,
-6.541187456e-33,
-0.0302785616,
0.0070296172,
-0.0122266794,
0.022501329,
-0.0114453575,
0.0066604158,
-0.0394112617,
0.0040649679,
0.0017654768,
-0.0053965077,
-0.0192375835,
0.0441807471,
0.0451003127,
0.0117635187,
-0.0089758374,
0.0438192785,
0.1027668938,
-0.0319737978,
-0.0120037626,
-0.0372949094,
0.0310889035,
0.0549653098,
0.0192834903,
-0.0615438148,
-0.001285225,
0.0022969954,
-0.0393505879,
-0.005553131,
0.0399125069,
0.0147591736,
-0.0008392889,
0.0649849921,
-0.0213315748,
-0.022369843,
-0.0228568651,
-0.0113628162,
-0.0788515061,
0.0492439605,
-0.0085042659,
0.0563096181,
0.0129813431,
-0.0020542867,
0.0720815063,
0.0297813043,
-0.0903723985,
0.0092274239,
0.0023871073,
-0.0563366786,
-0.0200353712,
-0.0295889098,
0.035972897,
0.0345875807,
0.0022134231,
-0.0357977599,
0.0671904087,
0.0450941399,
-0.0394363776,
-0.0003402122,
-0.1020625234,
0.0398011841,
0.0318662375,
0.0296194348,
0.0284056999,
0.0090817912,
-0.0444836766,
0.017125763,
0.0351774208,
-0.0715101808,
0.0085530737,
-0.0513917357,
0.0333546773,
0.0182386469,
0.034222737,
0.0372612737,
0.0290276296,
-0.0194673315,
-0.0235942099,
0.0179865006,
-0.0096231457,
0.0529675819,
0.0452218242,
0.0673206821,
-0.0172119476,
-0.0112058381,
-0.0588225536,
-0.0192466043,
-0.0289418474,
0.0049879248,
0.0253065005,
0.0312511586,
0.0071073016,
-0.0513692498,
0.0411464013,
-0.018227838,
0.0594894066,
-0.0107551143,
-0.0011419911,
-0.0149911959,
-0.0310290735,
0.0014346909,
-0.039129708,
-0.0200147945,
0.0593824163,
0.0636990517,
-0.0127227055,
-0.0340422839,
-0.0510521606,
0.0767023861,
-0.0519925058,
0.0148366513,
0.0098536266,
0.0096392976,
0.0068681496,
0.0214143731,
0.0421337038,
-0.0320858173,
0.0051962375,
0.0069120945,
-0.0120578837,
0.0296838935,
0.0062345224,
0.0017494628,
0.0168097615,
0.0090427678,
-0.034929011,
0.0347380973,
0.0299897641,
-0.0449786484,
0.0153687922,
-0.0079320418,
-0.0007770655,
-0.0064841509,
3.05e-7,
-0.0501627848,
0.019785095,
0.0292464737,
-0.0020832298,
-0.0035307042,
0.0033982976,
-0.0009842183,
-0.039219413,
0.0188902225,
0.0138314376,
0.0103564039,
0.0066914712,
0.0884044468,
0.0052064997,
-0.0597854927,
-0.0336846486,
-0.063844271,
-0.0081438497,
0.0223816112,
-0.003842449,
-0.0059948307,
0.0433975719,
0.0031441462,
-0.0607651249,
0.0031790747,
-0.0170667712,
0.0271047913,
-0.0252100341,
-0.0357890911,
-0.0080436571,
0.0434799567,
0.0071082329,
-0.0030347561,
0.0119779557,
-0.0237911046,
0.0238381866,
0.0137858652,
0.0186726265,
-0.00346138,
-0.053146515,
-0.0112692602,
0.0601595566,
-0.0251464508,
-0.0013162192,
0.0199251901,
0.0338327065,
0.0153052546,
-0.0046446868,
-0.0602891818,
-0.0065065157,
0.011402051,
-0.0184749439,
0.0013409961,
-0.0220630467,
0.0218412299,
0.0533107556,
-0.0480484292,
-0.038217321,
0.0049875355,
0.0189711489,
-0.0526672006,
-0.0096604181,
-0.0049951319,
-0.0031179693,
0.0329148024,
-0.023207793,
-0.0034155897,
2.47533029e-34,
0.0522812344,
-0.0475526042,
0.0183450058,
0.0226315651,
0.002478987,
-0.0125695337,
-0.02624554,
0.0258225128,
-0.0129514467,
-0.0759829134,
-0.0314843468
] |
http://arxiv.org/abs/2405.17475v1
|
http://arxiv.org/pdf/2405.17475v1
|
How Culturally Aware are Vision-Language Models?
| 2024-05-24T04:45:14 | 2024-05-24T04:45:14 |
Olena Burda-Lassen, Aman Chadha, Shashank Goswami, Vinija Jain
|
cs.CV
|
cs.CV, cs.AI, cs.CL, cs.LG
|
An image is often said to be worth a thousand words, and certain images can
tell rich and insightful stories. Can these stories be told via image
captioning? Images from folklore genres, such as mythology, folk dance,
cultural signs, and symbols, are vital to every culture. Our research compares
the performance of four popular vision-language models (GPT-4V, Gemini Pro
Vision, LLaVA, and OpenFlamingo) in identifying culturally specific information
in such images and creating accurate and culturally sensitive image captions.
We also propose a new evaluation metric, Cultural Awareness Score (CAS),
dedicated to measuring the degree of cultural awareness in image captions. We
provide a dataset MOSAIC-1.5k, labeled with ground truth for images containing
cultural background and context, as well as a labeled dataset with assigned
Cultural Awareness Scores that can be used with unseen data. Creating
culturally appropriate image captions is valuable for scientific research and
can be beneficial for many practical applications. We envision that our work
will promote a deeper integration of cultural sensitivity in AI applications
worldwide. By making the dataset and Cultural Awareness Score available to the
public, we aim to facilitate further research in this area, encouraging the
development of more culturally aware AI systems that respect and celebrate
global diversity.
|
How Culturally Aware are Vision-Language Models?: An image is often said to be worth a thousand words, and certain images can
tell rich and insightful stories. Can these stories be told via image
captioning? Images from folklore genres, such as mythology, folk dance,
cultural signs, and symbols, are vital to every culture. Our research compares
the performance of four popular vision-language models (GPT-4V, Gemini Pro
Vision, LLaVA, and OpenFlamingo) in identifying culturally specific information
in such images and creating accurate and culturally sensitive image captions.
We also propose a new evaluation metric, Cultural Awareness Score (CAS),
dedicated to measuring the degree of cultural awareness in image captions. We
provide a dataset MOSAIC-1.5k, labeled with ground truth for images containing
cultural background and context, as well as a labeled dataset with assigned
Cultural Awareness Scores that can be used with unseen data. Creating
culturally appropriate image captions is valuable for scientific research and
can be beneficial for many practical applications. We envision that our work
will promote a deeper integration of cultural sensitivity in AI applications
worldwide. By making the dataset and Cultural Awareness Score available to the
public, we aim to facilitate further research in this area, encouraging the
development of more culturally aware AI systems that respect and celebrate
global diversity.
|
[
0.0301482715,
0.0732814968,
-0.0373976566,
0.0494055487,
-0.0068052639,
0.0449889563,
0.026206987,
-0.0240834784,
0.0304531958,
-0.0420261361,
-0.0227212571,
-0.0447872728,
0.0221480727,
0.0502019748,
-0.009241702,
-0.1012370437,
0.009144621,
-0.0028580378,
0.0254861712,
-0.0044885036,
-0.0193029698,
0.0406507701,
0.0229600761,
0.0449049436,
-0.0456876196,
-0.0295764357,
0.046409741,
0.0104517527,
0.0451881364,
-0.0279437378,
-0.0396922752,
0.0175812747,
0.0163478348,
0.0171037875,
0.0000018009,
-0.0293481015,
0.0527876802,
-0.0207547285,
0.0700603426,
0.0287435278,
0.0633966699,
-0.0300645158,
-0.0255114622,
-0.0334254541,
-0.06169229,
0.0085645327,
0.0915083215,
0.0487029441,
0.042376373,
0.0270515494,
-0.0055067805,
-0.0375334397,
-0.0357126147,
-0.0282409005,
-0.0097766807,
0.007322914,
0.0236908551,
0.0700045079,
-0.0277626906,
-0.0153529076,
-0.0421947874,
0.0793528557,
0.0065955925,
-0.0019588566,
-0.0557616651,
0.0284754839,
-0.0797389224,
-0.0521421842,
0.0032143481,
0.0488103107,
0.0379817635,
0.0300385263,
-0.0032451581,
0.0176092014,
-0.0401906818,
-0.0163119286,
-0.0574092418,
0.0437116772,
0.0147602921,
0.0114389528,
-0.0228288732,
0.0241640117,
0.0212224219,
0.0244903397,
0.0306711067,
0.0188112594,
-0.0060713394,
0.0358078666,
-0.0644155368,
-0.0274885483,
-0.042272564,
-0.0541651323,
0.0250572599,
-0.0108609293,
0.0384977497,
0.0049813692,
-0.035390798,
0.005170275,
0.0006502413,
-0.0234124139,
0.0056555788,
0.0440372676,
-0.0754662305,
0.0688347518,
-0.0110775828,
0.0777834877,
0.0062383483,
0.0285733193,
-0.0280618593,
-0.0447366238,
-0.0117500527,
0.0063851224,
0.021747442,
-0.0426560789,
-0.0456270278,
-0.020197738,
-0.06183948,
-0.0016012066,
-0.0459968038,
0.0088720787,
-0.017624218,
-0.0163081158,
-0.0083662122,
-0.0005421138,
-0.025760008,
0.022291217,
-0.0160957593,
-0.0292894486,
-0.0573933683,
-0.0755333453,
0.0270996001,
0.0230519958,
0.0062232483,
-0.0061175432,
0.0126494681,
0.034491621,
0.0299802143,
-0.0281033758,
-0.0406259075,
0.0173643138,
-0.0177968163,
-0.0403900631,
0.0179712977,
-0.0560806021,
0.0188122354,
-0.0278306976,
0.0430078618,
-0.0429526642,
0.0174304899,
0.033977963,
-0.0209139381,
0.0664867163,
0.0130481264,
0.0245927274,
0.0367819183,
0.0129703293,
-0.0275632348,
0.0779637694,
0.0592574142,
0.0476373024,
0.0943645686,
-0.0204647668,
0.0059024803,
0.0163253434,
-0.0206049178,
-0.0014226311,
-0.0288781412,
0.0065402854,
0.0042130575,
-0.0230322983,
0.0286324136,
0.0451673791,
0.0397232175,
0.0012906452,
0.0628571734,
0.0961525142,
-0.0318699777,
0.049793452,
0.0548121072,
-0.0173928495,
-0.007095275,
0.0022359542,
-0.0584864318,
0.0650965944,
0.0402347036,
-0.0017361398,
-0.0102806939,
0.0990379527,
-0.0262649655,
-0.0260612052,
-0.0052039083,
0.0056816195,
0.0306848325,
0.0155652156,
0.0149807595,
-0.0101348534,
0.0123391068,
-0.0363004133,
0.0064552729,
-0.0500795692,
0.0019145599,
-0.0171383955,
0.0163011253,
0.0505327322,
0.0256960578,
-0.0091184629,
0.0467142388,
0.0014540073,
0.0514704362,
0.0522649698,
0.0275104754,
0.026804259,
0.0183849763,
-0.0491846092,
-0.0183284935,
0.0026385332,
-0.0018363572,
0.0496504158,
0.0079815611,
0.0611116886,
0.0389811434,
0.0227280315,
-0.081974797,
0.0129702454,
0.0049457648,
-0.0052412297,
-0.0155836139,
0.0078339316,
-0.0098034637,
0.051082328,
-0.0448334031,
-0.0328645371,
-0.0074088126,
-0.0337066613,
0.0107785296,
0.0646356121,
-0.0636327788,
-0.0112617752,
0.0478457808,
-0.1327403188,
-0.0035053899,
0.0192162506,
-0.0079355622,
-0.0252532735,
0.0583631322,
-0.0382765271,
-0.0374077708,
-0.0466014519,
-0.0296350401,
-0.0164232627,
0.0359563567,
-0.0051245233,
0.0337992981,
-0.0364894979,
-0.0117753828,
-0.0022794777,
-0.0504369624,
0.0270916652,
-0.0208890028,
-0.0581970848,
0.0445809662,
0.0095986146,
0.0059793619,
0.0200990383,
0.0472444892,
-0.0118914302,
0.0242865235,
-0.1163487658,
-0.0301539954,
-0.0367616303,
-0.0489682555,
-0.0359904245,
-0.0039052968,
-0.0229458641,
-0.0415639319,
-0.0721124336,
0.0001960227,
-0.0030237294,
0.0197487734,
0.0322237834,
-0.0050955717,
-0.0372071937,
-0.0086413836,
-0.0239929873,
0.0158948284,
0.0591239892,
0.0417362973,
-0.047936134,
-0.0242536645,
-0.0348631069,
-0.0622161701,
-0.0055773789,
-0.0143970707,
-0.0468514375,
0.0099862553,
0.0420490168,
-0.0006656468,
0.0512017421,
-0.0010958327,
-0.0204869229,
0.0307844244,
0.0138129788,
0.0108234091,
-0.0117628342,
0.0284353122,
-0.0099296607,
0.054943651,
-0.0908809379,
-0.0034296911,
0.0079427315,
-0.008721116,
0.0102535496,
-0.0514342599,
0.022900667,
-0.0497209318,
-0.0706008673,
-0.028318461,
0.0226645842,
0.0174827389,
0.0694278553,
-0.0376625359,
-0.0582734458,
0.0519517139,
-0.0183997508,
-0.0377524793,
0.0233098287,
-0.0068220454,
-0.0688142627,
0.0642098039,
0.0082132285,
0.0515405387,
-0.0117869955,
0.0001308523,
-0.0121742878,
-0.0129454508,
0.0052817408,
-0.0450700559,
0.0597886853,
-0.0408443399,
-0.0477508754,
-0.0271852463,
-0.0041515944,
-0.0341736823,
0.0267775878,
-0.0148473578,
0.0074732914,
-0.017215291,
-0.0022754427,
-0.0069798543,
0.0282440502,
0.0480141453,
0.0182797909,
0.0139116021,
0.0322557837,
0.0083761038,
-0.0079262555,
0.0411556549,
0.0019282512,
-0.0041019814,
0.0033154432,
0.011987499,
-0.005594837,
0.0535745062,
-0.0259451028,
-0.0068453373,
0.0423517376,
0.0264966097,
0.0298847761,
-0.0011120675,
-0.0188368596,
0.0670667961,
-0.0179219078,
0.0446355827,
0.0677118972,
0.0066667763,
-0.0481305122,
-0.0116747785,
-0.0137882,
-0.0129910121,
0.0407752767,
-0.0017637121,
-0.0897236317,
0.0143787675,
-0.0182789881,
-0.0153545542,
-0.008384943,
-0.0111690704,
-0.036678493,
-0.0417863317,
0.0362493023,
0.0305251982,
-0.039121069,
-0.0937394649,
0.0322305448,
-0.0202569272,
-0.0263805129,
-0.0193372127,
-0.0239409115,
-0.0017901562,
0.0071108104,
-0.009092588,
-0.0068666968,
-0.0422009677,
0.009735967,
-0.046613019,
0.0107746441,
0.0726215616,
0.0380904377,
0.0072961873,
-0.0628305823,
-0.0680095553,
-0.0164222289,
0.0024212589,
-0.0044195671,
0.0390413776,
-0.0159357786,
-0.0609638207,
0.0265188459,
0.0078920862,
-0.023734482,
-0.0178907961,
0.0261267815,
0.0035872492,
0.0182562005,
-0.0034220929,
-0.0398398824,
0.0152601385,
0.003076168,
0.0655605271,
0.0523286089,
-0.0299083441,
0.0004011456,
-0.0543309003,
-0.0139982179,
-0.0308457818,
-0.0528792031,
0.0222579334,
-0.0251763556,
-0.0500193685,
-0.0753596425,
-0.0320567787,
0.0196184944,
-0.004519715,
0.0504346564,
-0.0411502905,
-0.0198288895,
0.048180975,
-0.0679944605,
0.0514957793,
-0.0446993746,
0.0103381695,
0.0066855764,
-0.0129283434,
0.0149100237,
-0.01740136,
0.0021974149,
0.0379059426,
-0.0547534749,
-0.0034746586,
0.0189010985,
0.0473083295,
0.0110602994,
-0.0244871527,
0.0466258489,
-0.0293538868,
0.0079688132,
0.0153631885,
0.0544706173,
-0.0067868922,
-0.0026809429,
0.0250312127,
-0.0100968257,
0.0304708313,
-0.0209246948,
0.0341070257,
0.0206218269,
0.0292616189,
-0.0171731934,
0.0001035405,
-0.0328906067,
0.0114361811,
-0.0099089835,
-0.0680151209,
-0.0238069724,
-0.0132208075,
-0.0259841047,
-0.0206115432,
-0.0041770791,
-0.0432806648,
-0.0120614218,
0.0100206193,
0.0386027955,
-0.0017917972,
0.0079382332,
-0.0033848477,
-0.0257766452,
-0.0537665859,
0.0324185863,
0.04119635,
0.0114047723,
-0.045789469,
0.0245372746,
-0.0316377133,
0.0410718434,
-0.0204169881,
0.023784522,
0.0013977049,
0.0527257025,
-0.0037065917,
-0.0253121536,
0.0296987426,
0.0421539135,
-0.0058880448,
-0.0196001306,
0.0129221324,
0.0615188368,
0.1163483188,
-0.0298537351,
0.0530099571,
0.0024438514,
-0.0012319545,
0.0104464851,
0.03212329,
0.0436957255,
-0.045136597,
0.0060841264,
0.0075914389,
-0.0050707785,
-0.0151899941,
0.034419965,
-0.0471159443,
0.0025845855,
0.0112602143,
-0.0553048663,
-0.0174000058,
-0.0083691664,
0.0162488986,
-0.0336251408,
-0.0236211307,
0.0620339774,
-0.0605856553,
-0.0683645681,
0.0135388682,
-0.006222418,
0.0273765177,
0.016135985,
0.025783103,
0.0396886654,
0.019184161,
-0.0213953797,
-0.0281827413,
-0.041292727,
-0.0703358799,
-0.0041748462,
-0.014479802,
-5.454252278e-33,
-0.0478682294,
-0.0101000089,
-0.0076224091,
0.0055032554,
-0.0219562929,
-0.0092720855,
-0.0362544619,
0.0118245678,
-0.0325102843,
0.0241585486,
0.0113054384,
-0.0050878003,
0.0229640305,
0.0309851635,
0.0075650914,
0.0225159433,
0.0124254208,
0.0251809191,
-0.0262459833,
-0.0110921143,
0.013134188,
-0.0052221087,
0.0461925231,
-0.0970873684,
0.0282145794,
0.0111353435,
-0.0070546931,
-0.0327501372,
0.0695270225,
0.0692510009,
-0.0575306267,
0.0342729762,
0.0319889672,
0.020443568,
-0.0081810094,
0.0139623024,
-0.0436417535,
-0.0107596526,
0.0365890674,
0.0647197664,
0.0098579973,
-0.0345413424,
-0.0088993507,
-0.054654181,
-0.0409711599,
-0.0018582942,
0.0702065751,
-0.0606938042,
0.0196912531,
-0.006189262,
-0.0187863372,
0.0176041778,
-0.01216623,
0.0007770697,
0.042178642,
0.0629263371,
0.0268667359,
0.0501129106,
-0.0302901566,
0.0465858467,
0.0162804537,
-0.0023166684,
0.1103997082,
0.0346468091,
0.0256321728,
0.0665610284,
0.0226713941,
-0.0143187782,
-0.0060822717,
-0.025919158,
0.0248188972,
0.006531565,
0.0357965529,
0.0993118137,
-0.0010069237,
0.0216967352,
-0.0247840602,
0.015105566,
-0.0290029999,
0.0047801971,
0.0416534133,
0.0034175969,
-0.0379987918,
0.0397540554,
-0.0165503919,
-0.0395731591,
-0.0296971761,
-0.029592175,
0.0303731225,
-0.0125074238,
-0.0272157155,
0.0690556988,
0.038573198,
-0.0665333197,
0.0163163673,
-0.0082766265,
-0.002655481,
-0.0033407148,
0.0297758207,
-0.0074653882,
-0.0138450768,
-0.0138681205,
0.0324489586,
0.0314811207,
0.0379978381,
-0.0138190957,
0.045325309,
0.0142902136,
-0.0487725101,
-0.0174444709,
-0.002625077,
-0.001254229,
-0.0350290574,
0.006248923,
0.0221368149,
-0.0082431091,
-0.0652458891,
-0.0234111957,
0.0246420782,
-0.0085864794,
-0.013102469,
-0.0763344094,
-0.0223359931,
0.0461525805,
-0.0667028129,
0.0095293261,
-0.0179916006,
-0.022304425,
-0.0104813091,
-0.013896185,
-0.031706728,
0.0033743514,
2.447e-7,
0.000006993,
0.0069476049,
0.0143156415,
0.0525906831,
0.0190014299,
0.0170933455,
-0.045517318,
0.0097472444,
-0.0018460056,
-0.0286658239,
0.0391072333,
0.0066436394,
0.0648297518,
-0.007261306,
-0.057822682,
0.0007356603,
-0.0197249502,
0.0042765695,
-0.0461559743,
0.0017783776,
0.0165828075,
0.0701843351,
0.0026390627,
-0.0192361344,
-0.015145638,
-0.0691853613,
0.0496462807,
-0.0741999075,
-0.0018536556,
-0.0315292068,
-0.0168297831,
0.024481196,
-0.0148492027,
-0.0108608101,
0.000115952,
0.0139818778,
0.0976879448,
0.0410850719,
-0.0182879902,
0.0379771367,
-0.0275658127,
0.0646784455,
-0.0164367706,
-0.0826806426,
0.0478255749,
0.019778138,
-0.0493941978,
-0.0256984588,
0.0090713641,
0.0047948617,
0.0704962462,
-0.0242132638,
-0.0404350311,
0.0130534889,
0.0463650376,
-0.0367376618,
-0.007558574,
-0.042871926,
0.0086032012,
-0.043150302,
0.0099041658,
-0.0207415819,
-0.0484952182,
-0.0375946909,
-0.0336919539,
0.0348523743,
-0.0193652734,
1.849529699e-34,
0.0371387191,
-0.0267552026,
-0.0210968368,
-0.0055222213,
0.0419911221,
-0.0108093536,
-0.0077484748,
-0.0225123856,
-0.0035562294,
-0.0167637281,
0.0028006276
] |
http://arxiv.org/abs/2405.15208v1
|
http://arxiv.org/pdf/2405.15208v1
|
Decoding at the Speed of Thought: Harnessing Parallel Decoding of Lexical Units for LLMs
| 2024-05-24T04:35:13 | 2024-05-24T04:35:13 |
Chenxi Sun, Hongzhi Zhang, Zijia Lin, Jingyuan Zhang, Fuzheng Zhang, Zhongyuan Wang, Bin Chen, Chengru Song, Di Zhang, Kun Gai, Deyi Xiong
|
cs.CL
|
cs.CL, cs.AI
|
Large language models have demonstrated exceptional capability in natural
language understanding and generation. However, their generation speed is
limited by the inherently sequential nature of their decoding process, posing
challenges for real-time applications. This paper introduces Lexical Unit
Decoding (LUD), a novel decoding methodology implemented in a data-driven
manner, accelerating the decoding process without sacrificing output quality.
The core of our approach is the observation that a pre-trained language model
can confidently predict multiple contiguous tokens, forming the basis for a
\textit{lexical unit}, in which these contiguous tokens could be decoded in
parallel. Extensive experiments validate that our method substantially reduces
decoding time while maintaining generation quality, i.e., 33\% speed up on
natural language generation with no quality loss, and 30\% speed up on code
generation with a negligible quality loss of 3\%. Distinctively, LUD requires
no auxiliary models and does not require changes to existing architectures. It
can also be integrated with other decoding acceleration methods, thus achieving
an even more pronounced inference efficiency boost. We posit that the
foundational principles of LUD could define a new decoding paradigm for future
language models, enhancing their applicability for a broader spectrum of
applications. All codes are be publicly available at
https://github.com/tjunlp-lab/Lexical-Unit-Decoding-LUD-. Keywords: Parallel
Decoding, Lexical Unit Decoding, Large Language Model
|
Decoding at the Speed of Thought: Harnessing Parallel Decoding of Lexical Units for LLMs: Large language models have demonstrated exceptional capability in natural
language understanding and generation. However, their generation speed is
limited by the inherently sequential nature of their decoding process, posing
challenges for real-time applications. This paper introduces Lexical Unit
Decoding (LUD), a novel decoding methodology implemented in a data-driven
manner, accelerating the decoding process without sacrificing output quality.
The core of our approach is the observation that a pre-trained language model
can confidently predict multiple contiguous tokens, forming the basis for a
\textit{lexical unit}, in which these contiguous tokens could be decoded in
parallel. Extensive experiments validate that our method substantially reduces
decoding time while maintaining generation quality, i.e., 33\% speed up on
natural language generation with no quality loss, and 30\% speed up on code
generation with a negligible quality loss of 3\%. Distinctively, LUD requires
no auxiliary models and does not require changes to existing architectures. It
can also be integrated with other decoding acceleration methods, thus achieving
an even more pronounced inference efficiency boost. We posit that the
foundational principles of LUD could define a new decoding paradigm for future
language models, enhancing their applicability for a broader spectrum of
applications. All codes are be publicly available at
https://github.com/tjunlp-lab/Lexical-Unit-Decoding-LUD-. Keywords: Parallel
Decoding, Lexical Unit Decoding, Large Language Model
|
[
0.0125392526,
0.0931408703,
0.0112623591,
0.0694559887,
-0.0186876114,
0.026394587,
0.0176706705,
-0.0000889544,
-0.0449387208,
-0.0437109955,
0.0109197581,
-0.03946447,
-0.0097540924,
0.0269102547,
0.0203985404,
-0.009637326,
0.0468306094,
-0.0240253173,
-0.0035968642,
0.0026464455,
0.0042851372,
0.0339518115,
-0.0130529236,
0.0637706146,
-0.0272686929,
-0.028362399,
-0.0266481005,
0.0249539018,
-0.0129731661,
-0.0405308269,
0.0155067397,
0.0128896153,
0.0256384984,
0.0194636621,
0.0000020884,
-0.0241951309,
-0.027104767,
-0.0087306276,
-0.0130634587,
0.0273005515,
0.0355733112,
-0.0547483005,
-0.0085495161,
0.0253000837,
-0.0403265357,
-0.0319937132,
0.0507663004,
0.008522382,
0.0247386359,
0.0866587311,
-0.0092576593,
-0.0417109691,
0.0051869978,
-0.0114623914,
-0.0580634363,
-0.029308835,
-0.0115634166,
0.0205113944,
-0.0122598726,
0.0090350397,
-0.0215038285,
0.0298672505,
0.0301771183,
0.0274863224,
0.0081487652,
0.0521313585,
-0.0032567328,
-0.031953644,
-0.0620031841,
0.0534591377,
-0.0318231285,
-0.0213137623,
0.011403827,
-0.0011719281,
-0.0217839219,
-0.0494439192,
-0.018360272,
0.0723735318,
0.0086311046,
0.0705561489,
0.0730432644,
0.0015191787,
0.0407344475,
0.0087762848,
-0.0446582735,
0.0784341171,
0.0267699379,
-0.0077756909,
-0.0033742723,
0.0069456659,
-0.0072836876,
-0.0513065346,
0.0166081451,
0.03899974,
-0.0249180887,
0.0335355178,
-0.0629457906,
-0.0324881561,
0.0099628987,
-0.0487173535,
0.0316739008,
0.0613786168,
-0.0040879808,
0.0152447913,
-0.0281383172,
0.0361794494,
-0.0147736138,
-0.0102124009,
-0.0639952123,
-0.0034007018,
-0.0074860998,
-0.0100928564,
-0.0557793602,
0.0565175489,
-0.0658911318,
-0.0493399464,
-0.0066331485,
0.0453030206,
0.0565213114,
-0.0002483644,
0.0302110538,
-0.001764742,
-0.0381964669,
-0.0219542366,
0.0233996976,
0.0346652903,
-0.0328072906,
-0.0035239663,
0.0027467974,
-0.0853193775,
-0.0366223864,
0.0275896937,
0.0006409212,
-0.0195115469,
0.0053542173,
0.0656152889,
-0.0198856704,
-0.0381471701,
-0.0430514663,
-0.0570365414,
-0.0139193973,
-0.031287618,
-0.0035582085,
-0.0740072504,
0.0331346206,
0.0464083403,
0.0119577562,
-0.0163146034,
0.0336142294,
0.0045156633,
-0.0163140483,
0.0599673167,
-0.0942352712,
0.0149266729,
-0.007760074,
0.0090790913,
0.0354798213,
0.09762007,
0.031835597,
0.0533234663,
0.0236745216,
0.0062286053,
-0.0570836104,
0.0622683875,
0.0188481435,
-0.019346511,
-0.0513011292,
-0.010842572,
-0.0051114755,
0.0149244759,
0.0558858104,
0.0515309162,
-0.0480441451,
0.0353412069,
0.0423924699,
0.0364876874,
0.0155004542,
0.0449438617,
0.0693205893,
0.005551693,
-0.016386563,
-0.0086904466,
-0.013973061,
0.0045590545,
0.0237389002,
0.0224519838,
0.0528359935,
0.0468051024,
-0.017577095,
-0.047510054,
-0.0226690676,
-0.046044942,
0.0047478313,
-0.0090036532,
0.0321031176,
0.0332035944,
-0.067104578,
0.065835923,
0.037836805,
-0.0531104095,
-0.0022834491,
-0.0327759907,
0.0085823443,
0.0841218531,
0.0548575111,
0.0184270646,
-0.0499287099,
0.0254919771,
-0.0195261575,
0.0478511676,
0.020819515,
-0.0107376259,
0.038656842,
-0.0781810656,
0.0065098293,
0.0115853054,
0.0027757594,
0.0375715122,
0.021845011,
0.0052359942,
0.0272413082,
-0.0157355964,
-0.0125633907,
0.0369923264,
-0.0185073987,
-0.0427078195,
0.0432678089,
-0.014872265,
0.0316660926,
0.012206899,
-0.0140235787,
0.0103105828,
0.0287538357,
-0.0598526485,
-0.0154049955,
0.0210678037,
-0.0700872093,
-0.0971531123,
0.0292739533,
-0.0033393621,
0.0015237398,
0.0413401201,
-0.0137059605,
-0.0320802815,
-0.0057211304,
-0.0408373103,
0.0700164363,
-0.0526148379,
0.0190578233,
0.037404459,
0.0121846087,
-0.0193805248,
0.0447619855,
-0.0370254591,
-0.0161305461,
0.0226445124,
0.0379255787,
0.0145369656,
-0.0054690028,
-0.0072894031,
0.0193652064,
-0.0184636153,
-0.0065412843,
-0.0024140649,
0.0241843574,
-0.0159370266,
-0.0264687184,
-0.1049341857,
0.0163773857,
-0.045492921,
0.007723453,
0.0160414614,
0.0085250549,
0.0133933071,
-0.0240757,
-0.0271482542,
0.0107558528,
0.0371685326,
-0.0117575433,
0.026739398,
-0.0020007964,
-0.0263968483,
-0.0189472828,
-0.0022269902,
0.0027803669,
0.0398332961,
-0.0175645147,
-0.0094273556,
-0.0631746873,
0.0444950387,
0.0007178606,
0.0113745546,
0.0214483868,
-0.0403674133,
0.0166342873,
-0.0404059812,
0.0640647635,
0.0129120611,
0.0060131047,
0.0579936765,
0.0198807865,
-0.0290506948,
-0.0347464047,
0.0218866188,
-0.0245811287,
0.0131399473,
-0.0554826632,
-0.0299405009,
0.0470239893,
0.087585628,
0.0444878303,
0.0045599318,
-0.0792799294,
0.0017131091,
-0.009243736,
-0.0524362028,
0.003987574,
0.0104219941,
0.0002568632,
0.0094234068,
-0.0144236069,
-0.0650316626,
0.023247404,
-0.0532045886,
-0.0081774537,
-0.0288399514,
-0.0162479132,
-0.0440556109,
0.0708401054,
0.0089082625,
0.0386007093,
-0.078220576,
-0.0521860011,
0.0144841708,
0.0565921105,
-0.0553213619,
-0.0320479423,
0.0326908827,
-0.0774521455,
-0.002990691,
-0.0120387003,
-0.0421114899,
-0.0575220026,
-0.0136277843,
-0.0300554819,
0.0164584182,
0.0381962173,
-0.0187856015,
0.0188369472,
-0.0083115902,
-0.045991078,
0.0235756151,
-0.0235490687,
0.0095616588,
-0.0377233289,
0.0464063622,
0.0524395183,
-0.0087921452,
0.0191850364,
-0.0386361517,
0.0278986786,
0.0692610666,
0.0261085518,
0.022353787,
-0.0846804753,
0.0225309283,
0.047644861,
-0.008425789,
-0.0044685546,
0.0347991958,
0.0412880853,
-0.0622973032,
-0.0457202755,
0.1461108625,
0.0428062677,
-0.0154384943,
0.0167294517,
0.0162212234,
0.0283936411,
0.0013386649,
-0.0107132103,
-0.0611776412,
-0.0602673478,
0.0118933655,
0.0020458053,
0.0120867174,
-0.0076475744,
-0.0463290848,
-0.093693696,
0.0356041528,
0.0409352146,
-0.0120822899,
-0.047997646,
0.0015088983,
-0.0783857405,
0.017757114,
0.0231047906,
0.0653019547,
0.0148206977,
0.0227104984,
0.0289649013,
-0.0308169313,
-0.0184386615,
-0.0247361232,
-0.024422111,
0.0739869028,
0.0866979808,
0.0140281953,
0.022166688,
-0.012947496,
0.0131143862,
-0.0181561746,
0.0062975185,
-0.0385558121,
0.0126734469,
-0.040104311,
-0.0185029637,
0.0123740239,
0.0299500953,
-0.0584283136,
-0.0028210396,
0.011337026,
0.0453149416,
-0.0272539165,
-0.0034240226,
-0.0316025056,
-0.0109902546,
0.0132375266,
0.0744850114,
0.0059997053,
-0.0515223034,
-0.0481063128,
-0.0208732784,
-0.0144212572,
-0.0861648619,
-0.0045139412,
0.0071425368,
0.0017155986,
-0.0107417302,
-0.0655300543,
-0.0437038839,
0.0194080081,
0.0297224782,
-0.0043550306,
0.0066420417,
0.0012399971,
-0.0178480223,
-0.0613510422,
-0.0186162274,
-0.0056304173,
-0.0126071889,
-0.0665163025,
-0.0003482102,
0.0119508971,
-0.0403819755,
0.005984711,
0.0633290038,
-0.1108495519,
-0.0563395321,
0.004829511,
0.0273662023,
0.0094698127,
0.0409521908,
-0.0012466399,
-0.0578510948,
0.0234893896,
0.015983507,
-0.0436289124,
0.0195888411,
0.0030630326,
0.0338661112,
0.0197477397,
0.0523333922,
-0.0213065427,
0.0280317347,
-0.002356695,
0.067821689,
-0.0385567583,
-0.0201682374,
0.032780271,
0.0196401849,
0.0265321322,
-0.025286952,
0.0284321345,
0.0040828912,
0.0316179655,
-0.0388205759,
-0.0009451309,
-0.0781434625,
0.0284319613,
0.0371321738,
-0.011818612,
-0.0313540511,
-0.0374946743,
0.0334942043,
0.0031243309,
0.0132089304,
-0.0035460985,
-0.0352969095,
0.0087571107,
0.0491765924,
-0.0282523949,
-0.0300936792,
0.0092526227,
-0.0211120527,
-0.0021732764,
-0.0426738709,
0.0321225487,
0.029098846,
0.002382142,
-0.0288915671,
0.0136965495,
-0.0226279367,
-0.0306923501,
-0.0268834606,
0.0117703835,
0.0482870825,
0.0098004024,
0.0584092736,
-0.0534461029,
0.0196378529,
-0.025637202,
0.0525330976,
0.0008292071,
-0.0355359912,
0.0330818035,
0.0003951356,
-0.0496022403,
-0.0274271779,
0.0671105906,
0.0207836982,
0.0139322793,
0.0234790631,
0.0038348737,
-0.0484057851,
-0.0198614169,
0.0131411599,
0.0170404315,
-0.0127331717,
0.0331024528,
-0.0255673006,
-0.0452044494,
0.0096898843,
0.0319964588,
0.0295497477,
-0.0207672678,
0.029360909,
0.0572625324,
-0.0294037499,
-0.0404597558,
-0.0453152508,
0.0288001802,
-0.0193145014,
-0.0340250507,
-0.0322890691,
-6.526822915e-33,
-0.0378209166,
-0.0380929187,
0.0381443352,
-0.0322120078,
-0.0372998863,
0.0131368525,
-0.021158414,
-0.0213010367,
0.0306845848,
0.0084261904,
-0.0000644144,
0.0183035154,
0.0045948755,
0.0524799526,
0.0068986677,
0.0191276632,
0.0599613376,
-0.0234211963,
0.0131447241,
0.0111806374,
0.0324399956,
0.0614349507,
0.0216472708,
-0.0046441671,
-0.0252123885,
-0.0224756431,
-0.0090101063,
0.0078776199,
0.0008231684,
0.0192490742,
-0.0227614641,
0.0205043331,
0.0268396977,
-0.030485291,
-0.0446527675,
0.0265272837,
-0.0916422531,
0.0010325519,
0.0246116463,
0.0329264328,
-0.0557356887,
-0.0358526446,
0.020988198,
0.0041152416,
-0.0346572809,
-0.0096840542,
-0.0156953782,
-0.0271300301,
-0.0237695277,
-0.0590652153,
-0.0491215289,
0.0392011255,
-0.0252338182,
0.0088719577,
0.0555361621,
0.0150170736,
-0.0235328507,
0.0000827844,
-0.0372859053,
0.0011213982,
0.0215920117,
0.0575830154,
0.0525250025,
0.006479424,
-0.0138897588,
0.039636407,
0.0464206487,
-0.0436952859,
-0.0284248982,
0.0413416363,
0.0103536779,
0.0548814796,
0.0050284844,
0.0147923762,
0.0476893783,
-0.0330214798,
-0.0574964397,
0.0031719988,
0.0088437125,
-0.00428158,
0.0404292643,
0.0037221338,
-0.0354604349,
-0.0124474009,
-0.0153504992,
-0.0015674636,
-0.0213735886,
-0.0336943194,
0.0686976686,
0.0298592839,
-0.0128159551,
-0.0131276986,
0.0563981459,
-0.0341071673,
0.0253193546,
0.0097915223,
-0.0142501323,
-0.0060253744,
-0.0338579118,
0.0323873647,
-0.0428671725,
-0.0347849354,
0.0223660432,
-0.0227098707,
0.0044517787,
0.0142803211,
-0.0085090706,
0.0229773056,
-0.0627856776,
-0.0067817033,
-0.0442097895,
0.0103820432,
0.0020940744,
0.0187197309,
0.0632846579,
-0.0212857239,
-0.011191939,
0.0444242656,
0.0024757623,
-0.0192260984,
-0.0453880392,
-0.0365312733,
0.0129760904,
0.0236346219,
-0.0454192348,
0.007245936,
-0.0367878154,
0.0084143123,
0.0212325156,
-0.0493370146,
-0.0217481907,
0.0368016548,
2.643e-7,
0.0336492807,
0.0541888475,
0.0545236133,
0.0315539278,
0.0313069411,
0.000435688,
0.0078301672,
0.0488091595,
-0.0149959363,
0.0207449775,
0.0200416278,
0.0266541205,
0.0473197661,
0.0008174414,
-0.0866746828,
0.0393546931,
-0.0244810805,
0.0198141765,
-0.0622611642,
-0.00780618,
0.051960025,
0.0715376213,
-0.0177272279,
-0.0174822267,
0.0410422012,
-0.0114523759,
0.0467494316,
-0.0057466226,
-0.0281005595,
0.0181172956,
0.0049747233,
0.017363131,
-0.0327443704,
-0.0162521303,
-0.0042713499,
0.0219193287,
-0.0153890243,
0.0790788233,
0.0020900564,
-0.0362004228,
0.0351391211,
-0.0288366619,
-0.0106774336,
-0.0271880701,
0.0758881569,
-0.0017312487,
-0.0014974787,
-0.0379140936,
0.0072858441,
0.0320902914,
-0.0561033152,
-0.0092932414,
-0.0083482107,
0.0276852883,
0.0347468555,
-0.0162606686,
-0.0094683329,
-0.0750344843,
-0.0335887745,
-0.0311933775,
-0.0584828332,
-0.0329986401,
-0.057893496,
0.0354429595,
0.0267564971,
-0.0540091582,
-0.025562847,
2.822802517e-34,
0.0185680334,
-0.0291702859,
0.0283510275,
0.0879235044,
0.0353307351,
-0.03306669,
-0.0253371522,
0.0268971249,
0.0159017015,
-0.0649980158,
-0.0371048227
] |
http://arxiv.org/abs/2405.15202v1
|
http://arxiv.org/pdf/2405.15202v1
|
Cross-Task Defense: Instruction-Tuning LLMs for Content Safety
| 2024-05-24T04:14:32 | 2024-05-24T04:14:32 |
Yu Fu, Wen Xiao, Jia Chen, Jiachen Li, Evangelos Papalexakis, Aichi Chien, Yue Dong
|
cs.CL
|
cs.CL, cs.CR
|
Recent studies reveal that Large Language Models (LLMs) face challenges in
balancing safety with utility, particularly when processing long texts for NLP
tasks like summarization and translation. Despite defenses against malicious
short questions, the ability of LLMs to safely handle dangerous long content,
such as manuals teaching illicit activities, remains unclear. Our work aims to
develop robust defenses for LLMs in processing malicious documents alongside
benign NLP task queries. We introduce a defense dataset comprised of
safety-related examples and propose single-task and mixed-task losses for
instruction tuning. Our empirical results demonstrate that LLMs can
significantly enhance their capacity to safely manage dangerous content with
appropriate instruction tuning. Additionally, strengthening the defenses of
tasks most susceptible to misuse is effective in protecting LLMs against
processing harmful information. We also observe that trade-offs between utility
and safety exist in defense strategies, where Llama2, utilizing our proposed
approach, displays a significantly better balance compared to Llama1.
|
Cross-Task Defense: Instruction-Tuning LLMs for Content Safety: Recent studies reveal that Large Language Models (LLMs) face challenges in
balancing safety with utility, particularly when processing long texts for NLP
tasks like summarization and translation. Despite defenses against malicious
short questions, the ability of LLMs to safely handle dangerous long content,
such as manuals teaching illicit activities, remains unclear. Our work aims to
develop robust defenses for LLMs in processing malicious documents alongside
benign NLP task queries. We introduce a defense dataset comprised of
safety-related examples and propose single-task and mixed-task losses for
instruction tuning. Our empirical results demonstrate that LLMs can
significantly enhance their capacity to safely manage dangerous content with
appropriate instruction tuning. Additionally, strengthening the defenses of
tasks most susceptible to misuse is effective in protecting LLMs against
processing harmful information. We also observe that trade-offs between utility
and safety exist in defense strategies, where Llama2, utilizing our proposed
approach, displays a significantly better balance compared to Llama1.
|
[
0.0340741873,
0.0255111251,
-0.0104047917,
0.0471633822,
-0.0160009675,
-0.000935634,
-0.008325547,
-0.0078906249,
0.0203757975,
-0.0475834571,
-0.036117088,
-0.0512889586,
-0.0064343167,
-0.0084874323,
0.0499353521,
-0.0151968207,
0.039790649,
-0.0532600097,
-0.0014336668,
-0.0014137395,
0.027462136,
0.0016475136,
0.0183692127,
0.0233297423,
0.0293702818,
-0.0239045564,
-0.0041393619,
0.0200336557,
0.0260354578,
-0.0276351403,
0.0533776768,
0.0141130695,
0.0058066235,
0.0441265181,
0.0000020124,
-0.0394745469,
-0.044946488,
0.0116946073,
-0.0138656655,
0.0161547717,
0.0536196269,
-0.0362223685,
-0.0107197976,
0.0546494871,
-0.0222551189,
0.0372696891,
0.0318505466,
0.0668471307,
0.0902495533,
0.0998062193,
0.0088471873,
0.0055444627,
0.0461614802,
-0.0256042164,
0.0275854915,
-0.066130057,
0.0266487151,
-0.0181797445,
0.0044015832,
0.0392142497,
-0.0071371165,
0.0126067679,
0.0017087091,
-0.0284896065,
-0.0376750194,
-0.0067143543,
-0.0163247734,
-0.0044692154,
-0.0155010736,
0.0552829131,
-0.037390288,
-0.0022268591,
0.0184100121,
-0.0043008998,
0.0309704654,
-0.0025799638,
-0.0360397287,
0.0644634143,
-0.0088309674,
-0.0109069692,
0.0283523407,
-0.0019046665,
0.0354818031,
0.0205844715,
0.0226171464,
0.0343390033,
-0.0009167296,
-0.0414286964,
0.012909946,
-0.0129891438,
-0.0293795858,
-0.0400228053,
0.0157198459,
0.0359241925,
-0.0056182374,
0.0318204239,
-0.0742171407,
-0.0800340101,
0.0494328514,
-0.0808528736,
0.0105399815,
0.0188109372,
0.005518178,
0.041650217,
-0.0227625724,
0.0219085217,
0.0025679588,
0.0571139902,
-0.0004996218,
0.0181770977,
-0.0154404677,
-0.0048601497,
-0.0280659255,
0.0499692596,
-0.0126085347,
0.0365474559,
-0.0298357662,
0.0600109845,
-0.0389460921,
0.0200055987,
0.0221868828,
-0.0019229351,
-0.0277265608,
-0.0016989968,
0.0003660386,
-0.0010511717,
-0.0085764695,
0.0002577969,
-0.0053222603,
-0.040040262,
-0.0304958299,
0.0183910467,
-0.0270461235,
-0.0391350165,
0.0155937271,
0.1105687693,
0.0206753556,
-0.0285116583,
-0.0588285252,
-0.0877073407,
-0.0355202965,
-0.0322218947,
0.0113856727,
-0.0538151935,
-0.0120718069,
-0.0133810146,
0.0038395817,
-0.0580832362,
0.009281924,
0.0398618802,
-0.0716524422,
0.042336423,
0.0299700145,
-0.0195498262,
0.019392468,
0.0445662551,
-0.0278160609,
0.0753045976,
0.0226035547,
0.0340887569,
0.076595515,
0.0108546624,
-0.0000669699,
0.0391541794,
0.0329046398,
0.0517645441,
0.0097511187,
0.0060796533,
0.0472768471,
-0.0476611592,
0.0158235785,
0.0660128221,
-0.0020262892,
-0.0420459881,
0.0757278875,
0.0246801656,
-0.0484921485,
-0.0124481246,
0.0177603103,
0.0382915735,
-0.0142800491,
0.039958965,
-0.0654802397,
0.0156298745,
-0.0708986521,
-0.0361932293,
0.037000604,
0.03337707,
-0.015786916,
-0.0087803481,
-0.0106323268,
-0.0221079644,
0.0455099083,
-0.0017168149,
-0.0022740127,
0.0172360316,
-0.044567477,
0.0293494724,
0.0360462815,
-0.0236813575,
0.0083680591,
-0.0656518713,
-0.0177265219,
0.0626648217,
0.0527096689,
-0.0327681601,
-0.0421166904,
-0.0456371754,
0.0326126479,
0.0115459953,
0.0269705541,
0.0015517453,
0.0283921733,
-0.0376973599,
-0.0490639135,
-0.0345750004,
0.0206512865,
-0.0264587495,
0.0462229811,
0.002826208,
0.0315000638,
-0.0106735835,
-0.1014706567,
0.0312577896,
-0.0158520043,
-0.0619765185,
-0.0085241506,
0.0291014723,
0.018970564,
-0.0100970399,
-0.0067701852,
0.0207465291,
0.0355401337,
0.0041553755,
0.0003258083,
-0.0435695536,
-0.0192422848,
0.0018645951,
0.0339955874,
-0.0213508215,
-0.005108336,
0.0389290862,
-0.0166174211,
-0.008546317,
0.0045942599,
-0.0693968758,
-0.0137428418,
0.0145115945,
0.020557994,
-0.0078919437,
0.0029168425,
-0.0392680429,
0.0494901836,
-0.0016820634,
-0.0065960549,
-0.0020246156,
-0.0506451651,
0.0010893918,
-0.0037983272,
-0.0456777066,
-0.0203811862,
-0.0262422133,
0.0457079336,
0.027501395,
0.0241487399,
-0.0289208703,
-0.0201714337,
-0.0469373949,
0.0230220947,
-0.0249387715,
0.0506384633,
0.048473604,
-0.0029267098,
-0.0068570836,
-0.0054974533,
-0.0406790711,
-0.0100248158,
0.0028815398,
0.0451380387,
0.0124143902,
-0.0023971314,
0.0112362364,
-0.0131306788,
-0.0143736433,
-0.0172333419,
0.0305431988,
-0.0155438902,
-0.0079632141,
-0.0296268072,
0.0619627312,
0.0244740453,
0.0169586018,
-0.0042520505,
-0.0293383729,
-0.0161814261,
-0.0089487117,
0.0215909146,
0.0774756074,
0.0106703835,
-0.0013023152,
0.0717396438,
-0.0001967388,
-0.0512991473,
0.0084137581,
-0.0054740119,
-0.0366852246,
-0.0350377709,
-0.0343165994,
0.0132426545,
0.0654342026,
0.0040342528,
-0.0385157466,
-0.0364719257,
0.0021713094,
-0.0448845401,
-0.0179617386,
-0.0205577537,
-0.0412005,
-0.0139026763,
0.0162698533,
-0.0252672285,
-0.0466493592,
0.0840599239,
-0.0118729006,
0.0497083403,
-0.0650775507,
-0.0135627873,
-0.0221117716,
0.0396797471,
-0.0139956633,
0.0335799158,
0.0269355755,
-0.0197646767,
-0.0710681528,
0.0018028044,
0.0634148717,
-0.0293564741,
0.046113085,
-0.0380819701,
-0.0293136686,
-0.0425470583,
-0.0267050788,
-0.0356713906,
-0.0561494231,
0.0091296304,
-0.0036764201,
0.0437825285,
0.0011806467,
-0.0451115258,
-0.0216183141,
0.0502098501,
-0.0025026822,
-0.06294889,
-0.0128726447,
-0.0132458778,
0.0286466479,
0.0559316613,
-0.0512093753,
0.0096206581,
-0.0657824874,
0.030593032,
0.0996495411,
0.0498111509,
-0.0311185773,
0.0405111723,
-0.0216834247,
0.0458626486,
-0.0331992097,
-0.0173099302,
0.0516674481,
0.0608203672,
-0.0705954731,
-0.0252612215,
0.064822413,
0.0207504947,
-0.0335323662,
-0.0064380388,
-0.0105896695,
0.0047406289,
0.013359515,
0.0088027362,
-0.04100601,
-0.0172703732,
0.0215467867,
0.0574267358,
-0.0050523253,
-0.0260166805,
-0.0086350655,
-0.0995610133,
-0.0055472068,
0.0074179955,
-0.0162087921,
-0.0138521111,
-0.066527836,
-0.0183022749,
-0.0027623868,
0.0365076996,
-0.0161925349,
-0.0388445593,
-0.0305780489,
0.0290215351,
0.020290358,
-0.0063455035,
0.0138060097,
-0.0526263565,
0.0187411439,
0.0503344052,
0.0013106072,
-0.0191210732,
-0.0148866763,
0.054771997,
-0.0094512608,
-0.041169811,
0.0124519784,
0.0752099156,
-0.0234463662,
0.0109914085,
0.0204872489,
0.0388543643,
-0.0111679882,
0.0008193938,
0.0531428978,
0.073922202,
-0.0023023821,
-0.0029256993,
-0.0468304344,
0.0010148549,
0.0136369979,
0.0866835341,
0.0125992987,
-0.0529172979,
-0.0159288924,
-0.0524973534,
-0.0319965445,
-0.0639408901,
-0.0354820415,
0.0104879411,
-0.0461473241,
-0.0306592416,
-0.0628714412,
-0.0439151153,
0.0055915257,
0.0028648572,
0.0331845544,
-0.018315075,
-0.007608789,
0.0179908723,
-0.0489719473,
-0.012078804,
-0.0483132191,
0.0335251801,
-0.058881104,
0.0273158327,
0.0009340465,
-0.0098418538,
-0.0134133985,
0.0133515447,
-0.0999207869,
-0.0857353508,
-0.0263592191,
0.0681903288,
0.0222518425,
0.0131547134,
-0.0340749025,
-0.0511766933,
0.0583451129,
0.0396916084,
-0.0840241015,
-0.0123449406,
0.0517855026,
-0.0378667228,
0.0060098702,
0.0337697938,
-0.0628116578,
-0.0410766676,
0.0384442583,
-0.0095747132,
-0.0202968288,
-0.0348611772,
-0.0003646167,
0.0451920405,
0.0078344028,
-0.0008601691,
0.0304902997,
-0.009054983,
0.0432211272,
-0.0051572067,
-0.0345075615,
0.0252888612,
0.049298346,
0.0352927856,
-0.0254589189,
0.0070117065,
0.0582630821,
0.0136906644,
0.0172086246,
-0.0423709117,
0.0139129283,
-0.0135383783,
0.0097674225,
0.0340498015,
0.0146268839,
-0.0106605962,
0.0105697168,
0.0154142464,
0.0165784955,
-0.0127058746,
0.0203790963,
0.002117187,
0.023977343,
0.006684829,
0.0612812489,
-0.0584079474,
-0.0393535718,
-0.0279206149,
0.0014343732,
0.0124997655,
-0.022708822,
0.0472221002,
-0.015925603,
-0.0024610253,
-0.047280468,
-0.0204642732,
0.0044061458,
-0.0418260507,
0.004849873,
0.0177061521,
0.0129463403,
-0.013834591,
0.1244573444,
-0.0026715316,
0.0470094383,
0.0401631296,
-0.0554395169,
-0.0362073928,
-0.030823525,
0.0157888457,
0.0378155969,
-0.0367834605,
-0.0049563069,
-0.0608959086,
-0.0824909061,
-0.0019402801,
-0.0133242514,
0.037039306,
0.0143007878,
0.014417165,
0.0397794247,
0.0108936671,
0.0272291303,
-0.0583345182,
0.0131157264,
0.018954277,
-0.0060983063,
-0.0101669971,
-5.782067898e-33,
0.0109277619,
-0.0227394328,
-0.0077161775,
0.017147813,
0.0407427363,
0.0150824022,
-0.0152558088,
-0.0066683483,
-0.012388112,
-0.0219915248,
-0.0211135857,
0.0251211189,
-0.0147447316,
0.0329201706,
0.0040133116,
-0.0285605621,
-0.0028983708,
-0.0229320228,
-0.0348556824,
0.065441519,
0.062705189,
-0.0065997639,
0.0258961581,
-0.0634194314,
0.0512992926,
0.0084811505,
-0.0084251156,
0.0029085001,
-0.0319418237,
0.0719000697,
-0.0060420241,
0.0496715233,
0.0390122086,
-0.0517214015,
0.0041617444,
0.0355642922,
-0.036005713,
-0.0227940343,
-0.0045075007,
-0.0303422622,
-0.0340578929,
-0.0955350474,
0.056432683,
0.0060333484,
0.0152259255,
-0.0365174673,
0.0269001033,
-0.0180992559,
-0.0007704742,
-0.0541525073,
-0.0516113713,
0.0027919824,
-0.0065375813,
-0.0394852683,
0.015306442,
0.014390354,
0.0386085957,
0.0097864727,
-0.0256639719,
0.0340166688,
-0.0025854623,
0.0554949306,
0.0346831046,
-0.0334096774,
-0.0398727208,
0.0515191145,
-0.0564591289,
0.0442372039,
0.015945768,
-0.0232885387,
0.0075169615,
0.0283184089,
0.0380530134,
0.0238921698,
0.0550001636,
-0.0195108466,
-0.0937853679,
0.0635336265,
0.0099323178,
0.0575394481,
0.0273498241,
-0.0111889513,
-0.0340249464,
0.0017959556,
-0.0222230162,
0.0318399295,
-0.0294668861,
-0.0037214076,
0.0078438753,
0.013859014,
-0.0411446877,
-0.0384055078,
0.0067635784,
-0.0526351631,
0.0151208071,
0.0328364484,
-0.0051679173,
-0.0440717079,
0.0044602542,
0.0170762986,
-0.0288225822,
0.0111792609,
0.0124589289,
-0.0757267848,
0.0122137843,
0.0016085507,
0.0702927038,
0.0229075458,
-0.0685476139,
-0.006641259,
-0.048352547,
0.0645425394,
0.0305014495,
0.0658347011,
0.0250199325,
-0.0003792118,
0.0128632439,
0.0143073425,
0.040152479,
-0.0353192091,
-0.0330593511,
0.0134855332,
-0.0103342775,
-0.0041458062,
-0.020513054,
0.0032442904,
0.0007807438,
0.0360249318,
0.0141302124,
-0.0970340222,
-0.0084573487,
-0.0090720383,
2.782e-7,
0.0251555592,
0.0417279303,
0.0208805185,
0.0438413769,
0.0118214916,
0.0585052632,
-0.0218912158,
0.0765933767,
-0.0221061204,
-0.0372217335,
-0.0290183444,
0.0106424857,
0.0500011668,
0.0548496656,
-0.0736848041,
-0.0135116531,
-0.0513941236,
0.0212358274,
-0.0382298231,
-0.0261375606,
0.0702431574,
0.0901375487,
-0.0079167979,
-0.0219240654,
-0.0217308421,
-0.0648405328,
0.026106637,
0.0079540648,
0.0051238486,
0.0183655806,
0.052837085,
0.0433774553,
0.0037492686,
0.0188508406,
-0.0485942215,
-0.0013131959,
-0.0275725536,
0.0771304965,
0.0252618045,
-0.0318815522,
0.0005366305,
-0.0695462376,
-0.0017045022,
-0.00490946,
0.0770572796,
-0.0032612057,
0.0033688801,
0.0325758681,
-0.0040324056,
-0.0180765782,
-0.0035910322,
-0.0216095727,
-0.0148244677,
0.0180467423,
-0.0088880928,
-0.0024682512,
0.0165667739,
-0.0726446435,
-0.0357078649,
-0.0029992226,
0.0052705589,
-0.0675268993,
-0.0162420757,
0.0302207042,
0.0176514834,
-0.069157511,
0.0064021177,
2.932016274e-34,
0.033235319,
-0.0693650469,
-0.0073196422,
-0.011748191,
0.0272349808,
-0.0147936353,
-0.0308340993,
-0.0032461991,
0.0114692254,
-0.0189253967,
-0.0418062955
] |
http://arxiv.org/abs/2405.15198v1
|
http://arxiv.org/pdf/2405.15198v1
|
RAEE: A Training-Free Retrieval-Augmented Early Exiting Framework for Efficient Inference
| 2024-05-24T04:01:24 | 2024-05-24T04:01:24 |
Lianming Huang, Shangyu Wu, Yufei Cui, Ying Xiong, Xue Liu, Tei-Wei Kuo, Nan Guan, Chun Jason Xue
|
cs.CL
|
cs.CL
|
Deploying large language model inference remains challenging due to their
high computational overhead. Early exiting accelerates model inference by
adaptively reducing the number of inference layers. Existing methods require
training internal classifiers to determine whether to exit at each intermediate
layer. However, such classifier-based early exiting frameworks require
significant effort to design and train the classifiers. To address these
limitations, this paper proposes RAEE, a training-free Retrieval-Augmented
Early Exiting framework for efficient inference. First, this paper demonstrates
that the early exiting problem can be modeled as a distribution prediction
problem, where the distribution is approximated using similar data's existing
information. Next, the paper details the process of collecting existing
information to build the retrieval database. Finally, based on the pre-built
retrieval database, RAEE leverages the retrieved similar data's exiting
information to guide the backbone model to exit at the layer, which is
predicted by the approximated distribution. Experimental results demonstrate
that the proposed RAEE can significantly accelerate inference. RAEE also
achieves state-of-the-art zero-shot performance on 8 classification tasks.
|
RAEE: A Training-Free Retrieval-Augmented Early Exiting Framework for Efficient Inference: Deploying large language model inference remains challenging due to their
high computational overhead. Early exiting accelerates model inference by
adaptively reducing the number of inference layers. Existing methods require
training internal classifiers to determine whether to exit at each intermediate
layer. However, such classifier-based early exiting frameworks require
significant effort to design and train the classifiers. To address these
limitations, this paper proposes RAEE, a training-free Retrieval-Augmented
Early Exiting framework for efficient inference. First, this paper demonstrates
that the early exiting problem can be modeled as a distribution prediction
problem, where the distribution is approximated using similar data's existing
information. Next, the paper details the process of collecting existing
information to build the retrieval database. Finally, based on the pre-built
retrieval database, RAEE leverages the retrieved similar data's exiting
information to guide the backbone model to exit at the layer, which is
predicted by the approximated distribution. Experimental results demonstrate
that the proposed RAEE can significantly accelerate inference. RAEE also
achieves state-of-the-art zero-shot performance on 8 classification tasks.
|
[
0.0347394533,
0.0715949908,
0.0108757494,
0.022382332,
-0.0421732813,
0.0153751271,
-0.0221750811,
-0.012494158,
-0.0182842147,
-0.0259831809,
-0.0358679853,
-0.0232717395,
-0.0202648286,
-0.0101761762,
-0.008191783,
-0.023680076,
0.0296055768,
-0.0085233394,
0.0138005028,
-0.0131049491,
-0.008403522,
0.0282481778,
0.0121736564,
0.0449689701,
0.0252876915,
0.0201515183,
-0.0147394128,
0.0582188368,
-0.0125721712,
-0.0144148376,
0.0170123335,
0.0513835885,
0.0432397872,
0.0401306786,
0.0000019876,
-0.034775082,
-0.0319263227,
0.0158930309,
-0.0746392757,
-0.0183924902,
0.0377427861,
0.0038803495,
0.009663525,
0.0294902511,
0.0236600004,
-0.033911664,
0.0900304765,
-0.0026053288,
0.0025434329,
0.0408248752,
-0.0005713784,
-0.0415154696,
0.0000782153,
-0.0389225818,
-0.0143324984,
-0.0583796725,
0.0248882752,
-0.0320711397,
0.0113263801,
-0.0300414376,
-0.0116892504,
0.0549526438,
0.0340210274,
0.0337388515,
-0.0476557016,
0.0632941499,
-0.0461303443,
-0.0201433804,
-0.019949872,
0.0256337095,
-0.0207691435,
-0.0873011202,
0.0026094343,
-0.0250822399,
-0.0186286028,
0.0201809481,
-0.0353673659,
0.0333532542,
-0.0103660077,
0.0385609046,
0.0205594581,
0.020573169,
0.0191824436,
0.0352102779,
-0.0545395277,
0.0337392986,
0.02669839,
0.0050687133,
-0.0016103141,
0.0083673876,
-0.0270749144,
-0.0361920744,
-0.0099162869,
0.0047868909,
0.0262008552,
0.0144322626,
-0.0422810204,
-0.0277156606,
-0.0112929037,
-0.0555879138,
0.0353385359,
0.0568213277,
-0.0425595641,
0.0652448982,
-0.029124653,
0.0185805429,
-0.0458632186,
-0.0098950183,
-0.0450245887,
-0.0190824457,
-0.0255911406,
0.0022894004,
-0.0463869609,
0.030396495,
-0.027506385,
-0.019577384,
-0.0461146794,
0.0169668384,
0.0322320648,
-0.0387175418,
0.0163142048,
0.0313246883,
-0.049865678,
0.0035996768,
0.0054895575,
0.0231940672,
-0.0310240667,
-0.0125888241,
-0.0192152113,
-0.0616817661,
-0.0181292705,
0.05192158,
0.0310376547,
-0.0455998294,
0.0310731065,
0.0063805119,
0.0035043212,
-0.0539091639,
-0.0579942763,
0.0084285475,
-0.0032876313,
-0.0987440348,
0.0034809448,
-0.0885879248,
0.025064256,
0.0301804468,
-0.0163730644,
-0.019584313,
-0.0101788845,
0.0071352744,
-0.0222101621,
0.055439122,
-0.0022755589,
0.0242512133,
0.0181568991,
-0.018475797,
0.0028403078,
0.0928767323,
-0.0344642513,
0.0070885904,
0.0383080617,
0.0179621074,
0.0050101955,
0.0527482592,
0.0302092209,
0.0035226096,
-0.0687324926,
0.0087694582,
-0.0287923161,
-0.0342349187,
0.020458648,
0.0117601929,
-0.0294721406,
-0.0043752911,
0.0167191196,
0.0555619858,
0.0315449648,
-0.0268108342,
0.067893818,
0.0226289053,
-0.0204929747,
-0.0103617311,
-0.0612038821,
-0.0021733586,
-0.0311709661,
0.0044301697,
-0.0018859402,
0.0279659126,
-0.0558185354,
-0.0061863326,
-0.0134241618,
-0.0130511383,
-0.04419595,
0.0520555042,
0.0991945043,
-0.0044406806,
-0.0021800357,
0.0837300643,
0.0342854895,
-0.0334287323,
-0.0206493977,
-0.0758574829,
0.0089921989,
0.0283137579,
0.0239268076,
-0.0252334196,
-0.0407784954,
0.0636479557,
0.0139789935,
0.0172061883,
0.0005013805,
0.0045633069,
0.0250269771,
-0.0469424725,
-0.052756317,
0.0422211513,
0.0293582268,
0.0136320423,
0.0320151262,
-0.0014079447,
0.0353247151,
-0.0950071737,
0.0045942883,
0.0310117137,
-0.0400226302,
-0.0116604362,
0.0355238169,
0.0255248956,
0.0298142079,
0.0016479563,
-0.0168395117,
-0.0000854851,
0.039754197,
-0.0115739126,
-0.0234691594,
0.0117955701,
-0.003917173,
-0.0598534495,
0.0081982566,
0.0481030494,
0.0174033251,
0.0140291601,
-0.0184233766,
-0.0263929684,
0.0346709639,
-0.0150147527,
0.0809163377,
-0.0736192018,
0.0258765016,
-0.0218757521,
0.0081351064,
-0.0133544747,
0.0881780833,
-0.0171540603,
0.032553371,
0.0318421572,
-0.0242192596,
0.0120945694,
-0.0273111332,
-0.0455629565,
0.0004949818,
-0.0355128534,
0.0104497122,
0.0556132607,
-0.0059578437,
-0.0011437177,
0.0098671932,
-0.0611534864,
0.0296707042,
-0.02101155,
0.0273240563,
0.0074144998,
0.0330651179,
-0.0202547461,
-0.0433570892,
0.0217815842,
-0.0123687172,
0.0092998976,
0.0325255245,
-0.0375711434,
0.0186133403,
-0.0529691353,
-0.0242749248,
-0.033691369,
0.0080935312,
0.0933653861,
-0.0516974889,
-0.0630586967,
-0.0158547796,
-0.0209930651,
-0.050278578,
0.0672919527,
-0.0140563641,
-0.0387563035,
-0.0139791574,
-0.0626466796,
0.0049035624,
0.0284638219,
-0.0151921902,
0.0061142808,
0.0418167971,
-0.0017925286,
-0.072768867,
-0.0012537723,
-0.0480199642,
-0.0246063042,
-0.0081186621,
0.0076013468,
-0.0367443338,
0.0739388987,
0.0177646,
-0.0115086781,
-0.0422254428,
-0.0108625172,
-0.0081776129,
-0.0168506801,
-0.0178542621,
0.0295200441,
0.000631168,
0.0493157357,
-0.0305819754,
-0.0276365783,
0.0907047242,
-0.0213142596,
-0.002445661,
-0.0047422615,
-0.0209867563,
-0.0256970711,
0.0604337938,
-0.0046176654,
0.0099582998,
0.0347791538,
-0.0283848401,
-0.0503209457,
-0.0275765825,
0.0321896151,
-0.0220541824,
0.0404097624,
-0.0408838801,
0.0435621068,
0.0006255262,
-0.0257634707,
-0.0363705344,
-0.0313039199,
-0.0258697774,
0.0075329966,
-0.0536301658,
-0.0078524007,
-0.058126159,
0.0384359136,
-0.0060304143,
0.0535833426,
-0.0305081923,
0.0227057394,
0.0045149992,
0.0523194633,
0.0211790688,
-0.0312164649,
-0.0197169762,
-0.0527504832,
0.0293719973,
0.0357737131,
0.0129969744,
0.0507926829,
-0.041997198,
-0.0331815258,
0.06948746,
-0.0110370778,
-0.0501682125,
0.0399166495,
0.0098449532,
-0.0573756993,
-0.0191609636,
0.1104317829,
0.0099417018,
-0.0516087674,
-0.0384963192,
0.0019905905,
0.0559058264,
0.0319264084,
0.0350624546,
-0.0565902218,
-0.0281331446,
0.0198664553,
0.049494613,
0.0119443899,
-0.0107811345,
0.0099758869,
-0.1088843197,
0.0762380362,
-0.0214811917,
-0.0313923694,
0.0031322495,
-0.0416830294,
0.0387648642,
-0.0102866059,
-0.0429026261,
0.0287985764,
0.023268288,
-0.0102850609,
0.0226777289,
0.0047848541,
-0.0498196334,
0.0121371821,
-0.0451152176,
-0.0005552138,
0.0476619229,
0.0176199805,
0.0017035046,
0.0193792004,
0.0389665626,
-0.073153764,
-0.0138910515,
-0.01889419,
0.0690148398,
-0.0117263356,
0.0203722436,
-0.005877581,
0.0162952226,
0.0012954965,
-0.0061318865,
-0.0162941646,
0.0885957032,
-0.0264164731,
-0.023748843,
-0.0302800406,
-0.0143877333,
0.0013900764,
0.0720171407,
0.0253340732,
-0.0282060448,
-0.0264368951,
0.0154988207,
-0.0189727992,
-0.0340682417,
-0.0039861579,
0.0107648941,
-0.0229834858,
-0.085474059,
-0.0286426898,
0.0068801371,
-0.0025942377,
0.0601055063,
-0.0259514954,
0.0218456257,
-0.0374888331,
0.0014669212,
-0.0730010122,
0.0137238093,
-0.0319200084,
0.0142963231,
-0.050540749,
0.0210799798,
0.0311656334,
-0.03932124,
-0.035642162,
0.0163274407,
-0.0572671928,
-0.0596813187,
-0.012556796,
0.0287460536,
0.0253578909,
0.0447810143,
-0.0331701897,
-0.0335071944,
0.0757078528,
0.0572832637,
-0.0065773986,
0.0122036338,
0.0191226657,
-0.0693224743,
-0.0069831014,
0.039318312,
-0.0683371499,
0.0490933098,
0.0409343652,
0.0260220896,
-0.030784186,
-0.0513725281,
0.0268061403,
-0.004492254,
0.0597726181,
-0.0187150054,
-0.0642387867,
-0.0323976055,
0.0138203315,
-0.0076333405,
-0.050653886,
-0.0446761064,
0.0330865197,
0.0345338807,
0.0198784769,
-0.0030646469,
0.0222853348,
0.0174237322,
0.0494462736,
-0.0411086641,
-0.0020402884,
0.0652760491,
0.026751183,
0.0301755462,
-0.0238383804,
-0.0113265542,
-0.0198880695,
0.0146076465,
0.0239877291,
-0.0547739789,
0.0447177105,
0.0326524153,
0.0010945321,
-0.025158044,
-0.0068452507,
-0.0512786359,
-0.0269278493,
-0.0477360897,
0.0374933295,
0.0317064002,
-0.0046400744,
0.0150011936,
0.0029786131,
-0.0270651113,
-0.0402106494,
0.0366388224,
-0.0168378428,
-0.0404779054,
0.0405529328,
0.0336603522,
0.0002895801,
0.0153189851,
0.0360963829,
0.0293606762,
0.0504468083,
-0.019283032,
-0.0543901846,
-0.0121922251,
-0.0179996602,
-0.0117327897,
-0.0083309552,
-0.0246684626,
0.0248165987,
-0.0267934166,
-0.0826364309,
-0.0066210595,
0.0193651766,
-0.0227435045,
0.0216167439,
0.0269751586,
0.0813262165,
0.0151231289,
-0.0508628078,
-0.0372691341,
-0.0028374321,
-0.0168206245,
0.0072927373,
-0.0151195638,
-6.673797912e-33,
-0.0182648003,
0.003754402,
0.0328093581,
-0.0375454016,
-0.0095954733,
0.0278868135,
-0.0249599628,
-0.0125809442,
-0.0220257975,
-0.026146993,
0.0064107655,
0.0240777563,
0.0253925901,
0.0162824802,
-0.001108598,
0.0267962608,
0.0293791927,
0.0080034547,
-0.0171986669,
0.0222217608,
0.0379391946,
0.05718017,
0.0838503018,
-0.0187460817,
-0.0377960876,
0.0114369253,
0.015859805,
0.0407506712,
0.0130844424,
0.065775983,
-0.0016418147,
0.0206648912,
0.0118835522,
-0.0244914796,
0.0018296867,
-0.0163026545,
-0.0753525645,
-0.0372711867,
-0.00034659,
0.0084435726,
-0.0053904047,
0.0419425145,
0.027595818,
-0.0072166286,
0.0280568916,
-0.0283264667,
0.0433439352,
-0.0174545795,
0.0073521701,
-0.0223925449,
-0.0442415774,
0.0606953688,
0.0406311639,
-0.0395372771,
0.0477427654,
0.0587609373,
0.0071319132,
0.0213837605,
-0.0233822037,
-0.0095809652,
-0.0082353884,
0.0462408625,
0.0451465137,
0.0827011243,
-0.001790031,
0.0052399631,
0.0135662882,
-0.0289397743,
-0.0201237779,
-0.043930579,
0.0786481276,
0.0424680673,
-0.0089066364,
0.0223495681,
0.0634601638,
0.0306578893,
-0.0584272482,
0.0122573748,
0.0217746962,
0.0623208433,
0.0401710086,
0.0157004669,
-0.0528095141,
-0.0432963222,
-0.0351484194,
-0.0510989688,
-0.0489645712,
-0.0122615509,
0.0327215642,
0.0330774523,
-0.0415358134,
0.0060476712,
0.0688315779,
-0.0320028067,
-0.0067122225,
0.0456334651,
-0.0120200757,
0.0175151099,
-0.007402936,
0.0423984006,
-0.0552246086,
-0.0408866554,
-0.0091404514,
-0.0176748391,
-0.010375319,
-0.0058521344,
0.0207211655,
0.0512662753,
-0.0345396921,
-0.0060546421,
-0.0327996798,
0.0031019135,
-0.0180037953,
0.0484642647,
0.0351095982,
0.050358817,
0.0229510348,
0.0537188351,
0.0462596938,
0.0224767942,
-0.0191347562,
-0.0073452545,
0.0398441516,
0.0165500585,
-0.0408552848,
-0.0015668148,
-0.0386687703,
0.0312123895,
0.0043499726,
-0.0627126172,
-0.0173491407,
0.0685496405,
2.765e-7,
0.0726320595,
-0.032351695,
0.0238126684,
0.052054096,
-0.0416949391,
0.0034413091,
-0.0038872573,
-0.0238281246,
0.0415019691,
-0.0173498318,
0.0372998156,
0.0315425955,
0.0263178833,
0.0219277125,
-0.1163268909,
0.0238808692,
-0.0644972697,
-0.0000146706,
-0.0257931538,
-0.0586364977,
0.0267279334,
0.1012398079,
0.0311992895,
-0.0184811223,
-0.0085245781,
-0.0370449275,
0.0617143363,
0.0245061032,
0.0164644122,
0.0069298539,
0.0133319534,
0.0102988975,
0.0117727974,
0.0175863374,
-0.002290155,
0.0472990274,
-0.0140443509,
0.0135663953,
-0.0347225033,
-0.0076123504,
0.0304535646,
0.0091500422,
-0.0043573352,
-0.0469343849,
0.0820481256,
-0.0697598681,
0.0009787489,
-0.0263621025,
-0.041937802,
-0.0051682671,
-0.0386497304,
-0.0047424035,
-0.0091203619,
0.01786655,
0.0389189087,
-0.0510847755,
-0.0166659206,
-0.093271628,
-0.0219031535,
0.0226779133,
-0.0402112231,
-0.0162665434,
-0.0228099599,
-0.0270851497,
-0.005452544,
-0.0711953491,
0.0080303587,
2.801699868e-34,
0.0284251664,
0.0177857149,
-0.0107583376,
0.0203915909,
0.001913124,
-0.0230909865,
-0.0197497383,
0.033927422,
0.0132286642,
-0.0673855618,
-0.0401123278
] |
http://arxiv.org/abs/2405.15189v1
|
http://arxiv.org/pdf/2405.15189v1
|
SOAP: Enhancing Efficiency of Generated Code via Self-Optimization
| 2024-05-24T03:48:15 | 2024-05-24T03:48:15 |
Dong Huang, Jianbo Dai, Han Weng, Puzhen Wu, Yuhao Qing, Jie M. Zhang, Heming Cui, Zhijiang Guo
|
cs.SE
|
cs.SE, cs.CL
|
Large language models (LLMs) have shown remarkable progress in code
generation, but their generated code often suffers from inefficiency, resulting
in longer execution times and higher memory consumption. To address this issue,
we propose Self Optimization based on OverheAd Profile (SOAP), a
self-optimization framework that utilizes execution overhead profiles to
improve the efficiency of LLM-generated code. SOAP first generates code using
an LLM, then executes it locally to capture execution time and memory usage
profiles. These profiles are fed back to the LLM, which then revises the code
to reduce overhead. To evaluate the effectiveness of SOAP, we conduct extensive
experiments on the EffiBench, HumanEval, and MBPP with 16 open-source and 6
closed-source models. Our evaluation results demonstrate that through iterative
self-optimization, SOAP significantly enhances the efficiency of LLM-generated
code. For example, the execution time (ET) of StarCoder2-15B for the EffiBench
decreases from 0.93 (s) to 0.12 (s) which reduces 87.1% execution time
requirement compared with the initial code. The total memory usage (TMU) of
StarCoder2-15B also decreases from 22.02 (Mb*s) to 2.03 (Mb*s), which decreases
90.8% total memory consumption during the execution process. The source code of
SOAP was released in https://github.com/huangd1999/SOAP.
|
SOAP: Enhancing Efficiency of Generated Code via Self-Optimization: Large language models (LLMs) have shown remarkable progress in code
generation, but their generated code often suffers from inefficiency, resulting
in longer execution times and higher memory consumption. To address this issue,
we propose Self Optimization based on OverheAd Profile (SOAP), a
self-optimization framework that utilizes execution overhead profiles to
improve the efficiency of LLM-generated code. SOAP first generates code using
an LLM, then executes it locally to capture execution time and memory usage
profiles. These profiles are fed back to the LLM, which then revises the code
to reduce overhead. To evaluate the effectiveness of SOAP, we conduct extensive
experiments on the EffiBench, HumanEval, and MBPP with 16 open-source and 6
closed-source models. Our evaluation results demonstrate that through iterative
self-optimization, SOAP significantly enhances the efficiency of LLM-generated
code. For example, the execution time (ET) of StarCoder2-15B for the EffiBench
decreases from 0.93 (s) to 0.12 (s) which reduces 87.1% execution time
requirement compared with the initial code. The total memory usage (TMU) of
StarCoder2-15B also decreases from 22.02 (Mb*s) to 2.03 (Mb*s), which decreases
90.8% total memory consumption during the execution process. The source code of
SOAP was released in https://github.com/huangd1999/SOAP.
|
[
0.0234287605,
0.0382115431,
-0.0145919938,
0.010734953,
-0.0263096467,
-0.0392570086,
-0.0414432548,
0.009331407,
0.0292578395,
-0.0064163818,
-0.0049798735,
0.0319870636,
0.0048812055,
0.0478577837,
0.0093265539,
0.0151678361,
0.0539772511,
-0.0191371702,
-0.0284551885,
0.0103512229,
-0.0376004279,
-0.0035116421,
0.0408579856,
0.0109504629,
0.0182526559,
0.0152240973,
-0.0084562274,
0.0521925576,
-0.0075990399,
-0.0261387583,
0.0234351233,
0.0411787331,
-0.025985349,
-0.0063738986,
0.0000023601,
-0.0366062634,
0.012341287,
0.0379765928,
0.0110607017,
0.0929779857,
0.0204018615,
-0.0134924725,
-0.0287901051,
0.0596634634,
0.0123988567,
-0.0155265499,
0.0507688224,
-0.0808686912,
0.051166296,
0.0378248021,
0.007994568,
-0.0122784935,
-0.0368201248,
0.0208005793,
-0.0330413096,
0.0259696078,
-0.0321581848,
-0.0101286331,
-0.0318400227,
-0.0089272317,
-0.0050805458,
-0.0086243842,
-0.0016307092,
-0.0283511113,
0.0182835292,
0.0083253784,
-0.0646158457,
-0.0202386808,
0.0080815898,
-0.0010749177,
-0.0309064407,
-0.0359163955,
0.0170976389,
0.0012190469,
-0.0620539486,
-0.0571185872,
0.0079935892,
0.0263384022,
0.0183379278,
-0.0100842454,
0.0145299379,
0.0045021921,
-0.0093920436,
-0.007439238,
-0.0041255611,
0.0439344533,
0.0263824053,
0.0009706974,
-0.0165582467,
-0.0179594494,
0.0237032939,
-0.0760371163,
0.0294528194,
0.0453877188,
-0.0038105375,
0.0027125187,
-0.0515226349,
-0.0528055392,
0.0523888543,
-0.0812483653,
0.0581914373,
0.0868920907,
0.0409841835,
-0.0220780354,
-0.0416713022,
0.0505483598,
-0.000684632,
-0.0550691187,
-0.0216225572,
0.0186702125,
0.016992474,
0.0033587341,
-0.0351278372,
0.0425524563,
-0.0240856372,
0.0318911783,
-0.0495456904,
0.0389276482,
0.0250117797,
0.0539820045,
0.0048651388,
-0.0292764865,
0.0059074992,
0.0097472873,
0.0291243847,
0.0205760598,
-0.0197576303,
-0.0734175295,
-0.0185311735,
-0.0568232723,
0.0007897446,
0.0171528552,
-0.0456091389,
-0.0053720782,
0.0430903398,
0.0744822025,
-0.0008890166,
0.0115629425,
-0.0293948278,
-0.0497548766,
0.0377999544,
-0.0428692736,
-0.0216169432,
-0.0902985781,
-0.0541026518,
0.0098299282,
-0.0335689336,
-0.0064207101,
0.0122252395,
-0.00505138,
-0.0076609971,
-0.0009662474,
-0.0459961481,
0.0445855148,
-0.0118467575,
0.0321754105,
0.0401415564,
0.0575005412,
0.0181241818,
0.0214424673,
-0.0004323499,
-0.0733686313,
-0.0264717862,
0.0373249613,
-0.0039049215,
0.0149126379,
-0.0369190611,
0.036818292,
0.014296758,
0.0032123064,
0.0372768193,
0.0346256159,
0.0061635496,
0.0237394143,
-0.0080920067,
0.0341256037,
-0.0260263085,
0.0216634348,
0.0382202044,
0.0480309427,
0.0256202277,
-0.0291953851,
-0.0018655321,
0.0546724238,
0.0003400666,
0.004653513,
-0.0129493279,
0.0607995354,
-0.0014291373,
-0.0202211533,
-0.0229397956,
-0.0275706649,
-0.0382488593,
0.0229048859,
0.0354910418,
-0.0078612324,
-0.0667521134,
0.0738298744,
0.0267977826,
-0.1170208752,
-0.0663692281,
-0.04760997,
-0.0150478613,
0.0608113445,
0.0571472161,
0.0161609277,
-0.0060671777,
0.044089064,
-0.0304092132,
-0.0021953713,
0.0527751707,
-0.0100889355,
0.0178019013,
-0.0411750935,
-0.0216521509,
-0.0248939022,
-0.0182461068,
0.0581643581,
0.0648339912,
-0.0205697529,
0.042375531,
0.0175674334,
-0.0149498945,
0.0137677593,
-0.0188759509,
0.0127870198,
0.0470989086,
-0.0249720216,
0.030461533,
-0.0345879309,
-0.0365387872,
0.0011630123,
0.0588166043,
-0.0173819996,
-0.0072814403,
-0.0256349631,
-0.0275657047,
0.0087112105,
0.0137770995,
-0.0134834787,
-0.0021353713,
0.0239678845,
-0.0141707482,
-0.029827429,
0.0465204008,
-0.0440372899,
0.0491420962,
0.0265189968,
0.0599494167,
-0.0648585558,
0.0266949665,
-0.0286917593,
0.0320088416,
-0.0040436881,
0.0112416465,
-0.0060872138,
0.068270795,
0.0286208484,
0.0508974418,
0.0142833944,
-0.0068275081,
-0.0235346407,
0.0173752271,
0.0162396822,
0.0379949212,
-0.001955406,
-0.0445565283,
-0.119729571,
0.0696576089,
-0.0023851979,
0.013671359,
0.0344866104,
0.0527423024,
0.0161030572,
0.0113175586,
0.0169148576,
0.0247468576,
-0.0007023067,
0.012670109,
-0.0183443017,
-0.0508578867,
-0.072102949,
-0.0004247169,
-0.0315194167,
-0.0080324858,
0.0196148772,
-0.0458100736,
0.0178304166,
-0.0783554614,
0.0876690224,
0.0328017287,
0.01328097,
0.0011662623,
-0.0602518991,
-0.0233282354,
-0.0316124409,
0.0607189834,
0.0300748199,
-0.0008575367,
-0.010541724,
0.0601449721,
0.0071936571,
-0.0602057427,
0.0055355066,
0.0451660603,
-0.0225951299,
0.0176110286,
-0.0569943599,
0.0309280939,
0.0856903717,
0.0056232172,
-0.026239315,
-0.01183338,
-0.0488843098,
0.0345161483,
-0.0186109208,
0.0076313582,
-0.0029696531,
-0.0122640235,
0.0154307839,
-0.0099198949,
-0.0873557925,
0.0030728581,
-0.0572160967,
-0.022527026,
-0.002769734,
-0.0353449583,
-0.0486029088,
0.0312297791,
-0.0028991736,
0.0363322906,
-0.0500348955,
-0.018932661,
0.0225413218,
0.0256379154,
0.0005924435,
-0.0371937789,
0.0330477729,
-0.0087157208,
0.0184790604,
-0.0234932397,
-0.0164394658,
-0.0347926654,
0.0235688314,
-0.0009472063,
-0.0309739765,
0.0302130114,
-0.0347277336,
0.0371444039,
0.0293199345,
-0.0117190033,
0.0472932607,
0.0295622796,
-0.0373793356,
0.0126917753,
-0.0494968854,
-0.0385227911,
-0.0750824288,
-0.0200270023,
0.0421797782,
0.0096542025,
0.0515661836,
0.0211581774,
0.0168617442,
-0.0427432507,
0.0109045953,
0.0718186647,
-0.0018895388,
-0.0350456983,
0.0649202988,
0.072000511,
-0.0450386852,
-0.0411866792,
0.0478904247,
0.0325530954,
0.011506089,
0.0267453212,
-0.0351152755,
0.0678719282,
0.0686050877,
0.0689739883,
-0.0122681428,
-0.0212577097,
0.0108164549,
0.022037955,
-0.0253565777,
0.0255976655,
-0.0305661689,
-0.1161311418,
0.0298980866,
-0.0029318649,
-0.0219082516,
-0.0099005168,
-0.0520082787,
-0.0376619659,
0.0317156091,
-0.0246430486,
0.0540647395,
0.004613834,
0.0121863559,
-0.0299038272,
-0.0292205643,
0.0212697368,
0.0213011708,
-0.0078553399,
0.024300294,
0.0616072714,
0.00302167,
0.0037093668,
0.0145130726,
0.0474246778,
-0.0062780357,
-0.0156950429,
-0.0640173405,
-0.0042224508,
0.0587986074,
-0.0397205576,
0.0299010295,
-0.020612184,
-0.0442689732,
-0.0001917843,
0.0160381608,
0.0257148668,
-0.0007735768,
-0.0088639073,
-0.0470238924,
-0.0168070924,
-0.0147655755,
0.042615205,
-0.0009685283,
-0.0112328762,
-0.0533550382,
-0.0700727999,
-0.0494658947,
-0.0628864542,
-0.0121842176,
0.0035825956,
-0.0195086729,
-0.0628567785,
0.0181825515,
0.0629981458,
-0.0318785906,
0.0361471623,
0.0164217446,
-0.0163927171,
-0.014716655,
-0.0118338978,
-0.1062862128,
-0.0512179621,
-0.0334495157,
0.0087723061,
-0.0314323157,
-0.0518353768,
0.0511407778,
0.0473585762,
0.0235964321,
0.0001676242,
-0.0541399047,
-0.070960395,
-0.0394213796,
0.0996971726,
0.0132668605,
-0.0022177608,
0.0020272369,
-0.0197629761,
-0.008685533,
0.0203655455,
-0.0093670236,
0.0442728959,
-0.0129891457,
-0.0413882174,
0.0229184758,
-0.0408525318,
-0.0286331661,
0.0051429691,
0.0172131564,
0.0067422488,
0.0118209338,
-0.0341237076,
0.0166414548,
0.0698939785,
-0.0385327488,
-0.0357401781,
-0.0337604359,
0.0604927987,
0.0043912772,
-0.0406118818,
-0.0098105632,
-0.0401722267,
0.0130026918,
-0.0121222371,
-0.0126508726,
0.0385536887,
0.0406034887,
-0.0377314985,
-0.0005265993,
0.0328999385,
0.0602361821,
-0.0549788289,
0.0210920982,
0.0889091045,
0.0135715511,
-0.0152603285,
0.0155516258,
0.0385131612,
0.0120977722,
-0.0467176437,
-0.0158613678,
-0.007386934,
0.0563352443,
0.0192430373,
0.0262991916,
0.0095387362,
-0.0356307775,
-0.0218838025,
0.0009418542,
0.0271421373,
0.0411800779,
0.0196157526,
-0.0165514369,
0.0106163146,
-0.0346681215,
0.0226525161,
-0.0554921739,
-0.0185985528,
-0.0147907091,
-0.0049310867,
-0.0463498794,
-0.0208347254,
0.1060410812,
0.0278370436,
0.0102965925,
0.0154126259,
-0.0193201788,
-0.0390457176,
-0.0099546183,
0.0158128645,
-0.0297772326,
-0.0187414959,
-0.0297254287,
-0.0000617491,
-0.0537953749,
-0.012876655,
0.0279325191,
0.0539768524,
-0.0032467069,
0.0470365845,
0.0590819567,
0.022999309,
0.0008977909,
-0.0267887861,
0.0250763316,
-0.0172355231,
-0.0285953134,
0.0665370375,
-6.395942639e-33,
-0.0372232124,
-0.0175283793,
-0.0027413997,
0.0333356373,
-0.0011279229,
-0.0086513655,
-0.0752781406,
-0.0140178716,
0.0304244924,
-0.04230351,
0.0145633677,
0.011356838,
0.0168866049,
-0.0201475248,
0.0349733271,
0.0651353747,
0.0276996605,
0.0253793113,
-0.0536014177,
0.00240133,
0.0026906673,
0.020525666,
-0.0223984327,
-0.0340099968,
-0.1006410345,
0.0126706678,
0.0454355739,
-0.0164542645,
0.0336672924,
-0.0030782386,
0.0124165695,
0.0099495631,
-0.0204171129,
-0.0045180013,
-0.0352708064,
-0.0090551591,
-0.0245359968,
-0.0313149057,
0.0170154497,
0.0220740512,
0.0087745804,
-0.0852677748,
0.0423513465,
0.0313340835,
-0.0425332375,
0.0063279355,
-0.0259351302,
0.0110321781,
-0.0413124971,
-0.0477261245,
-0.0527861863,
0.013664186,
-0.0011411075,
0.0262935981,
0.0090578478,
-0.0389958173,
-0.020179389,
0.0220604204,
0.0498619713,
0.0129747642,
-0.0077947197,
0.0807620212,
0.0661464632,
-0.0078889066,
0.0238621868,
-0.013695416,
0.0608000644,
0.0281636324,
0.012576608,
0.0395532884,
0.067369841,
0.003351924,
0.0070508514,
-0.0242266953,
0.0201824773,
-0.0701570511,
-0.0359325781,
0.010076887,
0.0001928738,
-0.0359073542,
-0.0027421622,
0.0289464071,
0.0189253688,
-0.0245562866,
-0.0347012468,
0.0338862911,
-0.0183232054,
-0.0226528533,
0.0137667246,
-0.004386845,
-0.0412053056,
-0.0190347657,
0.0181385837,
-0.021979697,
0.0387857854,
0.0193861071,
-0.0488316976,
0.008242215,
0.0046772906,
0.0424881056,
0.0072584199,
-0.0431156158,
-0.0273678862,
-0.0171411335,
0.0726234242,
0.0286317542,
-0.0131078046,
0.0116823167,
-0.0813365206,
0.0231688805,
-0.0544810854,
0.0070640943,
-0.0014573443,
-0.0043908805,
-0.0379118957,
-0.0166596044,
0.0156534035,
0.0321465023,
0.0139723215,
-0.014023141,
-0.0486535132,
-0.0010411849,
0.0683557838,
0.0370453,
-0.0097187897,
0.0054929289,
-0.0791974962,
0.0113213258,
0.0393920392,
-0.0046486221,
-0.090439193,
0.0403281376,
3.082e-7,
-0.0177506153,
0.043288704,
-0.0224446487,
0.05169826,
0.0104239071,
0.0490900613,
0.0174794309,
0.0004756155,
-0.0121154571,
0.0037482434,
-0.0480323881,
0.0139424577,
0.0525455102,
0.0132608991,
0.0083315494,
0.0339768082,
0.0008590422,
-0.0169731695,
-0.0277755596,
0.0091314269,
-0.0358391702,
0.0309076346,
-0.0024718512,
-0.0341721997,
0.0247800462,
-0.0035061759,
-0.0081229825,
-0.0172844641,
0.0010379914,
0.0346496105,
-0.1003552973,
0.0297713317,
-0.0450224429,
-0.0124676926,
-0.0375683345,
-0.0295221619,
-0.0251125135,
0.0906323344,
-0.0084952787,
0.0517575219,
0.0981566533,
-0.0454994366,
-0.0437611192,
0.0204380769,
0.0327063501,
0.0257821381,
-0.0431807563,
-0.0384933129,
0.0234022345,
-0.0051726769,
-0.0282064173,
-0.0292231645,
0.0269959625,
0.0185477156,
0.0201372169,
-0.0251645781,
-0.0308009665,
-0.0281042363,
-0.0601846389,
0.0058462434,
-0.0098789381,
-0.0503456667,
-0.0414036736,
0.0442946889,
0.0258388352,
-0.027703695,
0.0022843785,
3.601374211e-34,
0.0350270234,
-0.0262536891,
0.0014605372,
0.0444230959,
0.0469811894,
-0.006525883,
-0.0316259414,
0.0150638521,
-0.0334939063,
0.007171791,
-0.0480745807
] |
http://arxiv.org/abs/2405.15185v1
|
http://arxiv.org/pdf/2405.15185v1
|
An Evaluation of Estimative Uncertainty in Large Language Models
| 2024-05-24T03:39:31 | 2024-05-24T03:39:31 |
Zhisheng Tang, Ke Shen, Mayank Kejriwal
|
cs.CL
|
cs.CL, cs.AI, cs.HC
|
Words of estimative probability (WEPs), such as ''maybe'' or ''probably not''
are ubiquitous in natural language for communicating estimative uncertainty,
compared with direct statements involving numerical probability. Human
estimative uncertainty, and its calibration with numerical estimates, has long
been an area of study -- including by intelligence agencies like the CIA. This
study compares estimative uncertainty in commonly used large language models
(LLMs) like GPT-4 and ERNIE-4 to that of humans, and to each other. Here we
show that LLMs like GPT-3.5 and GPT-4 align with human estimates for some, but
not all, WEPs presented in English. Divergence is also observed when the LLM is
presented with gendered roles and Chinese contexts. Further study shows that an
advanced LLM like GPT-4 can consistently map between statistical and estimative
uncertainty, but a significant performance gap remains. The results contribute
to a growing body of research on human-LLM alignment.
|
An Evaluation of Estimative Uncertainty in Large Language Models: Words of estimative probability (WEPs), such as ''maybe'' or ''probably not''
are ubiquitous in natural language for communicating estimative uncertainty,
compared with direct statements involving numerical probability. Human
estimative uncertainty, and its calibration with numerical estimates, has long
been an area of study -- including by intelligence agencies like the CIA. This
study compares estimative uncertainty in commonly used large language models
(LLMs) like GPT-4 and ERNIE-4 to that of humans, and to each other. Here we
show that LLMs like GPT-3.5 and GPT-4 align with human estimates for some, but
not all, WEPs presented in English. Divergence is also observed when the LLM is
presented with gendered roles and Chinese contexts. Further study shows that an
advanced LLM like GPT-4 can consistently map between statistical and estimative
uncertainty, but a significant performance gap remains. The results contribute
to a growing body of research on human-LLM alignment.
|
[
0.045715183,
0.1037964299,
-0.0206406806,
0.0074091195,
-0.0403920673,
-0.0115125049,
-0.0011435435,
-0.0139870411,
-0.0224142689,
-0.0420922935,
-0.008651196,
-0.0842839777,
0.061388731,
-0.0233129002,
0.0275496542,
-0.0408595391,
0.0278604273,
-0.0207524989,
0.0739655867,
-0.0214046016,
0.0130388001,
-0.0043389988,
-0.01568323,
0.0447511673,
-0.0009787079,
0.0152093079,
0.0148974899,
0.001222564,
0.0242197607,
-0.0295962691,
-0.0150978612,
0.0509699062,
0.0058833584,
0.0261011887,
0.0000018928,
-0.0719510913,
0.0197434556,
-0.0043201968,
-0.0211362075,
0.047544159,
0.0603032932,
-0.0102260262,
-0.0328474641,
-0.0145289628,
-0.0282448716,
0.0048703202,
0.0662797466,
-0.0042567258,
-0.0268617421,
0.0142949596,
-0.0020076111,
-0.002226552,
-0.0205853265,
-0.0095348526,
0.0054406426,
-0.0536889248,
0.0285156593,
-0.011378116,
-0.0140389204,
0.0085167894,
-0.0123289935,
0.0428839885,
0.0121719735,
0.0235654619,
-0.082125932,
0.0406881124,
-0.0206827447,
-0.0393748954,
-0.0088001667,
0.0064728744,
0.0518520288,
-0.0027488309,
-0.0398209095,
-0.0308437198,
0.0159722958,
0.0410735235,
-0.0166428089,
0.0237351321,
-0.0091955308,
-0.0106896739,
-0.0040687169,
0.0175752267,
-0.0021129716,
0.0224446133,
-0.0037462453,
0.1317960322,
0.0063142427,
-0.013679699,
0.0294884425,
0.0067169229,
-0.074154444,
-0.0437907912,
0.0053831963,
-0.015009189,
-0.0144950384,
0.0229397099,
-0.0227996483,
-0.0047450797,
0.0232796706,
-0.0469399765,
-0.0057901102,
0.047031749,
0.0120578529,
0.0021929531,
0.0213867519,
0.0558395125,
-0.003766055,
-0.0507226773,
-0.0217077099,
0.0139862066,
-0.0150031568,
-0.039965149,
-0.0151720187,
-0.0306817871,
-0.023029089,
-0.0062805009,
-0.094962731,
-0.0245925635,
-0.0247394685,
-0.0024327629,
-0.0036573149,
0.012221301,
-0.0472216122,
-0.020470826,
-0.0649848506,
0.016185645,
-0.0410920791,
0.0635739714,
-0.0237616319,
-0.0873130187,
0.0246299375,
0.0254443176,
-0.0143114878,
-0.0275002997,
-0.0262604859,
0.0362347998,
-0.0072281468,
0.0182067603,
0.0124368658,
-0.0251778625,
-0.060189724,
-0.0531578325,
-0.0518892333,
-0.0638221279,
-0.0101781869,
0.0127758356,
0.0096138325,
-0.0420940332,
0.0324823037,
0.0021923711,
-0.0057946565,
0.0772522539,
0.018206751,
0.036368683,
0.0162126515,
0.0039939685,
-0.0153414821,
0.0738932565,
0.0505691245,
0.0028331985,
0.0467804372,
-0.0058126296,
0.023894852,
0.0573272742,
-0.0280010011,
0.003941521,
-0.0199931394,
-0.0115688266,
-0.0194889512,
0.0383670181,
0.063385509,
0.0698027536,
-0.0018788561,
0.0084305527,
0.0860926285,
0.0610952936,
-0.036898341,
-0.0092472443,
0.043250978,
0.0374334157,
0.0366459303,
-0.0020059638,
-0.0471863449,
0.0081449756,
-0.0249520466,
0.0163591225,
0.001723035,
0.0220615231,
0.0098999357,
-0.0185859315,
-0.029348338,
0.0079304799,
0.0210051909,
0.0114581408,
0.0289133769,
-0.0604073741,
0.007368824,
0.0016000185,
0.0058933031,
-0.0097146984,
-0.0135088982,
-0.0577075854,
0.0127895018,
0.0188064352,
-0.0213663261,
0.022682352,
0.0228137709,
-0.0001390305,
-0.0017662996,
0.0453985408,
0.0531713106,
0.0232984945,
0.0008724229,
-0.0279744603,
0.006359444,
0.0112612899,
0.0348047614,
-0.0087227393,
0.0418755636,
0.0453819744,
0.0350621492,
-0.0477358177,
-0.0289176889,
0.00325012,
-0.0207005609,
-0.0031122791,
-0.0348598845,
-0.0277904756,
0.0161113497,
0.0032969182,
-0.0260076504,
0.044329267,
-0.0128490208,
0.0376713313,
-0.0202716812,
0.0295463204,
0.0034076576,
0.035391897,
0.0535083562,
-0.0250708666,
-0.0313827433,
-0.0186766498,
0.0122233806,
-0.0097136591,
0.0210409407,
-0.0611433722,
0.0222477894,
-0.0337825194,
0.0404367261,
-0.0258207694,
0.0361313485,
-0.0267456081,
0.0509714261,
-0.0024220126,
0.0143958842,
0.0164225679,
-0.0231734216,
0.0252101943,
-0.0666445941,
-0.0621996336,
0.0003405247,
0.0244978834,
0.0119303204,
0.0588305406,
-0.0238181427,
-0.0559985824,
-0.0304813571,
-0.0656961799,
-0.0169833209,
0.0152378865,
0.0480515845,
0.0085157016,
-0.0186778959,
-0.0411123782,
0.0293959342,
-0.0029088436,
-0.0277864262,
-0.0271903258,
0.0448988117,
-0.0013213998,
-0.0159398485,
-0.0128216408,
-0.0262359232,
-0.0032736291,
-0.0462365784,
-0.0032181598,
0.0360156186,
-0.0451725274,
-0.0570820607,
0.0420790873,
-0.0484402068,
-0.0162481107,
-0.0096809249,
-0.0526612401,
0.0311465915,
0.0058462378,
0.0654251128,
0.0250705034,
-0.0097194761,
0.0435751826,
0.0569628067,
-0.0591930524,
-0.0463357195,
-0.0203200411,
-0.0147437537,
-0.0388029926,
-0.0026779275,
-0.0918916687,
-0.045714777,
0.0294389743,
0.0497916155,
-0.0057431725,
-0.0407109037,
0.0573314056,
-0.0548460074,
-0.0701581612,
-0.0110441502,
-0.0133643532,
-0.0130374366,
0.0730266646,
-0.0039362367,
-0.0495642908,
0.0572130457,
-0.01555348,
0.0157434754,
-0.022407053,
0.0033051257,
0.0230412688,
0.0414020047,
-0.0065180771,
-0.001314893,
0.0136184925,
-0.0187679343,
-0.0596457161,
-0.007896224,
0.0114853056,
-0.0503694974,
0.0016767228,
-0.0810166821,
-0.0121749993,
-0.0033779331,
-0.0071923896,
0.0067302878,
-0.0552699268,
-0.0761076212,
0.0273924842,
0.041661758,
-0.0471272841,
-0.0074279648,
0.0255946461,
0.0387246013,
-0.019987233,
-0.0321182869,
-0.0136358663,
-0.0272716116,
-0.0453232415,
0.0482726693,
0.0092108054,
0.0484942757,
-0.0323851481,
-0.0071326895,
0.0373083614,
0.0122697726,
0.0645748228,
-0.0311172605,
-0.0170902479,
0.0454773866,
0.0192539431,
-0.0508057885,
0.0629385263,
0.0560101606,
-0.0271101389,
0.0255523007,
0.1262904555,
0.0186334364,
0.0107153337,
0.0106090056,
-0.0548742302,
0.0121889552,
0.0045170067,
0.0029012894,
-0.0708402023,
-0.0131777776,
0.005128358,
-0.0118291359,
-0.0659522936,
-0.0296833012,
-0.0228924602,
-0.1202528551,
0.0380881578,
0.0199573357,
-0.0161383729,
-0.018500749,
-0.0423496664,
0.0322795287,
0.0010154297,
0.0136214923,
0.0106889391,
0.0073328544,
0.0392416045,
0.0027597628,
-0.0776367933,
-0.0639052987,
-0.0035639277,
-0.0585678592,
0.0495566465,
0.0129642598,
0.0468356758,
-0.0572845191,
-0.020363396,
0.0309048183,
0.010701986,
-0.0135646248,
0.0156651177,
0.0043255202,
0.0231858771,
-0.0021846406,
-0.0648632124,
0.0081360713,
0.0024338816,
-0.0111974571,
0.0082192523,
0.0468933769,
-0.0307438597,
-0.0187339969,
0.0050641829,
0.0287238192,
0.0011199339,
0.0801928639,
0.0010345561,
-0.0228816159,
0.0079896394,
-0.0031259661,
0.0328424014,
0.0389220566,
-0.0209909659,
0.0144459149,
0.0208039116,
-0.044665847,
-0.0530570745,
-0.0507356077,
0.0156688783,
0.0102834515,
0.0191097371,
-0.0182491783,
0.0553236566,
0.010109121,
-0.0133960433,
-0.0006481787,
-0.0052213017,
0.0704364404,
0.0022806535,
0.036191579,
0.0315260142,
0.0270012114,
0.0024632902,
0.0591897555,
-0.0655616894,
-0.0688143671,
0.0242364183,
0.0138409408,
0.0394292921,
-0.0420616083,
-0.0065009906,
-0.0725285411,
0.0185161307,
0.0415418223,
-0.0158510059,
0.0269654896,
0.007636935,
0.0723624676,
0.0187661871,
0.0339611098,
0.0144279934,
-0.0293224715,
-0.0186533052,
0.0528534241,
-0.0503637306,
-0.078510657,
0.0344487242,
0.043753881,
0.0297398958,
-0.0555634722,
0.0046160202,
-0.0089655258,
-0.0274956934,
-0.0552673079,
-0.0369439945,
-0.0065256329,
-0.0293109249,
0.0496801436,
0.0291040968,
-0.0446617417,
0.0000173228,
-0.0130547369,
-0.0008385946,
-0.0226816945,
-0.0094609335,
0.0340135545,
0.0225438047,
0.015921792,
-0.0069660307,
0.0086081643,
-0.0169255678,
-0.0286253113,
0.0535491295,
0.0459365956,
0.0404516906,
-0.0035890031,
-0.0414524376,
0.0068944683,
-0.0141605334,
-0.0442330912,
-0.0811849162,
0.0025840532,
0.0258039627,
0.0084061436,
0.002476753,
0.0265110787,
-0.0573031642,
0.0034150756,
-0.0387473889,
0.0069347126,
0.0168820173,
0.0246339217,
0.0868877769,
-0.0071530165,
-0.0393612012,
0.0170351248,
0.0503216237,
0.0118660079,
0.0325869024,
-0.0404420607,
-0.081533879,
0.0125225335,
-0.0304867048,
-0.0262321495,
0.0126437573,
-0.0236978233,
0.0213567074,
-0.0866197348,
-0.0666082948,
-0.0177359413,
0.0129644629,
-0.0058967811,
-0.0289982222,
-0.017204633,
0.0641455874,
-0.0257646367,
-0.0420052409,
-0.0292660557,
0.0066797524,
-0.0139646148,
0.0223626513,
-0.0524329469,
-6.268450726e-33,
-0.0327674039,
-0.0455578268,
-0.0157335699,
-0.028029928,
-0.0083325366,
-0.0066403705,
-0.0124956593,
0.0612328425,
-0.0003037781,
0.0346890353,
0.0038592338,
0.0153617663,
0.0356156416,
-0.0154096037,
0.0143058402,
-0.0005972706,
-0.0084561659,
0.0153689738,
-0.0139832161,
-0.0976290852,
0.0368060656,
-0.0138022564,
0.0786459744,
-0.0854888707,
0.0861588195,
0.0393309109,
-0.0325689316,
0.0197858904,
0.0236558579,
0.0145982932,
-0.0058759525,
0.0695371255,
-0.0117519712,
-0.0365639664,
-0.0282745641,
0.0185088478,
0.0229236241,
-0.0453169197,
0.0795592517,
0.0360791869,
0.0336935669,
-0.0053221695,
-0.0072059655,
-0.0043157442,
-0.0834827349,
0.0358845703,
0.0245748516,
-0.0322051682,
0.0268965084,
-0.0026183452,
-0.0251231194,
0.0256682876,
0.0101511292,
0.0397636443,
-0.0029543324,
0.019267492,
-0.0504862405,
-0.0347762108,
-0.0286814943,
0.0235811006,
0.0151482811,
0.0343797468,
0.0540208705,
-0.0041514989,
0.0212032497,
0.030007327,
0.0564216524,
-0.0131460642,
-0.0437244698,
-0.0313468724,
0.0440957248,
-0.0025184054,
0.0062735123,
0.0886532962,
0.0067004878,
0.0124540413,
-0.0270438567,
-0.0045569651,
0.039547611,
0.0538511015,
0.047988683,
0.0260260142,
-0.0764420182,
-0.0562640689,
-0.0362270586,
-0.0277519003,
-0.0308746304,
-0.0080822511,
0.00184601,
-0.0492185578,
0.0095737381,
0.0297507979,
0.028935818,
-0.0099251559,
-0.0200115182,
-0.0216919091,
-0.0177730713,
0.0153042329,
-0.0000227503,
0.0602234751,
-0.0528863631,
0.0251754131,
0.0169558097,
0.0328752212,
0.0314258337,
0.0260386225,
-0.0283905137,
0.025783021,
-0.0654360652,
0.0101909442,
-0.0238135159,
-0.0354554355,
0.0097457217,
-0.065866895,
0.0406156145,
-0.0035630639,
-0.024730999,
-0.0238023475,
0.0184200257,
0.0182013959,
-0.0059180632,
0.0098682046,
-0.0136751076,
0.0003490924,
-0.0207365304,
-0.0198673103,
0.0035229682,
-0.0197218619,
0.0116267037,
-0.0260168537,
-0.0257287845,
-0.0113190655,
2.647e-7,
0.0446090214,
0.0432530828,
-0.0062144445,
0.0327524021,
0.0260749068,
0.0339608751,
-0.0047912197,
0.0005304683,
0.05230923,
-0.0313830599,
0.0121389739,
0.0111698294,
0.045361124,
0.0437039435,
0.0065128482,
-0.0051594875,
-0.0566134304,
0.0189466178,
-0.006970135,
0.0334983952,
0.0322179236,
0.0777042583,
0.0231477208,
0.0294795558,
-0.0348628759,
-0.0508839265,
0.0626880229,
-0.050973732,
-0.0278677866,
0.0504121371,
0.0316048153,
0.0329511464,
0.0134637374,
0.0606435686,
0.0060993987,
0.0140059078,
0.0556183383,
0.0519551188,
0.0029546758,
0.0503928736,
-0.0182320625,
-0.0305154826,
-0.0077444986,
-0.0643375441,
0.0361392908,
-0.0358254388,
-0.0581670962,
0.0247678794,
-0.0108301807,
-0.0330756381,
0.0360758714,
-0.0293450616,
-0.0186616983,
-0.0023193713,
0.0107120415,
0.0298599545,
-0.0427691266,
-0.005275683,
-0.0430271327,
-0.0009272577,
-0.0319129154,
-0.0918464884,
-0.0037334869,
0.0133908121,
0.0202436186,
-0.0651591271,
0.0309979357,
2.170713786e-34,
0.0351823084,
0.0450568199,
-0.028221922,
0.0066698203,
0.0754945204,
0.0287779048,
0.0176250767,
0.0231286585,
0.062284112,
-0.0500942357,
-0.0336785093
] |
http://arxiv.org/abs/2405.15179v2
|
http://arxiv.org/pdf/2405.15179v2
|
VB-LoRA: Extreme Parameter Efficient Fine-Tuning with Vector Banks
| 2024-05-24T03:24:34 | 2024-05-27T18:51:57 |
Yang Li, Shaobo Han, Shihao Ji
|
cs.CL
|
cs.CL
|
As the adoption of large language models increases and the need for per-user
or per-task model customization grows, the parameter-efficient fine-tuning
(PEFT) methods, such as low-rank adaptation (LoRA) and its variants, incur
substantial storage and transmission costs. To further reduce stored
parameters, we introduce a "divide-and-share" paradigm that breaks the barriers
of low-rank decomposition across matrix dimensions, modules and layers by
sharing parameters globally via a vector bank. As an instantiation of the
paradigm to LoRA, our proposed VB-LoRA composites all the low-rank matrices of
LoRA from a shared vector bank with a differentiable top-$k$ admixture module.
VB-LoRA achieves extreme parameter efficiency while maintaining comparable or
better performance compared to state-of-the-art PEFT methods. Extensive
experiments demonstrate the effectiveness of VB-LoRA on natural language
understanding, natural language generation, and instruction tuning tasks. When
fine-tuning the Llama2-13B model, VB-LoRA only uses 0.4% of LoRA's stored
parameters, yet achieves superior results. Our source code is available at
https://github.com/leo-yangli/VB-LoRA.
|
VB-LoRA: Extreme Parameter Efficient Fine-Tuning with Vector Banks: As the adoption of large language models increases and the need for per-user
or per-task model customization grows, the parameter-efficient fine-tuning
(PEFT) methods, such as low-rank adaptation (LoRA) and its variants, incur
substantial storage and transmission costs. To further reduce stored
parameters, we introduce a "divide-and-share" paradigm that breaks the barriers
of low-rank decomposition across matrix dimensions, modules and layers by
sharing parameters globally via a vector bank. As an instantiation of the
paradigm to LoRA, our proposed VB-LoRA composites all the low-rank matrices of
LoRA from a shared vector bank with a differentiable top-$k$ admixture module.
VB-LoRA achieves extreme parameter efficiency while maintaining comparable or
better performance compared to state-of-the-art PEFT methods. Extensive
experiments demonstrate the effectiveness of VB-LoRA on natural language
understanding, natural language generation, and instruction tuning tasks. When
fine-tuning the Llama2-13B model, VB-LoRA only uses 0.4% of LoRA's stored
parameters, yet achieves superior results. Our source code is available at
https://github.com/leo-yangli/VB-LoRA.
|
[
0.0506548509,
0.0426911898,
-0.0220848341,
0.0487991385,
0.0382482708,
0.0346862078,
0.0169790406,
-0.0230000317,
-0.0217410661,
-0.0186158642,
-0.0752379149,
-0.0512959324,
0.0111361686,
0.0250953827,
0.0435430296,
-0.0347188339,
0.0549986772,
-0.049893707,
-0.0071171098,
0.0098620104,
0.00049533,
0.015665235,
0.003354942,
0.0277013034,
0.096659258,
0.0029556591,
-0.0076212827,
0.0172862466,
0.0358990878,
0.045344796,
0.0281946361,
0.0486174449,
0.0287566148,
0.0099270688,
0.0000021745,
0.001885641,
-0.0166439749,
-0.0250289664,
-0.0131317312,
0.0491967164,
0.0489466414,
0.0224904101,
-0.0250598863,
0.0204056278,
-0.0231990758,
-0.0130770234,
0.0593640506,
-0.0220793039,
0.075457029,
0.0574421436,
0.0021747935,
-0.0273452513,
0.0019743666,
-0.0212961398,
-0.027395431,
-0.0522712246,
0.0029025075,
-0.0422635749,
-0.006697516,
-0.0179384965,
-0.0263157245,
0.0233116169,
-0.0023261306,
0.0076911827,
-0.0006897399,
0.0113828629,
-0.021571951,
-0.0129287364,
-0.0270656571,
0.0914100409,
-0.0623235106,
-0.0110149253,
0.0459705181,
-0.0050731944,
-0.0089457668,
-0.0682256967,
-0.0292793568,
0.0296393614,
0.0018795378,
-0.0114490949,
0.0330948047,
-0.0348249562,
0.0037887029,
0.0130312946,
-0.0283824112,
0.0444833562,
0.0435329042,
-0.0011798993,
-0.0359583013,
-0.0417755432,
-0.0214431565,
-0.0223914664,
0.0188735295,
0.0236751735,
-0.0403338484,
0.0008511712,
-0.1082247421,
-0.0612502433,
0.0053125373,
-0.0397202373,
0.1335743815,
0.0812561214,
-0.0105525339,
0.0256660152,
-0.0453867391,
0.0451622307,
-0.0421832316,
0.0260374714,
-0.0458535999,
0.0188969132,
-0.0092516206,
0.0112916259,
-0.0311716311,
0.015138871,
-0.1120518595,
-0.0120950285,
-0.0393466465,
0.053785935,
0.0034634969,
-0.0269115642,
-0.0168367941,
0.0158283748,
-0.0377367064,
-0.0108485082,
0.0012917856,
0.0274185631,
-0.0028775409,
0.0032337906,
-0.0131438253,
-0.0520239547,
-0.0273155812,
0.0065995338,
0.0132620884,
-0.0355795957,
0.0332717299,
0.0497424528,
-0.0433570705,
-0.0336107872,
-0.0378607847,
-0.0106332032,
-0.0071065463,
-0.0801242664,
0.0226107202,
-0.0291052442,
0.0493415259,
-0.0035267137,
-0.0102695646,
-0.0260883179,
-0.0168030355,
0.0313503183,
-0.0538991354,
0.0288580656,
0.0116358036,
-0.0090987356,
0.0262839496,
0.0069545205,
0.0212983824,
0.0582232177,
0.017940836,
0.0256727543,
-0.0274085272,
-0.0132754995,
-0.0397563651,
0.035708759,
0.0127057051,
-0.0050052386,
-0.011064522,
-0.0013573762,
0.006151245,
-0.0004063363,
0.064217627,
0.0214593373,
-0.019301651,
0.001368108,
0.0379810072,
0.0232825764,
0.0428537838,
0.0569045618,
0.0582557246,
-0.006777355,
-0.0022048254,
-0.0271280035,
-0.0299933348,
0.0809305161,
-0.0383186638,
-0.0105527621,
0.0372431912,
0.1104841605,
-0.0317742229,
-0.0283719916,
-0.0073240222,
0.0145619456,
0.0033463892,
0.0024188764,
0.0037113475,
-0.0071978625,
-0.0586241633,
0.0489630476,
0.040080674,
-0.0762779787,
-0.0106084859,
-0.0570742376,
0.0247209892,
0.0481932461,
0.0252791028,
0.0133473352,
-0.0308509097,
0.0097092232,
0.0412014723,
0.0089746611,
0.0414940305,
0.0302305017,
0.0154581927,
-0.0811282545,
-0.0171334911,
-0.0222961325,
0.0189737529,
0.0297614709,
-0.0037478518,
-0.003859004,
0.0295744371,
-0.0157268308,
-0.0204893332,
0.0173182189,
-0.0178708825,
0.0252687763,
0.0145049645,
-0.0012670862,
0.034941297,
0.00326978,
-0.0036668016,
-0.012304496,
0.0136431055,
-0.0305958092,
0.0237410292,
-0.0106495712,
-0.0502783619,
-0.0183432363,
0.007355737,
-0.0021948107,
0.0419363119,
0.0487857871,
-0.0340996385,
-0.0406026132,
0.0559228882,
-0.0436680727,
-0.0211912245,
-0.0251918137,
0.0341552757,
-0.0147508075,
0.0504616983,
-0.0176907852,
0.0457975678,
-0.0062786401,
-0.0064209099,
0.0721542314,
0.0250564925,
0.0092416778,
-0.0236647762,
-0.0309759621,
0.0157379489,
0.0084319012,
0.0354062058,
-0.0200241394,
0.0255452059,
-0.0303960983,
-0.0268848278,
-0.060298074,
0.0585631579,
-0.0418036059,
0.0368603654,
0.0348092318,
0.0328192525,
0.0180086792,
-0.006603119,
-0.0279460996,
0.0346076451,
-0.0075835097,
0.0215424336,
-0.0026239944,
0.0097310403,
-0.0438573062,
-0.0465319306,
-0.0153332688,
0.0115369288,
0.0220738482,
-0.0431245677,
-0.0129954796,
-0.0694651678,
0.01432632,
-0.0418628231,
0.0427996032,
-0.0287820324,
-0.0531257726,
-0.0198473241,
-0.0206840821,
0.0595188029,
0.0464934707,
-0.0256161559,
0.0210198741,
0.0559850596,
0.0099062501,
-0.0211078264,
0.0415609628,
0.0043166615,
-0.0429614,
-0.0115347588,
0.0153446235,
0.0017924076,
0.0703707635,
-0.005151093,
-0.0164896064,
-0.0866779238,
0.0216679834,
0.007927197,
-0.0331899673,
-0.0105756177,
0.0073253345,
-0.0252635516,
-0.0099860011,
-0.0389767103,
-0.0135647207,
0.0595292524,
-0.0261786617,
-0.0383115672,
-0.0353218019,
-0.0518524088,
-0.0388278887,
-0.0300854426,
-0.0105386199,
0.0390427224,
-0.0400396958,
-0.0036754301,
-0.0405592732,
0.0237440653,
-0.0032548073,
-0.0177059993,
0.0335827246,
-0.0496435128,
-0.0009785703,
-0.0385243371,
-0.0453735553,
0.005838179,
-0.0266549755,
0.0022631211,
0.017023297,
0.0468154326,
-0.0391029455,
0.0011968294,
0.0166362487,
0.0181195457,
0.0545459539,
-0.0407728292,
-0.0043209791,
-0.0594581179,
0.0134272287,
0.0626287907,
0.0660591722,
0.0245758872,
-0.0220676307,
0.0125844041,
0.007201768,
0.0361999758,
-0.0115536228,
-0.0267551448,
0.0598911531,
0.0438846461,
0.0197964422,
-0.0447421335,
0.0300721489,
0.0436558612,
-0.0451658256,
-0.0275628045,
0.0655508637,
0.0204249751,
-0.0398396216,
0.0074048671,
-0.0298584178,
0.0714529008,
0.0373422951,
0.0479688123,
-0.0563504621,
-0.0312586538,
-0.0056768986,
0.0458955616,
0.0251508262,
-0.0114889247,
-0.0310939681,
-0.0692214742,
0.034895815,
0.0315613076,
-0.0193892084,
-0.0238786079,
-0.0059322906,
-0.0058511612,
0.0138453115,
0.0386242121,
0.0309128538,
-0.0252015553,
0.0017765447,
0.0186059773,
-0.0133687006,
-0.0661537573,
0.0031677717,
0.0023109291,
0.0602593981,
0.0359951556,
0.0150163695,
0.0397042893,
-0.0545460694,
-0.0156422853,
-0.0849774033,
0.0244239569,
-0.0263054352,
0.0196017772,
0.0246340577,
-0.0507091768,
0.0019067012,
0.0330682471,
-0.0191308931,
0.0176821407,
-0.0505603068,
0.0575929768,
-0.0157707259,
0.0096332682,
0.0010007973,
-0.0115762865,
-0.0271839797,
0.041991163,
0.0423723347,
-0.0361097939,
-0.0205719024,
0.0038977964,
-0.0271724369,
-0.0703056976,
-0.033748921,
-0.0095136417,
-0.0385578312,
0.0320321694,
-0.0290882681,
-0.0039702877,
0.0443877093,
0.0529341176,
-0.0314722992,
-0.0041548177,
-0.0285524055,
-0.02046277,
-0.0462495685,
0.0225907285,
-0.0437380187,
-0.0024434812,
-0.0291997176,
-0.0302297492,
0.0363276675,
-0.023115389,
-0.0223823078,
0.0384232216,
-0.061436139,
-0.0907095596,
0.0209077653,
0.0644672513,
0.0209969115,
0.0443557315,
0.0237524454,
-0.0752524957,
0.0352757424,
0.0010975623,
-0.0014769908,
0.082807146,
0.0104656303,
-0.0207373053,
-0.0166435093,
0.0387382098,
-0.0655742288,
-0.0099514741,
0.0327744521,
0.0399803706,
0.0157495849,
-0.0276228301,
-0.0263281204,
0.0189785957,
0.002705774,
-0.0022493743,
-0.06306234,
-0.0281251017,
0.0553367287,
-0.0162704289,
0.004360341,
-0.0605141334,
-0.0290296692,
0.0068129352,
-0.0193480365,
-0.0009338512,
0.0040191128,
0.0256622024,
0.0379334502,
-0.0243180022,
0.0045934254,
0.0033724788,
0.0003393377,
0.0359606966,
0.0172104109,
-0.0226757787,
0.0416106582,
0.024513673,
0.0342724547,
-0.0397193283,
0.031004319,
-0.0226642396,
0.0017985835,
-0.0797008798,
0.0952121764,
-0.0361860543,
-0.1188602298,
-0.0465019122,
0.0054561202,
0.0606754832,
0.0218024794,
0.1006081775,
0.0221682917,
-0.0171403084,
0.0103939483,
0.0072813919,
-0.0113802543,
-0.018600557,
0.0371399224,
0.0070406478,
-0.0064389105,
-0.0357671119,
0.0511805229,
0.0502990969,
0.0577871948,
-0.047563985,
0.0345381983,
-0.0213246904,
-0.0192766562,
0.0201423671,
0.0213460587,
0.0071137915,
-0.0325679183,
-0.0191288795,
-0.0419909172,
-0.0060976087,
-0.0097026071,
-0.0060548601,
-0.050089974,
0.0301707629,
0.0448576175,
-0.0017247213,
-0.0188981686,
-0.0080816364,
-0.0524505749,
0.0032787754,
-0.0193782654,
0.0155374138,
-6.930641959e-33,
-0.0672301427,
0.0371369533,
0.0304539315,
0.0123040453,
-0.0137625365,
-0.0126803061,
-0.0357934088,
0.0058366447,
-0.0124534899,
-0.0229569245,
-0.0110135069,
-0.0110981409,
0.0192640107,
-0.0089249332,
-0.0143205598,
0.0143814264,
0.0618841164,
-0.0106981816,
-0.026731329,
0.0508630387,
0.0275015477,
0.0473192185,
0.037989419,
-0.013924405,
-0.0399064757,
0.0136594865,
-0.0169374272,
-0.0091274846,
-0.0175493434,
0.0295591522,
-0.028174486,
0.050327871,
0.0064452402,
-0.0263381302,
-0.0532311499,
-0.0020382521,
-0.0618281141,
-0.0203355625,
0.0586834736,
-0.0042826994,
-0.0168003719,
-0.0617072359,
0.0523996204,
0.0243500434,
-0.0330233462,
-0.0078344457,
0.0007246937,
-0.020042805,
-0.0249241441,
-0.0172228385,
-0.0721170828,
-0.0145105487,
0.0025709227,
-0.0068445518,
0.0602106154,
0.0263442658,
-0.0267603919,
-0.0093716998,
-0.0650114268,
-0.0125232656,
0.0269345623,
0.0602034777,
0.0788384229,
0.0220788587,
-0.0384646468,
0.0133468434,
0.0370232873,
0.0098597575,
0.0293623451,
0.0000997841,
0.0376668349,
0.0507749729,
0.0111501133,
0.0446202271,
-0.0080265142,
-0.0071124355,
-0.0406062454,
0.0138623491,
0.0170420092,
0.0372868516,
0.0365591049,
0.0339158997,
-0.0138070444,
-0.0307036508,
-0.0374637395,
0.0113257207,
-0.0635782182,
-0.0085096126,
0.0529064946,
0.0350925177,
-0.0315171778,
-0.0303536821,
0.0471408851,
-0.01758161,
0.0445310511,
0.0350044109,
-0.018218562,
0.0237796064,
0.025329873,
0.0288915467,
-0.0534026213,
-0.0797164664,
0.0363420993,
-0.0185286943,
0.0124945696,
-0.0036021753,
-0.0060427273,
0.0539183058,
-0.0623836219,
0.008124711,
-0.0328597315,
0.0255980287,
-0.0041682972,
0.0721897855,
-0.0153308641,
-0.0227098037,
-0.0142078511,
0.0336821079,
-0.0066340091,
0.0149155725,
-0.0178481322,
-0.0384537727,
0.0307235252,
0.0462774634,
-0.0367655009,
0.0296693686,
-0.0300131328,
0.0013758547,
0.031138869,
-0.0601451509,
-0.0279734675,
0.0090684146,
2.99e-7,
0.0070209391,
0.054154478,
-0.0371303149,
-0.0314950086,
0.0063659442,
0.0274321064,
-0.0045272429,
0.0235663857,
0.0383233838,
0.027505707,
-0.0102762887,
0.0265721623,
0.0227992758,
-0.0248957481,
-0.0303788222,
0.0302912425,
-0.0649114624,
0.0426504798,
-0.057981465,
-0.0628988892,
0.0134835104,
0.0778573453,
-0.0192038324,
-0.0325448252,
0.0077070771,
0.0098166876,
0.0491290428,
0.0123032453,
-0.0172633547,
-0.0051488532,
0.0539034121,
0.025269486,
-0.0642891228,
-0.0267119445,
-0.0105810454,
0.043253731,
-0.0377816297,
0.0864081606,
-0.0192269739,
-0.0036728065,
0.0650268123,
0.0233377684,
-0.015211883,
-0.0569918752,
0.0599442981,
-0.0614362247,
0.0086889481,
-0.0291917119,
-0.0103824139,
0.0066488083,
-0.0165014509,
-0.0412861742,
0.0208667554,
0.0124797598,
0.0102053089,
-0.0157362353,
-0.013701099,
-0.0883570611,
-0.0024415939,
0.0343829431,
-0.0307243634,
-0.0482341312,
-0.0792495832,
-0.0082036881,
0.0250701811,
-0.0794775635,
0.004626838,
3.015321859e-34,
0.0088475412,
-0.0091853356,
0.0600912608,
0.0341404453,
0.0241726786,
-0.0019475005,
0.002549188,
0.0233990252,
0.0162529126,
-0.112571694,
-0.0380106419
] |
http://arxiv.org/abs/2405.17470v1
|
http://arxiv.org/pdf/2405.17470v1
|
Athena: Efficient Block-Wise Post-Training Quantization for Large Language Models Using Second-Order Matrix Derivative Information
| 2024-05-24T03:14:29 | 2024-05-24T03:14:29 |
Yanshu Wang, Wenyang He, Tong Yang
|
cs.LG
|
cs.LG, cs.AI, cs.CL
|
Large Language Models (LLMs) have significantly advanced natural language
processing tasks such as machine translation, text generation, and sentiment
analysis. However, their large size, often consisting of billions of
parameters, poses challenges for storage, computation, and deployment,
particularly in resource-constrained environments like mobile devices and edge
computing platforms. Effective compression and quantization techniques are
crucial for addressing these issues, reducing memory footprint and
computational requirements without significantly compromising performance.
Traditional methods that uniformly map parameters to compressed spaces fail to
account for the uneven distribution of parameters, leading to substantial
accuracy loss. In this work, we propose Athena, a novel algorithm for efficient
block-wise post-training quantization of LLMs. Athena leverages Second-Order
Matrix Derivative Information to guide the quantization process using the
curvature information of the loss landscape. By grouping parameters by columns
or rows and iteratively optimizing the quantization process, Athena updates the
model parameters and Hessian matrix to achieve significant compression while
maintaining high accuracy. This makes Athena a practical solution for deploying
LLMs in various settings.
|
Athena: Efficient Block-Wise Post-Training Quantization for Large Language Models Using Second-Order Matrix Derivative Information: Large Language Models (LLMs) have significantly advanced natural language
processing tasks such as machine translation, text generation, and sentiment
analysis. However, their large size, often consisting of billions of
parameters, poses challenges for storage, computation, and deployment,
particularly in resource-constrained environments like mobile devices and edge
computing platforms. Effective compression and quantization techniques are
crucial for addressing these issues, reducing memory footprint and
computational requirements without significantly compromising performance.
Traditional methods that uniformly map parameters to compressed spaces fail to
account for the uneven distribution of parameters, leading to substantial
accuracy loss. In this work, we propose Athena, a novel algorithm for efficient
block-wise post-training quantization of LLMs. Athena leverages Second-Order
Matrix Derivative Information to guide the quantization process using the
curvature information of the loss landscape. By grouping parameters by columns
or rows and iteratively optimizing the quantization process, Athena updates the
model parameters and Hessian matrix to achieve significant compression while
maintaining high accuracy. This makes Athena a practical solution for deploying
LLMs in various settings.
|
[
-0.012545377,
0.0796339959,
-0.001732575,
0.0343495347,
0.0144236945,
0.0145519562,
0.019679321,
0.0271079727,
-0.0150671881,
-0.0373978019,
-0.0009294682,
-0.047929775,
0.0024985406,
-0.011001138,
0.0290773772,
-0.0403769128,
0.0331189036,
-0.0095084431,
-0.0091795772,
-0.0193200409,
-0.0143798487,
0.0256014057,
0.0469893292,
-0.004313685,
-0.0263403337,
-0.0239636302,
-0.0201742165,
0.0187415164,
0.0090318369,
-0.017825013,
0.0049914666,
0.0517812893,
0.061945647,
0.0226874556,
0.000002244,
0.0011636948,
-0.0432836078,
0.0045512933,
-0.0363520309,
0.057615757,
0.0814264342,
-0.0209985934,
-0.0277085621,
-0.0028959143,
-0.0444015339,
-0.0389253683,
0.0566080511,
-0.0047497824,
0.0964477509,
0.0378846824,
-0.0127981883,
-0.0341830738,
0.0089045446,
-0.0221869163,
-0.0225013215,
-0.0312850885,
0.0030134399,
-0.0075631933,
-0.0550475083,
0.020109335,
0.0095688533,
0.029701123,
-0.0049431403,
0.0297711398,
0.0139553882,
0.0144625809,
-0.0351149216,
0.0095426179,
-0.0215289108,
0.0595243834,
-0.0407101139,
-0.048613809,
-0.0096144704,
0.0316194333,
-0.0036050575,
-0.0515953265,
-0.0243769269,
0.0242671389,
0.02583896,
0.011329365,
-0.0213433281,
-0.0349644087,
0.0176429804,
0.0130100679,
-0.0434793122,
0.048982773,
0.0128715783,
-0.0238489155,
-0.0308956653,
-0.0294161178,
-0.0678182021,
-0.0124275107,
0.0255271178,
-0.0107282139,
0.0218095016,
-0.016363848,
-0.0754409209,
-0.0861471072,
-0.0275067054,
-0.0434660651,
0.0844406709,
0.0262787417,
0.0696985647,
0.024062831,
-0.0336825252,
0.0500697866,
-0.0087051205,
0.0027855979,
-0.0381894223,
0.0254464,
-0.0244307276,
-0.0208617616,
-0.0350403488,
0.0276060682,
-0.0294609908,
-0.0107894158,
-0.0274765491,
0.0205098316,
-0.0320136882,
-0.0077962698,
0.0168547779,
0.0433108695,
-0.0129904011,
-0.0400507078,
0.0246907305,
-0.0045731496,
-0.0050064046,
0.0104517527,
0.0042641796,
-0.0196750835,
-0.0202538446,
0.0169890374,
0.0370994322,
0.0076190443,
0.0286365133,
0.0319386646,
0.0173709076,
-0.0133647695,
-0.0023804628,
-0.0082272561,
-0.0052917283,
-0.0554251485,
-0.0296453983,
-0.0644220188,
0.0427322499,
0.0074715461,
-0.0183656346,
-0.054748226,
0.0049733608,
0.0007930942,
-0.0497964025,
0.0433603525,
-0.0267591402,
0.0132811777,
0.0555570312,
-0.0131053869,
0.0477442481,
0.0646516681,
0.0336258188,
0.0645661727,
0.0251933895,
-0.0008700082,
-0.0317235515,
0.0337556899,
0.0367844962,
0.0282619782,
-0.0493193157,
-0.0162884612,
0.0191967171,
0.0560869388,
0.0845077559,
0.0679472759,
-0.0458647422,
-0.0221439116,
0.0585570559,
0.0648512691,
0.0286231432,
0.0608042106,
0.0104434406,
0.0179276858,
0.0029675083,
-0.039774321,
-0.0000731472,
-0.0007283053,
0.0343477167,
-0.0052521024,
0.0510336943,
0.063714169,
-0.0191837884,
-0.0335996225,
-0.019372385,
-0.0174051821,
-0.0027539781,
0.0020728863,
-0.008940299,
-0.0127609465,
-0.0220071692,
0.1165531427,
0.0115439845,
-0.1024144888,
-0.0148835946,
-0.0304656904,
-0.0067057391,
0.049455978,
0.0199259836,
-0.0249747168,
-0.0188812017,
-0.0057035359,
0.0142657012,
-0.0156164207,
-0.0224202201,
0.0115397191,
0.009481756,
-0.0489149615,
0.0049625402,
0.0240482986,
0.0309913158,
0.0069786445,
0.0131287873,
0.0152234854,
0.0181610845,
-0.0301391575,
-0.0586634949,
0.0173034649,
0.0590966754,
-0.036185585,
0.0706613362,
-0.0061024395,
0.0305449124,
0.0293708332,
-0.0096242698,
-0.0269323476,
0.0426852852,
-0.0636936799,
0.0319756083,
-0.0570035577,
-0.004071279,
-0.0269448608,
0.0416899323,
0.0081033967,
0.033403527,
0.0873322412,
-0.045362968,
-0.0363420732,
0.0235156752,
-0.0300626699,
0.0149946036,
-0.0415211506,
0.0178994257,
-0.016030347,
0.0685115308,
-0.0147400498,
0.0158590898,
-0.0682617128,
-0.0130347135,
-0.0264609382,
-0.0229006186,
-0.0249083806,
-0.0027576732,
-0.0353507735,
0.0378064811,
-0.0013069402,
0.0582884885,
0.0074249706,
0.0064366525,
0.0212868303,
0.002941244,
-0.1085181832,
0.0111690434,
-0.0321880616,
0.029740205,
0.0400028229,
0.0322906896,
-0.009919256,
-0.0040077665,
-0.0346285291,
0.0087921452,
-0.017774161,
-0.0074090953,
0.0146385869,
-0.0210826993,
-0.0008008888,
-0.0366903245,
-0.0371679217,
0.0205103252,
0.0631081983,
-0.0697235316,
-0.0245055743,
-0.0713561699,
0.0350072607,
-0.0161487907,
0.0237393379,
0.0016397296,
-0.0475369655,
-0.0076752901,
-0.0361885428,
-0.0013589382,
0.0696422234,
0.0261945948,
0.0099294744,
0.0820461363,
0.0342769027,
-0.0348133743,
0.0209936574,
-0.042918738,
-0.0219565481,
-0.0065103248,
-0.0569631681,
0.0088616312,
0.0582306609,
0.0092403749,
-0.0111553706,
-0.0303291045,
-0.0091668079,
-0.0183328167,
-0.0122358548,
-0.0749717578,
-0.003990835,
-0.0203072429,
0.0468022339,
0.0158777833,
-0.0604544207,
0.0187574178,
-0.0437701531,
0.0094622765,
-0.0186769608,
-0.0255354084,
-0.0290340632,
0.0354708284,
0.0274391528,
0.0545247942,
-0.0420411341,
-0.0007016564,
-0.0192213804,
-0.0016323697,
0.0214419737,
-0.0372578464,
-0.0262480937,
-0.0235706288,
0.0001953457,
-0.0137429507,
-0.0242828019,
0.0263367016,
-0.0343185402,
0.0188269038,
-0.0083759753,
0.0407617763,
-0.0494472682,
0.0399645418,
0.0345497541,
-0.0086334916,
0.0401493944,
-0.0108712027,
-0.0005152829,
-0.007674058,
0.019487815,
0.0278833639,
0.0469487645,
0.0288674198,
0.0019973149,
0.0066916249,
0.0507338569,
0.0260986947,
0.0342380702,
-0.0550474748,
0.0093693379,
0.0608845763,
0.0121488981,
-0.0253527276,
0.0135413026,
0.045755174,
-0.0372900739,
0.0252916161,
0.0805530995,
0.0321816765,
-0.0988888815,
-0.0020452531,
0.0220640805,
0.0757559389,
-0.0022401926,
0.031864658,
-0.0746410713,
-0.0117278695,
0.000737318,
0.0342985168,
0.0137212323,
-0.0140261147,
-0.0130231678,
-0.0729450658,
0.0415670685,
0.0451931618,
0.0153574925,
-0.0176159088,
-0.0365632959,
-0.0248596929,
-0.0150979506,
0.0086676097,
-0.0054393956,
-0.0054232241,
0.0122154383,
0.0237785634,
-0.0529433973,
-0.044345893,
-0.0109231602,
-0.0015425043,
0.0139178559,
0.0182210989,
0.0463612303,
-0.0003098467,
0.0116542978,
-0.036586225,
0.0370829105,
-0.0205204301,
-0.0331758708,
0.0009820939,
-0.0020179516,
0.0123435911,
0.014443418,
0.047995355,
-0.0385911316,
0.0163596403,
-0.0119177559,
0.0826030746,
0.006651538,
-0.0008303325,
-0.0059845108,
-0.0388277769,
0.0028405534,
0.0374625586,
0.0235531162,
-0.0201960783,
-0.0365865938,
-0.0180833209,
0.0022426839,
-0.0152789345,
-0.0173149705,
-0.0382052138,
-0.0370143205,
-0.0355616845,
-0.0676178187,
0.0445103534,
-0.0090672579,
0.0269457102,
-0.0445556827,
-0.0340858884,
0.0051474548,
-0.0115035828,
-0.1115129068,
-0.005857422,
0.0220447574,
-0.00030047,
-0.0010949632,
0.0101958914,
-0.0012035178,
0.0115924245,
-0.0181175396,
0.0123671181,
-0.1268211901,
-0.0644279122,
0.0091104433,
0.025532946,
0.0162790995,
0.0673170984,
0.0075850752,
-0.05394179,
0.0583398491,
0.0712238401,
-0.0610012673,
0.0425293483,
0.0398975536,
-0.0218228176,
0.0060238866,
0.0341411456,
-0.005060141,
0.0110785486,
-0.0245856643,
0.0134318322,
-0.0617785826,
-0.0429421812,
0.0044346903,
0.0126464907,
0.0529714674,
-0.0055067926,
-0.055084981,
0.0026858363,
0.0312801674,
-0.0088402471,
-0.0206995569,
-0.025918629,
0.0056798942,
0.0209274106,
-0.0080692731,
-0.0274776425,
0.0151270088,
0.0699936077,
-0.0574811175,
0.0224589463,
-0.0191124715,
-0.0539357513,
0.0168198626,
0.0464075021,
0.0069685029,
-0.0319680609,
-0.0070952317,
-0.0122018615,
0.0127311787,
-0.0436563157,
0.0092071379,
-0.0333114155,
0.027355276,
-0.0473842584,
0.0770420879,
-0.0257081278,
-0.0723493099,
-0.063425079,
-0.0100416504,
0.0311041754,
0.020827068,
0.054122597,
-0.0389680415,
-0.0364847258,
-0.0016970426,
0.0165688638,
-0.0168077648,
0.0169449728,
0.028786689,
0.0022614303,
-0.0378193893,
0.0371871851,
0.0617313646,
0.0213993751,
0.0047226618,
-0.0384710692,
-0.0371555649,
-0.027841812,
-0.0124099823,
-0.073674351,
0.0325181261,
0.0164845381,
-0.0061404933,
-0.0065751355,
-0.0514648445,
0.0118409609,
0.0084492397,
-0.0034720467,
-0.0381582677,
0.0416250676,
0.0518568717,
-0.0078506917,
0.0047618523,
-0.0712849125,
-0.0296297595,
-0.0325187109,
-0.0225138571,
-0.0193275306,
-6.795857572e-33,
-0.0164830759,
0.0244303737,
-0.007052605,
-0.0101112248,
-0.0647621229,
0.0121152904,
0.0114849396,
0.0197521541,
-0.0082858745,
0.0027426744,
-0.0236615296,
0.0256541856,
0.0084698005,
0.0088195223,
-0.0119375903,
0.0309848078,
0.0421756729,
-0.0226161405,
-0.0145756947,
0.06247833,
0.038149938,
0.0422874428,
0.0362195745,
0.0236642845,
-0.0283137523,
-0.0062037506,
-0.0094560431,
-0.0064907949,
0.0281489156,
0.0107823191,
-0.0718160197,
0.025431795,
-0.0090492014,
-0.0219099279,
-0.0127610955,
-0.0075101997,
-0.0374383666,
0.0180645511,
-0.0059235906,
0.0165378097,
-0.0264920592,
-0.0599615984,
0.0390338078,
0.0244787466,
-0.052495271,
-0.0318841971,
-0.0035805551,
-0.0060267784,
-0.0677293539,
-0.0676465854,
0.0061019915,
0.0307712313,
-0.0142150782,
0.0151665229,
0.0497168452,
0.0110609746,
-0.0413649529,
0.0013027028,
-0.0803755298,
0.0080835372,
-0.0021404529,
0.0885193869,
0.0477243438,
-0.010762942,
0.0243436303,
0.0438358858,
0.0195126534,
-0.0143888034,
-0.0474570729,
0.0912887082,
0.0145475622,
0.0322825573,
0.0091815256,
0.0472589321,
0.0535919555,
-0.0287738908,
-0.0521328896,
-0.0123961968,
-0.0438663736,
0.0384164676,
0.0146423923,
0.0238333251,
-0.0070817973,
-0.0536450148,
-0.0469253808,
-0.0340342857,
-0.0192688201,
-0.0216125902,
-0.0140106976,
-0.0138412043,
-0.0538907722,
0.000453323,
0.0130800549,
-0.0384455062,
0.0172834601,
-0.0140115246,
-0.0188863333,
0.0370677337,
0.0374739207,
0.0306087714,
0.0023914773,
-0.0144043965,
-0.0478684418,
-0.0154059287,
0.0269709248,
-0.0139692798,
0.0247785207,
0.0424933098,
-0.0614886619,
-0.0070313015,
-0.0119856959,
-0.004906096,
-0.0150461523,
0.0348130763,
0.019317152,
0.0135689192,
0.0233599134,
0.0828425512,
-0.0055766702,
0.061939925,
-0.0105040316,
-0.030601358,
0.0138284797,
0.0452242643,
-0.0209347941,
0.0063519822,
-0.0472787097,
-0.0256300829,
0.0477600545,
-0.061435245,
-0.042291142,
0.0115926443,
2.954e-7,
0.0180796422,
0.0755411237,
-0.0159897432,
-0.0222692173,
0.0186199248,
0.0162412692,
0.0385013893,
0.0328941308,
-0.001688482,
-0.001016442,
0.0129467724,
0.0266172271,
0.0604118146,
0.0104311956,
-0.071417056,
0.0340703428,
-0.0737215057,
0.0063791987,
-0.0635945573,
-0.0085055875,
0.0230415054,
0.1269716918,
-0.0118294926,
-0.024036102,
0.0137555888,
0.0129172001,
0.0556117631,
-0.0354504213,
0.0168892704,
0.0282436106,
0.0651539639,
-0.0016438027,
-0.0357131809,
-0.0495470613,
-0.01495953,
0.0039966172,
-0.0309339035,
0.0938817635,
0.0137958257,
0.0141799841,
0.0309751127,
-0.0109750936,
-0.0127057312,
-0.0517456867,
0.0605384223,
0.0266284291,
-0.002833924,
-0.1073276177,
0.011327601,
-0.0045732432,
0.0454510376,
-0.0476897843,
0.0448672213,
0.0357312337,
0.010739916,
-0.0181341469,
-0.005400856,
-0.0888948664,
-0.0014225723,
0.0640598461,
-0.0273498129,
-0.027653629,
-0.0209954996,
0.006106677,
0.0675815493,
-0.0404520109,
0.0331810527,
2.828106247e-34,
0.0158025064,
0.0037456097,
0.03863547,
0.0299462788,
0.0039230869,
-0.0287981797,
0.0322788171,
0.0055109574,
-0.0016520916,
-0.0987999365,
-0.0145542463
] |
http://arxiv.org/abs/2406.00023v1
|
http://arxiv.org/pdf/2406.00023v1
|
LocMoE+: Enhanced Router with Token Feature Awareness for Efficient LLM Pre-Training
| 2024-05-24T02:50:44 | 2024-05-24T02:50:44 |
Jing Li, Zhijie Sun, Dachao Lin, Xuan He, Yi Lin, Binfan Zheng, Li Zeng, Rongqian Zhao, Xin Chen
|
cs.CL
|
cs.CL
|
Mixture-of-Experts (MoE) architectures have recently gained increasing
popularity within the domain of large language models (LLMs) due to their
ability to significantly reduce training and inference overhead. However, MoE
architectures face challenges, such as significant disparities in the number of
tokens assigned to each expert and a tendency toward homogenization among
experts, which adversely affects the model's semantic generation capabilities.
In this paper, we introduce LocMoE+, a refined version of the low-overhead
LocMoE, incorporating the following enhancements: (1) Quantification and
definition of the affinity between experts and tokens. (2) Implementation of a
global-level adaptive routing strategy to rearrange tokens based on their
affinity scores. (3) Reestimation of the lower bound for expert capacity, which
has been shown to progressively decrease as the token feature distribution
evolves. Experimental results demonstrate that, without compromising model
convergence or efficacy, the number of tokens each expert processes can be
reduced by over 60%. Combined with communication optimizations, this leads to
an average improvement in training efficiency ranging from 5.4% to 46.6%. After
fine-tuning, LocMoE+ exhibits a performance improvement of 9.7% to 14.1% across
the GDAD, C-Eval, and TeleQnA datasets.
|
LocMoE+: Enhanced Router with Token Feature Awareness for Efficient LLM Pre-Training: Mixture-of-Experts (MoE) architectures have recently gained increasing
popularity within the domain of large language models (LLMs) due to their
ability to significantly reduce training and inference overhead. However, MoE
architectures face challenges, such as significant disparities in the number of
tokens assigned to each expert and a tendency toward homogenization among
experts, which adversely affects the model's semantic generation capabilities.
In this paper, we introduce LocMoE+, a refined version of the low-overhead
LocMoE, incorporating the following enhancements: (1) Quantification and
definition of the affinity between experts and tokens. (2) Implementation of a
global-level adaptive routing strategy to rearrange tokens based on their
affinity scores. (3) Reestimation of the lower bound for expert capacity, which
has been shown to progressively decrease as the token feature distribution
evolves. Experimental results demonstrate that, without compromising model
convergence or efficacy, the number of tokens each expert processes can be
reduced by over 60%. Combined with communication optimizations, this leads to
an average improvement in training efficiency ranging from 5.4% to 46.6%. After
fine-tuning, LocMoE+ exhibits a performance improvement of 9.7% to 14.1% across
the GDAD, C-Eval, and TeleQnA datasets.
|
[
0.0389860459,
0.0311604105,
0.0060117082,
0.0256406255,
-0.0311512481,
-0.017010767,
0.0441929549,
-0.037246529,
-0.0125794364,
-0.0653243363,
-0.0642072409,
-0.0158174597,
-0.0152586792,
0.0254345573,
0.0334535874,
0.0002672811,
0.0326012298,
-0.0429345928,
0.0109168021,
-0.0210815035,
-0.015167499,
0.0371050872,
-0.0082330471,
0.0585367158,
-0.0351699553,
0.0307442173,
-0.0166027788,
0.0213054903,
0.0081328275,
-0.0233584251,
0.0097727021,
0.0522368476,
0.0281192716,
0.0199813657,
0.000002203,
-0.0490953997,
-0.0301681869,
0.0049437187,
-0.0211381093,
-0.0071744313,
0.0342507698,
0.0235561952,
0.00481252,
0.0505000986,
-0.0015003389,
-0.0205628425,
0.0497928075,
0.0430058837,
0.0325706489,
0.1155059114,
-0.0141117871,
-0.0410573669,
0.0500327498,
-0.021636581,
-0.0063284859,
-0.0409838147,
0.025397025,
-0.0012013627,
-0.0323106684,
-0.0129247019,
-0.0111113517,
0.0244647171,
0.0192127004,
0.0419013202,
0.0215648133,
0.029172821,
-0.0028340667,
-0.0276587196,
-0.0356112421,
0.0679461509,
-0.0409174301,
-0.0198774505,
0.0149937999,
0.0152712855,
-0.0270178951,
0.0063523431,
-0.0145439347,
0.0081792446,
0.0115716178,
0.0208424944,
0.0274499785,
-0.0176871587,
0.03871499,
0.0298967399,
-0.0916170254,
0.0556032509,
0.0338775404,
0.0084654121,
0.0113251088,
-0.0128036328,
-0.0305471867,
-0.0602657162,
0.0191146955,
0.0231574737,
-0.0329628214,
0.0224161968,
-0.0535885468,
-0.0182371866,
-0.0023295814,
-0.0217011385,
-0.0055110739,
0.047564216,
-0.0179944523,
0.041261591,
-0.0709519237,
0.0095868176,
-0.0188382287,
0.0434557982,
-0.0559378006,
-0.0379249118,
-0.0359018408,
-0.0134263756,
-0.0517735705,
0.052857291,
-0.0302165337,
-0.0473985374,
-0.0447585098,
0.0347932726,
0.010610735,
-0.0121093821,
0.0257165544,
-0.0122814598,
-0.0019648124,
-0.000663512,
-0.0280574113,
0.0037277071,
-0.0385993756,
0.0282468349,
0.0080914125,
-0.0657232478,
0.0112240221,
0.0603620671,
0.0037084573,
-0.0079229688,
0.0249853488,
0.0948898196,
-0.0298305321,
-0.0548004285,
-0.0467994511,
-0.029122822,
0.0133333933,
-0.072726652,
0.0124016758,
-0.0821145922,
0.0647546798,
0.0246152095,
0.0079757068,
-0.0469490588,
-0.0051626307,
-0.0017768402,
-0.0385639518,
0.0173725393,
-0.0401886888,
0.0216225162,
0.035073556,
-0.0014987956,
-0.00766727,
0.083044298,
0.0063815215,
0.0062745018,
0.0132016148,
0.0064649647,
-0.0451323427,
0.0259931348,
0.0296403319,
-0.0136492131,
-0.0228034966,
-0.0023146966,
-0.0099277357,
0.0493005142,
0.0481546111,
0.0573669747,
-0.0527444109,
-0.0057186866,
0.0118896849,
0.0609237812,
0.0207757466,
0.0391478315,
0.0589974374,
0.0245335139,
-0.003161845,
-0.019429544,
-0.0393100232,
0.006214025,
-0.0351200067,
-0.0144406678,
0.0730204806,
0.0539064854,
-0.039184548,
-0.033941824,
-0.0081240488,
-0.0225708056,
-0.0535201356,
0.0189743116,
0.0493310057,
-0.0142747657,
-0.0221110806,
0.0799776092,
0.0484559014,
-0.0366363041,
-0.0048113922,
-0.0793343931,
0.0206069574,
0.0756696463,
0.0310371406,
0.0014160644,
-0.0590824857,
0.0136871515,
-0.0004810204,
-0.0014954006,
0.0032359103,
0.0106112314,
-0.0085088788,
-0.0769425631,
-0.0525124706,
-0.0231766999,
-0.0019629041,
0.0332172997,
0.0367540978,
0.0194856059,
0.0379113853,
-0.0322557464,
-0.0068328301,
0.0328415819,
-0.0179305729,
0.0280679036,
0.0263818484,
0.0041276505,
0.0357349142,
0.0248767938,
-0.0274970308,
-0.0026737554,
0.023610428,
-0.0388097987,
-0.0152004883,
0.0062897685,
-0.0743904784,
-0.0284736641,
0.0316072442,
-0.0116764214,
0.0058076079,
-0.029826723,
-0.0384546816,
-0.0374690928,
0.0041812221,
-0.0267679766,
0.0686155334,
-0.0631037503,
0.0135203162,
0.0219330546,
0.0157960691,
-0.0053518005,
0.058593072,
-0.022010535,
-0.0070943628,
0.0321642421,
0.0482465848,
-0.0036100408,
-0.026567271,
-0.006971831,
-0.0390205234,
0.0251863301,
0.0157747213,
-0.0101072611,
0.0313730389,
-0.0294106957,
-0.0631730929,
-0.1234169528,
0.0196720287,
-0.0559296459,
0.0012908308,
0.0233751293,
-0.0060057314,
-0.0053859646,
0.0140797701,
-0.0518075004,
0.0012099385,
-0.0227442998,
0.0105550438,
-0.0282242689,
-0.0012137892,
-0.0586326048,
-0.0109519819,
-0.0267550759,
0.0080451509,
0.0572956949,
-0.0197317693,
-0.0141391903,
-0.032052923,
0.0004615353,
-0.0627258867,
0.0142049668,
-0.0054893652,
-0.0171816908,
-0.0226316601,
-0.0295541212,
0.0642957166,
-0.005037358,
0.0076241717,
0.0225457083,
0.0622593798,
-0.0123528382,
-0.046795208,
0.0099885128,
-0.015065209,
0.0068477248,
-0.0613617748,
0.0065595708,
-0.0047560139,
0.0350987352,
-0.0113503812,
-0.0175501108,
-0.0364474393,
-0.022942286,
0.0123421522,
-0.0740846545,
-0.0148412725,
0.0024360679,
0.0019902799,
0.007778096,
-0.0050610043,
-0.0196321234,
0.0412606746,
-0.0391318165,
-0.0012761138,
-0.0393421277,
-0.0342597179,
-0.0336968526,
0.0242647603,
0.0104890484,
0.0618469007,
-0.0569792129,
-0.0252188779,
-0.0397754386,
0.0061286814,
-0.0459751487,
-0.0095925061,
0.0466453247,
-0.0396195278,
0.0329521336,
0.0174174439,
-0.008289244,
-0.0267492048,
0.0008956393,
-0.0263693463,
0.0413615815,
0.0279192477,
-0.0433132723,
-0.0330466218,
0.0277212467,
-0.0002122976,
0.0468576998,
-0.0078074411,
0.0130713806,
-0.014140903,
0.0324666873,
0.0513997674,
0.0228355806,
0.0034274263,
-0.0252406448,
-0.0016652253,
0.0394944884,
0.0400846377,
-0.0589817651,
-0.0290340222,
0.0116863344,
0.0220548231,
-0.0296570845,
-0.0083968043,
0.0291222427,
0.0482051782,
-0.0403054841,
0.0097319633,
0.1341848522,
0.0217161663,
-0.0559444763,
-0.0132111283,
0.0003524914,
0.0434460454,
0.0184517149,
0.0223934148,
-0.0685767531,
-0.0404346026,
0.0110603413,
0.0029382631,
-0.0176204387,
-0.0129865557,
-0.0275714006,
-0.0721704066,
0.0246774927,
0.0074580414,
-0.0754421949,
-0.0052452325,
-0.0508099347,
-0.0231299661,
0.0042253966,
0.0757053122,
-0.0037399309,
0.0109965885,
-0.0317121297,
0.0313288979,
-0.0260364581,
-0.0550214425,
0.008696802,
-0.0280833934,
0.0672788098,
0.0262312107,
-0.0180735104,
0.0210559499,
-0.0253851283,
0.0615276285,
0.0154019026,
0.0357838646,
-0.0283155516,
0.0146661093,
-0.060613703,
0.0042002602,
0.0002788369,
0.0142608462,
-0.0012418847,
0.0110475514,
-0.0152731501,
0.0844948515,
-0.0208761506,
0.0068275267,
-0.0085798558,
-0.012143123,
-0.0298345052,
0.0714553818,
0.0202489961,
-0.0280215349,
-0.0307126716,
0.0036606048,
-0.0263987519,
-0.0001498696,
-0.0203178171,
0.0200292449,
-0.0303430893,
-0.0758932382,
-0.0357292853,
0.0012891603,
0.0590910874,
0.0676833317,
0.0021335736,
0.0227945391,
-0.0312986709,
0.0074446048,
-0.0504942425,
0.0425567664,
-0.0447551571,
-0.010822786,
-0.1182028726,
-0.0348126292,
0.0561515093,
-0.012275679,
-0.0161338523,
0.0380623005,
-0.0662222803,
-0.0747932047,
-0.0260592792,
0.0480234809,
0.0169744529,
0.0527674295,
0.0120351268,
-0.0551429391,
0.1062337682,
0.0449539535,
-0.0689123645,
0.024805434,
0.0111052999,
0.009262383,
0.0009243477,
0.0402696542,
-0.0610821173,
0.0046826759,
0.0261313189,
0.0513283275,
-0.001327068,
-0.0155042093,
-0.0093018599,
0.0214828979,
0.0270895921,
-0.0340739749,
0.0335133299,
-0.0363846421,
0.062752068,
-0.0221021622,
0.0089975558,
-0.032912612,
0.0247355849,
0.0164948106,
0.0182699598,
0.0137142455,
0.0034043856,
0.0225741453,
0.0121502178,
0.0235990193,
0.0225316603,
-0.0429119132,
0.0012856387,
0.0391639471,
-0.0062887138,
-0.0340289213,
-0.027267551,
0.0221460089,
-0.0166922174,
-0.0631638393,
0.0415008888,
0.0052468786,
0.0031019596,
-0.0327904113,
0.0422428697,
-0.0224178992,
-0.0761805624,
-0.0230477508,
0.0854564235,
0.0757883042,
0.0210545212,
0.0500439927,
-0.0283765346,
0.0042701764,
-0.0336800963,
0.0217147395,
-0.0155515447,
-0.0174793545,
0.0456386246,
-0.0030114197,
-0.0052580116,
-0.0059460071,
0.0639032498,
0.0183335692,
0.0335634463,
-0.0039492953,
-0.0020183118,
-0.0354363509,
-0.0202608854,
0.0478590876,
0.0081750657,
-0.0324880257,
0.0191990975,
-0.0246227179,
-0.0459918827,
-0.0023953361,
-0.0392006747,
-0.0199690703,
0.0193204023,
0.0767427161,
0.0672970936,
-0.0290920492,
-0.0240555741,
-0.01249013,
-0.0432412028,
0.0142203709,
-0.0182066709,
0.0400248766,
-7.323539925e-33,
-0.0134067135,
-0.017034281,
0.0522794127,
-0.0436792262,
-0.0147298947,
0.0006353994,
-0.0169518255,
-0.0044936743,
-0.0002875653,
-0.0200172905,
-0.0023019325,
-0.0153703997,
0.015854653,
0.0409240462,
-0.0176208243,
0.0036764462,
0.0810034126,
-0.0067791305,
0.0212995578,
0.0465153717,
0.0324502625,
0.0224685222,
0.0352895372,
-0.0369831882,
-0.0471415743,
0.0383472331,
0.0125764934,
0.0231739264,
-0.0030715941,
0.0341080464,
-0.0152413789,
0.0397193693,
0.0209380537,
0.0069506294,
-0.0234912466,
-0.0336727612,
-0.0546859354,
-0.0562344827,
0.0228003692,
0.0106684631,
-0.0091810431,
-0.0508570522,
0.0547673143,
0.0301704649,
-0.0189451333,
0.0071913549,
0.0234441534,
0.0220666882,
-0.0294854101,
-0.0281230211,
-0.0268708915,
0.0357032381,
0.0028410642,
0.0191116817,
0.0281996652,
-0.0210032035,
-0.0049134586,
-0.0344693214,
-0.0382958613,
0.0052035935,
0.0150895622,
0.0460660048,
0.0622866787,
0.0544334613,
0.0157194678,
0.057297539,
0.0610143915,
-0.0248799641,
-0.0069966032,
0.014429274,
0.0504198596,
0.0508892275,
-0.0176003464,
-0.0211943891,
0.0047143246,
-0.0100248763,
-0.076792255,
0.002335177,
0.0107531073,
0.0358923785,
0.058924444,
0.0064758142,
-0.0112998653,
-0.0124421595,
-0.0213943515,
0.0104114972,
-0.0417094156,
-0.0542761981,
0.0331167355,
0.0391045511,
-0.0079567209,
-0.0304657426,
0.0422622561,
-0.0224219505,
-0.0103115747,
0.0147788096,
-0.0097526666,
0.0112123061,
-0.0079859216,
0.0445116088,
-0.1015871167,
-0.0663917735,
0.0238810256,
-0.0453158505,
-0.0150287533,
-0.0199343115,
0.0000932763,
0.0304608997,
-0.0808698237,
0.0025907755,
-0.0216702074,
0.024723066,
0.0029167682,
0.0186917745,
0.0162027143,
0.0058555324,
0.0029381807,
0.0546529852,
-0.0017066137,
0.0007374375,
0.0008521769,
0.0009434474,
0.0262016263,
0.0321292616,
-0.0499489196,
0.0363286994,
-0.0407902971,
-0.033463195,
0.0246391576,
-0.0516626574,
-0.027117528,
0.0121131977,
3.03e-7,
0.0177429263,
0.0362975486,
0.0051438459,
-0.0184938684,
-0.0087085664,
-0.0251772366,
-0.0000389667,
0.0103894807,
0.0071568866,
0.002369191,
0.0412087627,
0.0639279932,
0.0003689306,
0.0032552318,
-0.1179634929,
0.0436556935,
-0.0442675091,
0.0058503104,
-0.0584522337,
0.0131976008,
0.0467723422,
0.1039326787,
0.0049603842,
-0.019995803,
-0.0124518331,
-0.0305939522,
0.0391428918,
-0.0257971473,
-0.0061809449,
0.044049114,
0.0601864308,
0.0398486145,
0.0045699365,
0.0309936479,
0.0097142262,
0.0530483052,
-0.0230134372,
0.0729971603,
-0.0213262904,
-0.0494241752,
0.0455261357,
0.001579901,
-0.0094167618,
-0.0410149395,
0.0690881461,
-0.0324779376,
-0.0259151012,
-0.0136848046,
0.0056034951,
-0.0096215503,
-0.0163965542,
-0.0292202868,
-0.0272423197,
0.038162563,
0.0246927775,
-0.033167351,
0.0096325539,
-0.088928774,
0.0306334067,
0.011621003,
-0.0482897051,
-0.0838847682,
-0.0264266394,
0.0295178853,
0.048007369,
-0.0191829354,
-0.0021789861,
3.087284613e-34,
0.0572465546,
0.0040071183,
-0.0053688819,
0.0609806739,
0.0266262181,
-0.0374450684,
-0.0334378891,
0.0045203981,
-0.0209222659,
-0.1003661156,
-0.0263787024
] |
http://arxiv.org/abs/2405.15165v1
|
http://arxiv.org/pdf/2405.15165v1
|
A Solution-based LLM API-using Methodology for Academic Information Seeking
| 2024-05-24T02:44:14 | 2024-05-24T02:44:14 |
Yuanchun Wang, Jifan Yu, Zijun Yao, Jing Zhang, Yuyang Xie, Shangqing Tu, Yiyang Fu, Youhe Feng, Jinkai Zhang, Jingyao Zhang, Bowen Huang, Yuanyao Li, Huihui Yuan, Lei Hou, Juanzi Li, Jie Tang
|
cs.CL
|
cs.CL, cs.AI, cs.SE
|
Applying large language models (LLMs) for academic API usage shows promise in
reducing researchers' academic information seeking efforts. However, current
LLM API-using methods struggle with complex API coupling commonly encountered
in academic queries. To address this, we introduce SoAy, a solution-based LLM
API-using methodology for academic information seeking. It uses code with a
solution as the reasoning method, where a solution is a pre-constructed API
calling sequence. The addition of the solution reduces the difficulty for the
model to understand the complex relationships between APIs. Code improves the
efficiency of reasoning.
To evaluate SoAy, we introduce SoAyBench, an evaluation benchmark accompanied
by SoAyEval, built upon a cloned environment of APIs from AMiner. Experimental
results demonstrate a 34.58-75.99\% performance improvement compared to
state-of-the-art LLM API-based baselines. All datasets, codes, tuned models,
and deployed online services are publicly accessible at
https://github.com/RUCKBReasoning/SoAy.
|
A Solution-based LLM API-using Methodology for Academic Information Seeking: Applying large language models (LLMs) for academic API usage shows promise in
reducing researchers' academic information seeking efforts. However, current
LLM API-using methods struggle with complex API coupling commonly encountered
in academic queries. To address this, we introduce SoAy, a solution-based LLM
API-using methodology for academic information seeking. It uses code with a
solution as the reasoning method, where a solution is a pre-constructed API
calling sequence. The addition of the solution reduces the difficulty for the
model to understand the complex relationships between APIs. Code improves the
efficiency of reasoning.
To evaluate SoAy, we introduce SoAyBench, an evaluation benchmark accompanied
by SoAyEval, built upon a cloned environment of APIs from AMiner. Experimental
results demonstrate a 34.58-75.99\% performance improvement compared to
state-of-the-art LLM API-based baselines. All datasets, codes, tuned models,
and deployed online services are publicly accessible at
https://github.com/RUCKBReasoning/SoAy.
|
[
0.0493127517,
0.0857768878,
-0.0163318217,
0.0610214286,
-0.043053329,
-0.0269211065,
0.0112895053,
0.0183711182,
0.0480257422,
-0.0068137348,
-0.034031678,
0.0436502583,
0.0203983858,
0.0283674244,
-0.0139154457,
0.0144626014,
0.014587339,
-0.0110287005,
0.0435212813,
-0.023932999,
-0.0088038603,
0.0261092503,
-0.0216109268,
0.0044035902,
0.0684640259,
-0.0345332772,
0.0139112603,
-0.0086207408,
-0.0181601401,
-0.0156221082,
0.0210942328,
0.0301185288,
-0.024337329,
0.0318097882,
0.0000022352,
-0.0879874304,
-0.0425623618,
0.07148619,
-0.0243867822,
0.0461513363,
0.036556039,
0.0267302711,
0.0299781077,
0.0131477974,
-0.0232230779,
0.0087098489,
0.0267463066,
-0.0446894541,
0.0292735174,
0.0289204959,
0.0222958066,
-0.0599092729,
0.0195263065,
0.0116527453,
-0.0435237736,
-0.0088035101,
0.0422249325,
-0.0198657177,
0.1035531536,
0.0635866448,
0.0182607658,
0.0134601118,
-0.0271948315,
-0.0182344988,
0.0120588345,
0.0114345718,
-0.0671006516,
-0.0292472765,
0.0207874868,
0.0265878066,
0.0784837306,
-0.0104410183,
0.0313606858,
-0.0064950036,
-0.0286198966,
0.0476613678,
-0.0564786457,
0.0036282949,
-0.0011271042,
-0.0103866542,
0.0080339313,
-0.0060114581,
0.0162567198,
0.0433682278,
-0.0178177338,
0.081779927,
0.0003632719,
0.0057296944,
0.0034177005,
-0.0431805924,
0.025525419,
-0.0050173025,
0.012872925,
0.0604854897,
-0.033393167,
-0.0143786585,
0.0443647578,
-0.0491385572,
0.0667644888,
-0.0007438546,
-0.0070222756,
0.0365055613,
-0.0212855525,
0.0314982831,
0.0114390366,
0.0429505222,
-0.0180953071,
-0.037983574,
0.0084158964,
-0.0072774696,
-0.0353252739,
0.0062139761,
0.0116642453,
0.0616682358,
-0.0507420003,
0.0089404229,
-0.0159758665,
0.0223161019,
0.0283604246,
0.0367491841,
-0.0312649198,
-0.0292386636,
-0.0218261629,
0.0022378466,
-0.0426414013,
0.0275446717,
0.0164325461,
-0.0288778897,
-0.0059577208,
-0.0486719087,
0.0082512898,
-0.0242581032,
0.0183563214,
-0.0117094247,
0.0423468575,
0.0813158751,
0.0496778078,
-0.0126175201,
-0.0370849073,
-0.0704552084,
0.0778447166,
-0.0494248942,
-0.0073083588,
-0.0861631706,
0.0526461825,
0.0684565976,
-0.0070192413,
-0.032794103,
-0.0123023521,
0.0082417941,
-0.0484689474,
0.0213262998,
-0.0456791259,
0.0426978432,
0.044471547,
0.0361508317,
0.0080328304,
0.0825112686,
-0.0458346307,
-0.0172183737,
0.0584414192,
-0.0182888974,
0.0017101214,
0.0340004116,
0.0091096386,
-0.0018380523,
0.0179677755,
0.0754783005,
0.0117233032,
0.0300935786,
-0.0358534567,
0.0094476445,
-0.025711054,
-0.005152347,
0.0189386047,
0.048798833,
0.008940422,
-0.0016129097,
-0.0353020877,
0.0567621216,
0.0274815261,
-0.0620473884,
-0.0208224431,
0.0384638198,
0.0075114234,
0.0098986821,
0.0623302832,
0.064099215,
-0.0007199678,
-0.0469509102,
-0.0309554655,
0.0114550637,
-0.1137659922,
0.0352326483,
0.0673502907,
-0.0366526209,
-0.012604692,
0.0564646125,
0.034723606,
-0.0408782065,
-0.0086965207,
-0.0187553279,
0.0449586585,
0.0536339134,
0.1071308777,
-0.0772573948,
-0.0402564555,
0.0153544843,
0.0191243738,
0.0602749847,
0.0260984022,
-0.0060290475,
0.0109197143,
-0.0216168966,
-0.0287485812,
0.0010834823,
-0.0014616385,
-0.0210864358,
0.0562900342,
0.0240663067,
-0.0278583691,
-0.0596314073,
-0.0624699704,
0.0380586572,
-0.0846837685,
0.0199043304,
-0.0037765526,
-0.0204504356,
0.0184016321,
0.0332529321,
0.0119151259,
0.0162681006,
0.0484427549,
-0.0302947648,
0.0102983555,
-0.053267736,
0.0639146343,
0.0664321408,
0.0365294181,
0.009584168,
0.0197724998,
0.0249550287,
-0.0343231708,
-0.002883374,
0.01415205,
0.0294563342,
0.0487679467,
0.0376370177,
0.0215952937,
-0.0573819466,
0.0169702917,
-0.021522969,
0.011896003,
-0.0205843672,
0.0600468516,
0.0178365838,
0.0133653292,
0.0227373764,
0.0205844119,
-0.0303666126,
0.0188568048,
-0.0280677434,
0.01159155,
0.021262398,
-0.0204281062,
-0.0305860899,
-0.0027301754,
-0.0408064872,
0.0213742796,
0.0230901875,
0.0119701046,
0.0048965439,
0.0259679686,
-0.0184110589,
0.0086078783,
0.0276870392,
-0.0116575575,
-0.0490754321,
0.0061299475,
0.0043220078,
0.0127405543,
-0.0635698661,
-0.0281650964,
-0.0123624904,
-0.0053540384,
0.017767461,
0.0035733611,
-0.020357687,
-0.0678058937,
0.0161282513,
-0.0331237502,
-0.0093227616,
-0.0311169233,
-0.0095738275,
-0.000114192,
-0.0308490563,
-0.0245658103,
0.043324627,
-0.0402302556,
-0.0233130008,
0.0648738444,
-0.00731058,
-0.0640427321,
0.0016536529,
-0.0298414566,
-0.0090093622,
-0.0033712336,
-0.0436910428,
0.0163605046,
-0.0353650227,
-0.036559023,
-0.0121960379,
0.0019031146,
0.0207082089,
-0.0016002083,
0.0100674061,
0.0186558869,
0.0289809145,
0.0332098044,
0.0649586618,
0.0361409672,
-0.044936832,
0.0331634842,
-0.0458350927,
0.0035618283,
-0.0585662015,
-0.0229141694,
-0.0565022267,
0.0472584963,
-0.0262953844,
-0.006517211,
-0.0268206242,
0.0108363116,
0.0093254102,
0.0114185568,
0.0063632876,
-0.0578457676,
0.0672361851,
-0.0215703007,
0.0184570123,
-0.0109664397,
0.0028515842,
0.0142326858,
0.0333433449,
-0.0501206927,
0.0146396235,
-0.046150174,
-0.0578755401,
-0.0090217041,
-0.0042628036,
0.042166695,
0.0410276428,
-0.0141197816,
-0.0323145278,
-0.0141280387,
-0.0522672869,
0.0161917713,
-0.0115652811,
0.0062112873,
0.0546258539,
0.0513238385,
0.028437661,
0.0145705901,
-0.0128516424,
-0.02635088,
-0.0748559311,
0.0567279607,
-0.059452951,
-0.0133000026,
0.0400455818,
-0.0352862366,
-0.0685248747,
-0.0372348428,
0.0877542421,
0.0374171175,
-0.077907078,
-0.026513895,
-0.06753923,
0.0253781173,
0.0329817384,
0.0733927861,
-0.0647204071,
0.0071111135,
0.0010018407,
0.0709151849,
-0.0252173506,
-0.0110085793,
-0.0197695699,
-0.0873841792,
0.089763619,
-0.0357828438,
-0.0324487649,
-0.0221906304,
-0.0933752134,
0.0223652795,
-0.0158151742,
-0.0266938582,
0.0318971537,
-0.0028919927,
0.0022261459,
-0.0084544569,
-0.046764195,
-0.0518263914,
-0.0056896158,
-0.0122719137,
0.0210435856,
0.006590141,
0.0414300598,
-0.0379539691,
-0.0069001289,
0.0341789238,
0.0124877291,
-0.0216517393,
-0.0587165467,
0.0032254488,
0.0282781478,
-0.0098823803,
0.0030873322,
0.0216259994,
-0.0861646309,
-0.0021380384,
0.0336327776,
0.0957807079,
-0.039761506,
0.0685706362,
-0.0869609937,
-0.0065163565,
-0.0099013289,
0.0697858036,
0.0051394179,
-0.0373283848,
-0.0025777433,
-0.0000080109,
-0.0090643102,
0.0176257323,
-0.0223924816,
0.0061313272,
-0.0652906746,
0.0002362596,
-0.0181404818,
0.0236769449,
0.0004355444,
0.0021261936,
0.0215147473,
0.0873942226,
0.0313988477,
-0.0122136287,
-0.0611448921,
-0.0032910022,
-0.0125272544,
0.02881312,
-0.0517452843,
-0.0246221479,
0.0891985744,
-0.0038090509,
-0.0211130474,
0.0222174749,
-0.0095801782,
-0.0790761784,
-0.0197133832,
0.0417605303,
0.0287947785,
0.0441730917,
-0.0185618531,
-0.0492780618,
-0.0185103808,
0.093075633,
0.0038382635,
0.0050803348,
-0.017652519,
-0.0082205981,
-0.0117412824,
-0.0160324406,
-0.0750106499,
-0.0306318011,
-0.0105522564,
0.0042386251,
-0.0149829267,
-0.0086571714,
0.0111397421,
0.0766019374,
0.0288880784,
0.000549773,
-0.0256404411,
-0.0188953336,
0.0285227764,
-0.0280003957,
-0.0149970772,
-0.0149444994,
-0.0139766829,
-0.017281346,
-0.016855251,
0.0198915284,
0.0286914445,
0.0082362387,
0.0060604955,
0.0056290776,
0.0540611856,
-0.0548128039,
0.0063751461,
0.055949796,
-0.0361260101,
-0.0202230066,
-0.0340580903,
-0.0263606664,
0.0574661717,
-0.0346437246,
0.0349776782,
0.0183605794,
0.0076873009,
0.0378143601,
-0.0012736948,
-0.0386352427,
-0.1026954725,
0.0115158278,
-0.0420001708,
0.0008488111,
0.0180469118,
0.0269167479,
-0.029930912,
-0.0028225721,
-0.0282052606,
0.0376956202,
-0.004162889,
-0.001690113,
0.03033337,
0.0206461139,
-0.0108326096,
-0.0222469419,
0.0802486986,
-0.0218073875,
-0.0040575881,
-0.0074980902,
-0.0319427438,
-0.0211226344,
-0.0489251018,
0.0194806997,
0.0033226791,
-0.0330009572,
-0.042972099,
-0.0302704591,
-0.0261325389,
0.0070067924,
-0.0156253707,
-0.0359550342,
-0.032569211,
0.0005847187,
0.0653800517,
0.051615674,
-0.0234839525,
-0.0471646078,
0.0361795686,
-0.0846965089,
-0.051218152,
0.0338079967,
-6.655922317e-33,
0.0106189176,
-0.029447278,
0.0397801623,
0.0253652371,
-0.0173018482,
0.0160366111,
-0.00135154,
-0.0396776609,
-0.0576801375,
-0.0274442658,
-0.0176049881,
0.017310312,
0.0061904145,
-0.0005316477,
0.0441482738,
0.0170867927,
-0.0156295486,
0.0115021169,
-0.022246033,
0.0088097146,
0.0481241196,
0.007487549,
0.0027323691,
-0.0742411911,
0.0036673627,
-0.0663764551,
0.0426920354,
0.023660671,
-0.0114591308,
0.0639327466,
0.0043187183,
0.0230228323,
0.0027864578,
-0.0261629559,
0.0233419202,
-0.0482067391,
-0.0384761505,
-0.0359481052,
0.0441632979,
0.037996348,
0.030185746,
-0.0106240995,
0.0465498492,
-0.0026601758,
-0.0092819585,
-0.0110310502,
0.0301364567,
-0.0295895096,
-0.0420545638,
-0.0048811939,
-0.0307503659,
0.0035307491,
-0.0052666538,
0.0743157417,
0.0079613701,
0.0177093428,
0.001866932,
0.069363974,
-0.0273866374,
-0.001365562,
0.0148821147,
0.0402079523,
0.035322167,
0.0258462429,
0.0061772331,
0.0169928037,
-0.0054803221,
-0.0173920915,
0.0446933582,
-0.0060235369,
0.0264463853,
0.0429069512,
0.035552457,
0.0026663893,
-0.0355545469,
-0.0129872914,
-0.0606311262,
0.0090671387,
0.0297445469,
-0.0015864705,
0.0549377128,
0.0015633428,
-0.0012522246,
-0.0107955933,
-0.0105505195,
-0.0190672949,
-0.0491614714,
0.0047021592,
0.0023944923,
-0.0403244421,
-0.0160867441,
-0.0170013979,
0.0056859762,
-0.0603526011,
-0.0296932962,
0.0285586361,
-0.0126656024,
0.0170537736,
-0.0130664594,
0.0227112975,
-0.0807324052,
-0.0329206698,
-0.0642489791,
-0.0464703552,
0.0118640028,
0.0050698956,
-0.0032883186,
0.0082367994,
-0.0217961725,
-0.0129715353,
-0.0382861122,
0.0122485729,
-0.010130737,
-0.0367862694,
0.0121464441,
0.0427732654,
0.0273338724,
0.0522858202,
0.0512359142,
0.0437843092,
-0.0272931214,
0.0352888219,
0.0164977834,
-0.0193416681,
-0.0028735613,
-0.0217194427,
0.0003816244,
0.035993401,
-0.0019377252,
-0.0392285176,
-0.0336447172,
0.0045075058,
3.029e-7,
0.0445882976,
0.0434715673,
-0.0246276092,
0.0611927658,
0.0169407111,
-0.0129620712,
-0.0189104211,
-0.0273462236,
0.000348625,
-0.0067359349,
-0.0231534094,
-0.0313478447,
0.0475980304,
0.0165124219,
-0.0789135545,
-0.0273750443,
-0.0135409078,
-0.0725445673,
-0.0938576311,
-0.0254857838,
-0.013926358,
0.0093967337,
0.0303403325,
-0.0045032641,
-0.0316350348,
-0.0360275656,
0.0010035415,
-0.032318484,
0.0269099995,
0.0872608945,
-0.0117057795,
0.044144202,
0.0378380939,
0.0005120979,
-0.0170799643,
0.0095678,
-0.0322680883,
0.0459419154,
0.0161811505,
-0.0066095851,
0.0491716117,
-0.040561866,
-0.0020717534,
-0.037344683,
0.0353040025,
0.0183683187,
-0.0385649316,
-0.0127524445,
-0.0388108194,
-0.03415142,
-0.0214687344,
-0.0391282812,
0.0183767751,
0.045645684,
-0.0096452869,
-0.0543201566,
-0.0180266332,
-0.0342538729,
-0.012697719,
-0.0137342596,
-0.0107727861,
-0.0524854437,
-0.0158419106,
-0.0287370738,
-0.0157407522,
-0.0142771555,
0.007110652,
3.019974475e-34,
0.0392232761,
-0.0152541278,
0.0010400291,
0.0861615688,
0.0086705303,
0.0068838447,
-0.0231457315,
0.0318106674,
-0.0262444876,
-0.0394316502,
0.0027195262
] |
http://arxiv.org/abs/2405.15152v1
|
http://arxiv.org/pdf/2405.15152v1
|
Machine Unlearning in Large Language Models
| 2024-05-24T02:12:51 | 2024-05-24T02:12:51 |
Saaketh Koundinya Gundavarapu, Shreya Agarwal, Arushi Arora, Chandana Thimmalapura Jagadeeshaiah
|
cs.CL
|
cs.CL, cs.AI
|
Machine unlearning, a novel area within artificial intelligence, focuses on
addressing the challenge of selectively forgetting or reducing undesirable
knowledge or behaviors in machine learning models, particularly in the context
of large language models (LLMs). This paper introduces a methodology to align
LLMs, such as Open Pre-trained Transformer Language Models, with ethical,
privacy, and safety standards by leveraging the gradient ascent algorithm for
knowledge unlearning. Our approach aims to selectively erase or modify learned
information in LLMs, targeting harmful responses and copyrighted content. This
paper presents a dual-pronged approach to enhance the ethical and safe behavior
of large language models (LLMs) by addressing the issues of harmful responses
and copyrighted content. To mitigate harmful responses, we applied gradient
ascent on the PKU dataset, achieving a 75\% reduction in harmful responses for
Open Pre-trained Transformer Language Models (OPT1.3b and OPT2.7b)
\citet{zhang2022opt} while retaining previous knowledge using the TruthfulQA
dataset \citet{DBLP:journals/corr/abs-2109-07958}. For handling copyrighted
content, we constructed a custom dataset based on the Lord of the Rings corpus
and aligned LLMs (OPT1.3b and OPT2.7b) \citet{zhang2022opt} through LoRA:
Low-Rank Adaptation of Large Language Models
\citet{DBLP:journals/corr/abs-2106-09685} finetuning. Subsequently, we employed
gradient ascent to unlearn the Lord of the Rings content, resulting in a
remarkable reduction in the presence of copyrighted material. To maintain a
diverse knowledge base, we utilized the Book Corpus dataset. Additionally, we
propose a new evaluation technique for assessing the effectiveness of harmful
unlearning.
|
Machine Unlearning in Large Language Models: Machine unlearning, a novel area within artificial intelligence, focuses on
addressing the challenge of selectively forgetting or reducing undesirable
knowledge or behaviors in machine learning models, particularly in the context
of large language models (LLMs). This paper introduces a methodology to align
LLMs, such as Open Pre-trained Transformer Language Models, with ethical,
privacy, and safety standards by leveraging the gradient ascent algorithm for
knowledge unlearning. Our approach aims to selectively erase or modify learned
information in LLMs, targeting harmful responses and copyrighted content. This
paper presents a dual-pronged approach to enhance the ethical and safe behavior
of large language models (LLMs) by addressing the issues of harmful responses
and copyrighted content. To mitigate harmful responses, we applied gradient
ascent on the PKU dataset, achieving a 75\% reduction in harmful responses for
Open Pre-trained Transformer Language Models (OPT1.3b and OPT2.7b)
\citet{zhang2022opt} while retaining previous knowledge using the TruthfulQA
dataset \citet{DBLP:journals/corr/abs-2109-07958}. For handling copyrighted
content, we constructed a custom dataset based on the Lord of the Rings corpus
and aligned LLMs (OPT1.3b and OPT2.7b) \citet{zhang2022opt} through LoRA:
Low-Rank Adaptation of Large Language Models
\citet{DBLP:journals/corr/abs-2106-09685} finetuning. Subsequently, we employed
gradient ascent to unlearn the Lord of the Rings content, resulting in a
remarkable reduction in the presence of copyrighted material. To maintain a
diverse knowledge base, we utilized the Book Corpus dataset. Additionally, we
propose a new evaluation technique for assessing the effectiveness of harmful
unlearning.
|
[
0.0406031571,
0.1198176816,
-0.0287182089,
0.0717186779,
-0.049565129,
0.0092922878,
0.0334289409,
0.0097922226,
0.0308100861,
-0.0296274815,
-0.0476224013,
0.0308652464,
-0.0078820707,
0.0090318201,
-0.0203993525,
0.015725594,
0.0272059608,
-0.0183479395,
-0.0008258436,
-0.0260411706,
0.0063265543,
0.0343101211,
0.0447647423,
0.0096725859,
0.0224766247,
-0.0531167835,
0.0238752533,
-0.027821159,
0.0365440696,
0.0132316761,
0.0431809425,
0.0242760479,
0.0175142381,
0.0907203332,
0.0000020684,
-0.056595251,
-0.0336207151,
0.0243824981,
-0.017333461,
0.0308958367,
0.0734329969,
-0.0097149294,
-0.0040415814,
-0.002986121,
-0.0273729376,
0.0015866614,
0.0447988622,
0.0459700339,
0.0677657872,
0.0784182325,
-0.0179774016,
-0.0360703245,
0.0339429639,
0.0030853362,
-0.0269172043,
-0.0373391807,
0.06609492,
0.0018891314,
-0.0439477414,
0.0093092797,
-0.0197635125,
0.0537475497,
-0.005639175,
-0.0052807662,
-0.0141651593,
-0.0253256205,
-0.0174854007,
-0.0341264904,
0.0108392453,
0.0540128276,
-0.0238262601,
-0.0612313822,
0.0364337899,
-0.0008565465,
0.0103191212,
0.0222932976,
-0.0104360757,
0.0577158295,
-0.0073523517,
0.0281833988,
-0.003221184,
-0.0402107574,
0.0621002205,
0.0373487696,
0.0232755113,
0.0360078998,
0.0135536706,
-0.0265216753,
0.0167725235,
-0.0390974842,
0.0005559276,
-0.0176107828,
0.0378239714,
0.0112252468,
0.0134388674,
0.0262982119,
-0.032980945,
-0.0462853983,
0.0189510453,
-0.0578807481,
0.0478568785,
0.0015173309,
0.0103345104,
0.0529851504,
0.0013350223,
0.0179187898,
-0.0042441734,
0.0533863418,
-0.0196671169,
0.0113294246,
0.0310970135,
-0.0074722259,
0.0051778103,
0.0455392785,
-0.0262235887,
0.033763133,
-0.0291081145,
0.0076618367,
-0.0326274969,
-0.0052583599,
-0.0030744802,
0.0207912941,
-0.0270995814,
-0.0093521737,
-0.0328483656,
-0.0148192178,
0.0303569995,
0.038131129,
-0.027267009,
-0.0263758451,
0.0042857858,
-0.0006426419,
0.0003511503,
-0.0228281897,
0.0476721227,
0.1032250524,
0.0163446069,
-0.0077468869,
-0.0319062956,
-0.0300859809,
-0.0038638397,
-0.0473518781,
0.0040347031,
-0.0317927413,
0.0146642467,
-0.0458926074,
-0.0234113801,
-0.0852958187,
-0.0037041446,
0.0382789336,
-0.0970870033,
0.04064494,
-0.0175894611,
-0.0045020557,
0.0323159881,
0.0185155384,
-0.0291272663,
0.0152103566,
0.0223709121,
0.0406759195,
0.0583773218,
-0.021111479,
0.0199868847,
0.0537636839,
0.0271277372,
-0.0162806455,
-0.010261436,
-0.0154671585,
-0.0085156374,
-0.0097535392,
-0.0019579448,
0.0674609691,
-0.0120298089,
-0.0542427599,
0.0710431635,
0.0358525403,
-0.016240349,
0.0524826869,
-0.0135455411,
0.0095721092,
0.0113754543,
0.0013725295,
-0.0541177839,
-0.0271381326,
-0.0432867073,
-0.0186821446,
-0.0398320071,
0.0572732352,
0.0216525216,
-0.0208070762,
0.0057443688,
-0.0355941728,
0.043947719,
0.0243782327,
0.0393497944,
0.0484495126,
-0.0360218957,
0.0299991928,
0.0564259961,
-0.0660859346,
0.0322064012,
-0.0263658538,
0.0493192859,
0.0425804593,
0.063900575,
-0.0479489751,
0.0261132363,
-0.0390139967,
-0.0421850458,
-0.0032223586,
-0.020571772,
-0.0279177185,
0.0321436003,
-0.0632705688,
-0.0481532477,
0.014636348,
-0.0049820039,
-0.0096806986,
-0.0082963211,
-0.015541397,
0.0096813655,
-0.0199574213,
-0.06486056,
0.0284665115,
-0.062868394,
-0.0421776511,
-0.0237157159,
0.0116843497,
0.030559307,
0.0121449698,
-0.0172042213,
-0.0311083384,
0.0552101359,
0.0092485938,
-0.0291682035,
-0.0222886894,
-0.0197888426,
0.0244212989,
0.0553103797,
-0.0595525578,
0.022387106,
0.0779850706,
-0.0287106652,
0.0193109103,
0.0076062102,
0.0116652725,
0.0389014557,
-0.042185884,
0.0180982631,
-0.0710557774,
0.0040102135,
-0.0399019904,
0.0488538556,
-0.0263817478,
0.0005524036,
0.022944577,
-0.0774025619,
0.0083342474,
0.0130364047,
-0.0507367626,
0.0058033378,
-0.0168307722,
0.0482100435,
0.0482257344,
-0.0045927474,
-0.0050469036,
-0.0187044572,
-0.0186534077,
-0.0073111188,
0.0013144236,
0.0244942661,
-0.0025955008,
-0.0133743966,
-0.0260420106,
-0.0068693697,
-0.0023396823,
-0.0092497654,
-0.0158836972,
0.0243525337,
0.0070252265,
-0.0271070246,
-0.0041258247,
0.0085217394,
-0.0228674207,
-0.0041853567,
0.0544133075,
-0.0013285694,
0.0148412576,
-0.0106450189,
0.0379797667,
-0.0418352857,
0.0240012761,
-0.0151927155,
-0.022479251,
-0.0175934769,
0.0118075842,
-0.0192985479,
0.0285619795,
0.0156151727,
0.0270787813,
0.055536814,
0.0231976714,
-0.0342898816,
-0.0229363535,
-0.0175131503,
-0.0851272941,
-0.0158339832,
-0.0015977635,
0.0154397953,
0.0022174625,
0.0156367254,
-0.0353976227,
-0.017871378,
0.0447777323,
-0.046890121,
-0.0721159875,
-0.0134134227,
-0.0489087813,
-0.0376826674,
0.0198758263,
0.0073597045,
-0.0566804372,
0.0489907451,
0.0169814099,
0.0500621274,
-0.0403713435,
0.0202742014,
-0.0101894606,
0.0316056982,
0.0146613969,
0.0691702813,
0.0378708206,
-0.0333662145,
-0.0361699611,
0.018647613,
0.0786893815,
-0.0693932176,
0.0247840192,
-0.0153829576,
-0.0126283895,
-0.0098092919,
-0.0392229185,
-0.0382949971,
-0.0540313087,
0.0325349271,
-0.0222313218,
0.0023716034,
0.0036803875,
-0.0291037615,
-0.0359143838,
0.0331542604,
-0.0124287065,
0.0187012181,
0.0232714731,
-0.0002592932,
0.020051932,
0.0520512201,
-0.0141374217,
0.0148087293,
0.0054922528,
-0.0002407188,
0.076945506,
-0.013703783,
-0.0018811212,
0.0329763889,
0.0324972421,
0.0352363624,
-0.0203217641,
0.0250586364,
0.0190037806,
-0.00612651,
-0.0373824798,
0.0150759835,
0.0703804344,
-0.0008564691,
-0.055478625,
-0.0050311601,
0.0083799269,
-0.0291834082,
-0.0034124474,
-0.0117920497,
-0.1327683628,
0.0189330913,
0.000625954,
0.031437885,
-0.0269952528,
-0.0573250987,
0.0179919749,
-0.1052063778,
0.0170938428,
0.0288770851,
-0.0252939276,
-0.026746748,
-0.0211778153,
-0.0137772225,
-0.0368399024,
0.026016036,
-0.022154931,
-0.0173346959,
-0.0208021626,
-0.0035916422,
0.0181348007,
-0.0433458164,
0.0092970822,
-0.0382742733,
0.032465402,
0.0510307588,
0.0357594788,
-0.0634826124,
-0.0141823478,
0.0043243971,
-0.002480471,
-0.0296970587,
0.0058554816,
0.0243166946,
0.0059632338,
-0.0149204992,
-0.0341868773,
0.0505721383,
0.0117357261,
-0.0554650463,
0.0012579652,
0.0880122557,
0.031692192,
-0.053918954,
-0.0331681408,
-0.0215250924,
0.0073231654,
0.0787871927,
-0.0269595142,
-0.0453629084,
0.0104400888,
-0.0272794422,
-0.0126417242,
0.0216770824,
-0.0400241055,
0.0132047944,
-0.0062740371,
0.0182072762,
-0.0903519019,
0.0104334252,
-0.037031047,
-0.0197800584,
0.0688143745,
0.0135880755,
0.0656514913,
-0.0077063274,
-0.0506785773,
-0.0092823356,
-0.0369192064,
-0.0247166492,
-0.0380898304,
0.0127717452,
0.0161612388,
-0.0020152917,
-0.0105779348,
0.0029503377,
-0.030322019,
-0.0796261206,
0.0572122484,
0.0714164823,
0.0188208092,
0.0288205836,
0.0051360247,
-0.0533935055,
0.116927959,
0.0875079259,
-0.0491525456,
0.0249212105,
0.0346922241,
-0.0400237739,
-0.0351864509,
0.0408858322,
-0.046098277,
-0.0016496486,
0.0465358011,
-0.0254678354,
-0.0176874604,
-0.0293020494,
0.0144749703,
0.0324694142,
0.0070746224,
-0.0321356468,
-0.0037260107,
-0.0413548127,
-0.0001870464,
-0.010086854,
-0.0332775339,
-0.0121450508,
0.0208520927,
0.0575275533,
0.0016028554,
-0.0168845393,
0.0475978814,
0.0191455893,
0.0113834441,
-0.0646792352,
0.0182629675,
0.0273034666,
0.0385389328,
0.0175356604,
0.0412243009,
-0.0232872386,
0.0370013937,
-0.0239882283,
0.0480823778,
-0.0045237499,
0.0404223092,
0.0051446846,
0.0346131884,
0.006949285,
0.0312970392,
-0.021808831,
-0.0440933928,
-0.048795037,
-0.0043890597,
0.0210637655,
-0.0187284648,
0.024842035,
-0.0083020087,
-0.0243998095,
-0.0218554381,
0.0155092925,
-0.0004658568,
-0.0289834179,
-0.0127434926,
0.0012734369,
-0.0397253372,
0.0595417656,
0.0699426681,
0.0125395842,
0.0501926504,
0.026738422,
-0.0119326971,
-0.0195811149,
-0.0507012755,
0.0143752294,
-0.0194384027,
-0.0073868702,
-0.0077624233,
-0.0266131107,
-0.0827293396,
-0.0253738612,
-0.0035821861,
0.0196080543,
-0.0407543182,
0.0118930936,
0.028162064,
-0.0006420015,
-0.0112783862,
-0.0530788265,
0.0374617949,
-0.0058080768,
-0.0507336222,
-0.0159718711,
-5.821091004e-33,
0.0363366157,
0.0145937903,
-0.0322404429,
0.0404320173,
-0.0128034716,
-0.015976496,
0.0169981811,
-0.0393722169,
-0.0371402949,
0.0164971985,
-0.0275569074,
-0.013036822,
-0.0141334673,
0.0150847221,
-0.0149163818,
-0.0387310572,
0.0296075232,
-0.0196280032,
-0.0351037234,
0.0354186371,
0.0485828631,
0.0387829505,
0.0762443617,
-0.1176210791,
0.0395120233,
-0.0189991537,
-0.0488719344,
0.0098359212,
-0.0420648232,
0.058032535,
-0.0332295001,
0.031366881,
0.0045080697,
-0.0261263456,
0.0056062201,
0.0292851198,
-0.0629073456,
0.0268511139,
0.0159118827,
0.0259628482,
0.0100509869,
-0.0620451383,
0.0748987943,
-0.0291099139,
-0.033749342,
-0.0360531621,
0.0445092246,
-0.0197489969,
-0.0322438404,
-0.0438685827,
-0.0595542528,
-0.019394597,
-0.0229289234,
-0.0102482615,
0.0440819375,
0.0195468925,
0.0120571833,
0.0263959877,
-0.0856912509,
0.0092705525,
0.0276249107,
-0.0256750919,
0.0562622286,
0.0116507607,
0.0289062057,
0.0776490048,
-0.0012790321,
-0.0028625939,
0.0105091464,
-0.0104319882,
0.0090683121,
0.0314321555,
0.0307924729,
0.0796209127,
0.0372131094,
-0.041746337,
-0.0577970184,
0.0083420258,
-0.0424035266,
0.0362237878,
0.0156643819,
-0.0014123307,
-0.0098043755,
-0.018770745,
-0.0543898642,
-0.0323180296,
-0.027748229,
0.0054798708,
0.0074122525,
-0.02472266,
-0.0543391034,
-0.0008154413,
0.0100676268,
-0.1023731008,
-0.0208809488,
0.0372212939,
0.0147202406,
-0.0118243415,
0.0098274779,
0.0076221474,
-0.0512115248,
0.0063486234,
0.0040684314,
-0.047793407,
0.052996859,
0.0099608013,
0.0398795158,
0.0754915848,
-0.0364084803,
0.0089869127,
-0.0056962799,
0.0175915807,
0.0518398322,
0.0529180579,
0.0126936734,
0.0144197959,
0.0127334027,
-0.0019836775,
0.0579930171,
0.0570687726,
-0.0593935363,
0.0053336523,
0.0035376388,
-0.0219424609,
-0.0138372537,
-0.0270381588,
0.0119626662,
0.0360525809,
0.0692089424,
-0.0838373154,
-0.004236605,
0.0048510055,
2.843e-7,
0.0202707723,
0.003708869,
0.0112169599,
-0.0092408629,
-0.0051031108,
0.0585723706,
0.0080242371,
0.070693858,
0.0060151075,
-0.0051726056,
0.0043840632,
-0.0215832628,
0.0435398296,
0.0122199748,
-0.0873582661,
-0.0402465463,
-0.0663037822,
-0.0291282516,
-0.0472214632,
-0.0055363402,
0.008254475,
0.0856125429,
0.0309009366,
-0.0624522306,
-0.009356563,
-0.0179727394,
0.0359248258,
-0.0541882962,
-0.0188231859,
0.0113434261,
0.0451905392,
0.0367695764,
0.004589343,
-0.0366780013,
-0.0222571641,
0.0248973966,
-0.000965244,
0.0703161433,
0.012517401,
-0.0199299846,
-0.0198327601,
-0.0239235144,
0.0077019893,
-0.035842102,
0.0648586974,
0.0311952382,
-0.036713995,
-0.0468644686,
-0.0258267149,
-0.0406524986,
0.0275969133,
-0.0354128219,
-0.0128284674,
0.022635771,
0.0176889934,
0.0476646684,
-0.0089820437,
-0.0812936947,
-0.000481284,
-0.0140577741,
-0.0286037102,
-0.0408543237,
0.0249914508,
-0.0074585634,
-0.0228376202,
-0.0755846277,
0.0750782192,
2.544906469e-34,
-0.0046721026,
-0.0622691512,
0.0005955876,
-0.0092993053,
0.0199933425,
0.0108836843,
0.0300819259,
0.0186779369,
-0.0208306313,
-0.0566957407,
0.0050105103
] |
http://arxiv.org/abs/2405.15145v1
|
http://arxiv.org/pdf/2405.15145v1
|
CulturePark: Boosting Cross-cultural Understanding in Large Language Models
| 2024-05-24T01:49:02 | 2024-05-24T01:49:02 |
Cheng Li, Damien Teney, Linyi Yang, Qingsong Wen, Xing Xie, Jindong Wang
|
cs.AI
|
cs.AI, cs.CL, cs.MA
|
Cultural bias is pervasive in many large language models (LLMs), largely due
to the deficiency of data representative of different cultures. Typically,
cultural datasets and benchmarks are constructed either by extracting subsets
of existing datasets or by aggregating from platforms such as Wikipedia and
social media. However, these approaches are highly dependent on real-world data
and human annotations, making them costly and difficult to scale. Inspired by
cognitive theories on social communication, this paper introduces CulturePark,
an LLM-powered multi-agent communication framework for cultural data
collection. CulturePark simulates cross-cultural human communication with
LLM-based agents playing roles in different cultures. It generates high-quality
cross-cultural dialogues encapsulating human beliefs, norms, and customs. Using
CulturePark, we generated 41,000 cultural samples to fine-tune eight
culture-specific LLMs. We evaluated these models across three downstream tasks:
content moderation, cultural alignment, and cultural education. Results show
that for content moderation, our GPT-3.5-based models either match or
outperform GPT-4 on datasets. Regarding cultural alignment, our models surpass
GPT-4 on Hofstede's VSM 13 framework. Furthermore, for cultural education of
human participants, our models demonstrate superior outcomes in both learning
efficacy and user experience compared to GPT-4. CulturePark proves an important
step in addressing cultural bias and advancing the democratization of AI,
highlighting the critical role of culturally inclusive data in model training.
|
CulturePark: Boosting Cross-cultural Understanding in Large Language Models: Cultural bias is pervasive in many large language models (LLMs), largely due
to the deficiency of data representative of different cultures. Typically,
cultural datasets and benchmarks are constructed either by extracting subsets
of existing datasets or by aggregating from platforms such as Wikipedia and
social media. However, these approaches are highly dependent on real-world data
and human annotations, making them costly and difficult to scale. Inspired by
cognitive theories on social communication, this paper introduces CulturePark,
an LLM-powered multi-agent communication framework for cultural data
collection. CulturePark simulates cross-cultural human communication with
LLM-based agents playing roles in different cultures. It generates high-quality
cross-cultural dialogues encapsulating human beliefs, norms, and customs. Using
CulturePark, we generated 41,000 cultural samples to fine-tune eight
culture-specific LLMs. We evaluated these models across three downstream tasks:
content moderation, cultural alignment, and cultural education. Results show
that for content moderation, our GPT-3.5-based models either match or
outperform GPT-4 on datasets. Regarding cultural alignment, our models surpass
GPT-4 on Hofstede's VSM 13 framework. Furthermore, for cultural education of
human participants, our models demonstrate superior outcomes in both learning
efficacy and user experience compared to GPT-4. CulturePark proves an important
step in addressing cultural bias and advancing the democratization of AI,
highlighting the critical role of culturally inclusive data in model training.
|
[
0.0540125109,
0.0445437953,
-0.0388169438,
0.0347582512,
-0.0073331231,
0.030727528,
0.0259558428,
-0.0404320508,
0.058728639,
-0.0097702593,
-0.0387906507,
-0.0490319133,
-0.0036659683,
0.0337497666,
-0.0067726197,
-0.0513432324,
0.0429844968,
-0.0092034377,
0.0496341102,
-0.0311501231,
0.0450747125,
0.0480684564,
0.0173976328,
0.0552802645,
-0.0374676771,
-0.0179024711,
0.0071876291,
-0.0114902398,
0.0533888638,
-0.0328020081,
0.0283862371,
0.0306001697,
0.0448451899,
0.0597280152,
0.0000022359,
-0.0352486037,
0.014721225,
-0.0347221941,
0.027174959,
0.0130291795,
0.0751283914,
-0.011161793,
-0.0084011424,
-0.0085408799,
-0.0709266812,
0.01361572,
0.0310587697,
0.0601178072,
0.0483596027,
0.0482223369,
0.0035721594,
-0.0107738944,
-0.0341494791,
-0.0166895967,
-0.0424483083,
-0.0286571532,
0.0281919632,
0.0589777566,
-0.0199602637,
-0.0072504226,
-0.0271710586,
0.0139109911,
0.0223893952,
-0.0325421877,
-0.0194191691,
-0.0025925466,
-0.028339332,
-0.0459659062,
0.0028495474,
0.0857436955,
0.0657394528,
-0.0175811555,
0.0091599077,
-0.0037836095,
0.0137185846,
0.0127856303,
-0.0518101864,
0.0401792526,
0.0081298091,
0.032427907,
0.0163010713,
-0.0136868954,
0.0343420543,
0.0398579687,
0.024329273,
0.0808840171,
-0.0196525045,
-0.015536136,
-0.0459048636,
0.0060574971,
0.014929899,
-0.0561891161,
0.0092066266,
0.0189897697,
-0.0544604287,
0.0217249151,
-0.0210476518,
0.0134452796,
0.0254588295,
-0.0537240095,
-0.0281716119,
0.0691922531,
-0.0039399271,
0.0436063819,
0.0324010663,
0.0200544205,
-0.0074791736,
0.0272831116,
0.0002689541,
-0.0359673575,
-0.0431376435,
-0.004267178,
0.0166611392,
-0.0070918319,
-0.0109215379,
-0.0053998125,
-0.0463528857,
0.0191185065,
-0.0580365919,
0.0389009789,
-0.0001478533,
-0.023014769,
-0.0200145897,
-0.0077837813,
-0.0343873575,
0.0759948939,
-0.059639927,
0.0038448335,
0.0006584155,
-0.0664025396,
0.0344710387,
-0.0118356971,
-0.0390557237,
-0.0090791266,
0.0102213575,
0.0151283471,
0.0009343207,
0.0165745188,
-0.0376116857,
0.0151988994,
0.0225839391,
-0.0709747225,
-0.0151783647,
-0.0674903318,
0.0551801361,
0.0133145014,
0.0183268283,
-0.0886283666,
0.018564973,
0.013470944,
-0.036111623,
0.0471908636,
0.0097204754,
0.0351573154,
0.0931277946,
0.0360316597,
-0.0448839664,
0.1065263748,
0.0341443121,
0.0076162005,
0.0696602687,
0.0001830166,
-0.0338958576,
0.0276243016,
0.0082488135,
0.0211707857,
-0.0025232432,
0.0089548472,
-0.0037003027,
0.0216929577,
0.0215811301,
0.0463086702,
-0.0099408757,
0.0044343015,
0.0573843755,
0.0523863994,
-0.0629639551,
0.0157305002,
0.0443635732,
0.0377763696,
0.0025221421,
-0.0045277202,
-0.0547039621,
0.0353747346,
0.0057384875,
-0.0195770469,
0.0034449727,
0.083684504,
-0.0504618995,
-0.0393781848,
0.013149824,
-0.014191539,
0.0873023421,
0.0321690403,
0.0116882706,
-0.0400177911,
0.0110142585,
0.0231474265,
0.0469691977,
-0.063143447,
0.0100284135,
-0.0173974466,
0.0637153685,
0.0742251873,
0.0019312782,
-0.0204143394,
0.0012987694,
-0.0092664771,
0.0571922213,
0.0691859201,
0.0876412019,
-0.0324649177,
-0.021211639,
-0.0457499139,
-0.0533468463,
-0.021795569,
-0.0121156573,
0.0274594612,
0.0553859249,
0.0375619307,
-0.0011006811,
0.0097179906,
-0.0623224378,
-0.0023913556,
-0.0615680665,
0.0151366759,
-0.041268032,
0.0090344129,
0.0231620166,
0.0555933714,
-0.0072426042,
-0.0381824262,
-0.0127837118,
-0.0062641357,
-0.01958788,
0.0335862748,
-0.0552517883,
-0.0034236505,
0.0515815876,
-0.0428987294,
-0.0521751046,
0.0272031371,
-0.0213532597,
-0.0149790114,
0.047786165,
0.0340566225,
0.0025732829,
-0.0366650708,
-0.0215648171,
-0.0247543007,
0.0000263965,
-0.0002083756,
0.0275917314,
-0.008984996,
-0.0203912575,
0.0228631273,
-0.0228688754,
0.0277906768,
-0.009486069,
-0.0277298763,
0.0383656099,
0.0084525505,
-0.0014226763,
0.0690819994,
0.0139711155,
-0.0456144214,
0.0030898284,
-0.1054997593,
-0.0271431953,
-0.0237144511,
-0.0130117266,
-0.0019685896,
0.0221808385,
-0.028738156,
-0.0104404818,
-0.0918582529,
0.0099396883,
0.0346885398,
0.034873791,
0.0241560526,
-0.0121572083,
-0.0087540215,
-0.0160924662,
0.0026870833,
0.0007882055,
-0.0000010877,
-0.0047587948,
-0.0382181816,
-0.0252169073,
-0.0133746108,
-0.0746348649,
0.0053415224,
-0.0562520996,
-0.0537747629,
0.0183414053,
0.0317507945,
0.0520858653,
0.049543079,
0.0364735052,
0.0029977167,
0.044386439,
0.0106233899,
-0.0638903677,
0.0021972284,
0.0117235631,
0.0267996304,
-0.0560788065,
-0.0689264089,
-0.0066044373,
0.0507050492,
-0.0170353763,
0.0142745543,
-0.0650388822,
0.0472663157,
-0.019337384,
-0.0400698856,
0.000013935,
-0.0312821679,
0.0040463125,
0.0405319408,
-0.0489412807,
-0.0172451697,
0.0845175385,
-0.0015182197,
0.0048188185,
-0.0111846216,
-0.0062789284,
-0.0468232036,
0.0310401078,
-0.030712055,
0.0446235724,
-0.025974717,
0.0033614871,
-0.0371561348,
-0.0264880545,
0.0531068817,
-0.0479293279,
0.0260893162,
-0.0294703897,
-0.0257646013,
-0.0182125997,
-0.0312200468,
-0.0386607759,
0.0323137678,
-0.0355204269,
0.0170772038,
0.0184354708,
-0.0096890628,
-0.0129630025,
-0.0091382693,
0.0625191554,
-0.0225104913,
0.0036965653,
0.0218095016,
0.0145689966,
-0.0402818024,
0.0559444539,
0.0194956735,
0.0074371877,
-0.0582561791,
0.0665460974,
0.0080364589,
0.046435032,
-0.0472583994,
0.0323830917,
-0.0214212462,
0.0397284143,
-0.0006299567,
-0.0074027884,
-0.009764133,
0.0829581693,
-0.0355778225,
0.0370453075,
0.1055476591,
0.0080536278,
-0.0637335926,
-0.0147983683,
-0.0483763777,
0.0202344023,
0.0253139902,
0.036108233,
-0.112409547,
0.007862878,
0.0050638276,
-0.0070045646,
-0.0210044924,
-0.0254620742,
0.0024804987,
-0.0993657559,
0.0096581196,
0.0152041866,
-0.0465578027,
-0.0377494022,
-0.0160320774,
-0.0097505935,
-0.0292803179,
0.0314814225,
-0.0439924672,
0.0279898085,
-0.0204929542,
-0.0179369692,
-0.0360622518,
-0.0504423045,
0.0338574089,
-0.0606719851,
0.0813923553,
0.0299268737,
0.0500440821,
0.0196602643,
-0.068407245,
-0.018494321,
0.0341374204,
-0.0154956263,
-0.0025092973,
0.0019070276,
-0.0279525202,
-0.0190434419,
-0.0026568258,
0.0215498991,
-0.0061221696,
-0.0178596899,
0.0072862138,
0.0649982095,
0.0050330763,
-0.0118427668,
-0.0092385616,
0.0135983806,
0.0312436428,
0.0878933966,
-0.0005700955,
-0.038995266,
-0.0092463419,
-0.0432689227,
-0.0315531492,
-0.0054372041,
-0.0471843258,
0.0118739065,
-0.0195611827,
-0.0107641527,
-0.03020861,
-0.0602974333,
0.0160157215,
-0.0309278071,
0.0398964882,
-0.0070615388,
0.0238623917,
0.0143944891,
-0.0091607347,
0.0284304675,
-0.0644216985,
0.0310693439,
-0.009813007,
-0.0367807709,
-0.0136189871,
-0.0462651923,
0.0240085479,
0.0161153488,
-0.0211859159,
-0.0654937252,
0.0260413699,
0.0473575965,
0.0030109249,
0.0579215214,
0.0331130251,
-0.0303155575,
0.0216427315,
0.0113284755,
-0.047626853,
-0.023529185,
0.0184002016,
0.0436096191,
0.0070681069,
0.0492844731,
-0.0467443131,
0.0196975376,
0.0055458392,
0.0368187502,
-0.0510190092,
-0.0095207635,
0.005134753,
0.0173035655,
-0.0311358217,
-0.0178159177,
-0.0160323307,
-0.0678740442,
-0.0286684278,
-0.0411363766,
-0.0387005471,
-0.0167554785,
-0.0189430658,
0.019538505,
-0.0362089276,
0.0248657912,
0.0196392834,
0.0053963661,
-0.0034103915,
-0.0422688909,
0.0455174409,
0.0015976097,
0.0516099334,
-0.0180499032,
-0.0041661086,
-0.0203410219,
0.0375217237,
0.0051339124,
0.017193567,
-0.009984755,
0.1014213786,
0.011027962,
-0.0200820658,
0.0459385253,
0.0091146352,
-0.0260433666,
-0.0386337452,
0.0291282535,
0.0271036141,
0.0580600277,
-0.0097115124,
0.0716257468,
-0.02086569,
-0.0215047579,
0.0024610669,
0.0198515467,
0.0280021373,
-0.057579942,
0.0254358966,
0.0077654785,
-0.0365556292,
0.0120549509,
0.0541254729,
-0.0391860753,
0.0112888757,
0.0028891072,
-0.0318717808,
-0.0216720831,
-0.0248039532,
0.051246684,
0.032886304,
-0.0468108058,
0.0094137629,
-0.0603908896,
-0.0369739756,
-0.0368863605,
0.0103932805,
-0.028208537,
-0.0294420142,
0.0002770442,
0.0686133727,
-0.0193809755,
-0.0157631338,
-0.0688221082,
-0.0087873228,
-0.0444957949,
-0.0176018402,
0.0064253109,
-6.385448414e-33,
-0.0196422301,
-0.0095159635,
0.0232018381,
0.0431410186,
-0.0377377756,
-0.0421135612,
-0.0398011766,
0.0148226144,
-0.042620834,
0.0388653092,
-0.0043762866,
0.0237684175,
0.0040968163,
-0.0055501973,
0.0159599204,
0.0019897304,
0.0421188399,
0.0040383809,
-0.0315135457,
0.049077671,
0.0258419961,
0.0124014365,
0.0445409939,
-0.0855081826,
0.0554360673,
-0.0021078058,
0.008748496,
0.0095010716,
0.0243768748,
0.0563422889,
-0.0352924988,
0.0408906527,
0.0116466833,
-0.0425644703,
0.0063950918,
0.0103190672,
-0.0628848299,
-0.0180337764,
0.0340709761,
0.0195789263,
0.0185760483,
-0.0618631914,
0.003140551,
-0.0466290899,
-0.0333535932,
0.0280861557,
0.0403658822,
-0.026959721,
0.0412782766,
0.0042659659,
-0.0710283518,
0.0213927832,
-0.0256426148,
0.0008558441,
0.0071248785,
0.0486794189,
0.009837701,
0.026536569,
-0.0673923939,
0.0431084111,
-0.011371607,
-0.0033894747,
0.1013425812,
-0.0160227083,
0.0057314537,
0.0594660118,
0.0137522332,
0.0575480387,
-0.0276639964,
-0.0284385514,
0.0256403834,
0.015858287,
-0.0105424924,
0.1054934487,
-0.0306845661,
-0.0426846072,
-0.0378131494,
0.0215332545,
0.0674229264,
0.005162532,
0.0232802983,
0.0003772865,
-0.0323320702,
-0.0171407126,
-0.0295812692,
-0.0113321086,
-0.0059367432,
-0.051161129,
0.012901362,
0.0177373793,
-0.025249593,
-0.035398189,
0.0246267077,
-0.0610346459,
-0.0231851842,
0.0381893814,
0.0062682419,
-0.0419426449,
0.0277038142,
-0.0027253926,
-0.0767864957,
-0.017391419,
-0.0277896486,
-0.0047158548,
-0.0048297532,
0.0214581005,
0.014249309,
0.0204751547,
-0.0290760212,
0.0077562132,
-0.016819926,
0.0144033106,
-0.0252101645,
-0.0208938401,
0.0136762783,
-0.0005682508,
-0.0194899216,
-0.0092947688,
0.0372777618,
-0.0156949684,
-0.0090708081,
-0.0902027637,
-0.032654088,
0.0045186672,
-0.0479630604,
-0.0147786792,
-0.0527207889,
0.0315453783,
0.0128817391,
-0.0263844114,
-0.0160891619,
-0.0206468552,
2.868e-7,
-0.0133072296,
0.0320065059,
-0.0175966732,
0.0599604435,
0.0182286706,
0.0110705942,
-0.0225485936,
0.036480844,
0.004061277,
-0.0447489955,
0.0158289913,
0.0676659197,
0.0769636855,
0.0421703085,
-0.0897779837,
-0.0130401943,
-0.0338421762,
0.005760367,
-0.0553603098,
0.0282329042,
-0.026303459,
0.0395909064,
0.0101543348,
0.0000595546,
-0.0225427225,
-0.0025123148,
0.0608814023,
-0.0571689978,
-0.0189097207,
0.0299239513,
0.019364547,
-0.0207103007,
0.0154609922,
0.0026955924,
0.0066610975,
0.0196652915,
0.0043751509,
0.0760892853,
0.0206429195,
0.0018929337,
0.0022454867,
0.0180251542,
0.001564549,
-0.0824788511,
0.0371722504,
-0.0113986228,
-0.0414046124,
-0.0282363445,
0.0039077401,
-0.0324285924,
0.0371258371,
-0.0176332574,
-0.0326434523,
0.0046815029,
0.0145076737,
-0.0169411786,
0.0040205116,
-0.0281558838,
-0.0103814211,
-0.0488513522,
-0.0060898731,
-0.0693767071,
-0.0229236465,
-0.0342729837,
-0.0514572896,
0.0204836819,
0.0309470259,
2.251244562e-34,
0.039441213,
-0.0215430204,
-0.015318281,
0.0067708152,
0.0554930456,
-0.0072145602,
0.015747888,
-0.0133454828,
0.0245400239,
-0.030633999,
0.0059584999
] |
http://arxiv.org/abs/2405.15143v2
|
http://arxiv.org/pdf/2405.15143v2
|
Intelligent Go-Explore: Standing on the Shoulders of Giant Foundation Models
| 2024-05-24T01:45:27 | 2024-05-30T06:48:44 |
Cong Lu, Shengran Hu, Jeff Clune
|
cs.LG
|
cs.LG, cs.AI, cs.CL
|
Go-Explore is a powerful family of algorithms designed to solve
hard-exploration problems, built on the principle of archiving discovered
states, and iteratively returning to and exploring from the most promising
states. This approach has led to superhuman performance across a wide variety
of challenging problems including Atari games and robotic control, but requires
manually designing heuristics to guide exploration, which is time-consuming and
infeasible in general. To resolve this, we propose Intelligent Go-Explore (IGE)
which greatly extends the scope of the original Go-Explore by replacing these
heuristics with the intelligence and internalized human notions of
interestingness captured by giant foundation models (FMs). This provides IGE
with a human-like ability to instinctively identify how interesting or
promising any new state is (e.g. discovering new objects, locations, or
behaviors), even in complex environments where heuristics are hard to define.
Moreover, IGE offers the exciting and previously impossible opportunity to
recognize and capitalize on serendipitous discoveries that cannot be predicted
ahead of time. We evaluate IGE on a range of language-based tasks that require
search and exploration. In Game of 24, a multistep mathematical reasoning
problem, IGE reaches 100% success rate 70.8% faster than the best classic graph
search baseline. Next, in BabyAI-Text, a challenging partially observable
gridworld, IGE exceeds the previous SOTA with orders of magnitude fewer online
samples. Finally, in TextWorld, we show the unique ability of IGE to succeed in
settings requiring long-horizon exploration where prior SOTA FM agents like
Reflexion completely fail. Overall, IGE combines the tremendous strengths of
FMs and the powerful Go-Explore algorithm, opening up a new frontier of
research into creating more generally capable agents with impressive
exploration capabilities.
|
Intelligent Go-Explore: Standing on the Shoulders of Giant Foundation Models: Go-Explore is a powerful family of algorithms designed to solve
hard-exploration problems, built on the principle of archiving discovered
states, and iteratively returning to and exploring from the most promising
states. This approach has led to superhuman performance across a wide variety
of challenging problems including Atari games and robotic control, but requires
manually designing heuristics to guide exploration, which is time-consuming and
infeasible in general. To resolve this, we propose Intelligent Go-Explore (IGE)
which greatly extends the scope of the original Go-Explore by replacing these
heuristics with the intelligence and internalized human notions of
interestingness captured by giant foundation models (FMs). This provides IGE
with a human-like ability to instinctively identify how interesting or
promising any new state is (e.g. discovering new objects, locations, or
behaviors), even in complex environments where heuristics are hard to define.
Moreover, IGE offers the exciting and previously impossible opportunity to
recognize and capitalize on serendipitous discoveries that cannot be predicted
ahead of time. We evaluate IGE on a range of language-based tasks that require
search and exploration. In Game of 24, a multistep mathematical reasoning
problem, IGE reaches 100% success rate 70.8% faster than the best classic graph
search baseline. Next, in BabyAI-Text, a challenging partially observable
gridworld, IGE exceeds the previous SOTA with orders of magnitude fewer online
samples. Finally, in TextWorld, we show the unique ability of IGE to succeed in
settings requiring long-horizon exploration where prior SOTA FM agents like
Reflexion completely fail. Overall, IGE combines the tremendous strengths of
FMs and the powerful Go-Explore algorithm, opening up a new frontier of
research into creating more generally capable agents with impressive
exploration capabilities.
|
[
0.0038618688,
0.0648403317,
-0.0160250589,
0.0229006056,
-0.0339370109,
-0.0081251673,
-0.0563339815,
0.0031686299,
0.0068339217,
0.0266076308,
-0.0102071147,
-0.0378801003,
-0.0459236354,
0.0761071369,
-0.024656212,
-0.0451991446,
0.0207090806,
-0.0379853025,
0.0170194693,
-0.0027314406,
-0.0591411144,
-0.0486631356,
-0.0365558155,
0.0199137013,
0.0034107466,
-0.0272227768,
-0.0765974224,
-0.0212973077,
0.0306846835,
-0.073832348,
-0.056032747,
-0.0048516118,
0.0644476637,
0.0645238832,
0.0000022359,
-0.0377404355,
-0.066190131,
0.0003435415,
-0.0074155363,
-0.0289711561,
0.0181409661,
0.0213135164,
-0.0206181947,
-0.0200619847,
-0.065111652,
-0.0280083474,
0.0542506427,
0.0016709138,
0.0139771551,
0.0504560024,
0.0104162339,
-0.0085672308,
0.0063116485,
0.004785622,
0.0234789029,
-0.0814728066,
0.0119185252,
0.0182201825,
0.0534994043,
0.0105916299,
-0.0429017656,
-0.0235336088,
0.0316218548,
0.0011013512,
0.0006453012,
0.0613992773,
0.0079164058,
0.0072713383,
-0.0021852241,
0.0446744524,
-0.0224237368,
-0.0187365636,
0.0230197087,
-0.0023710665,
-0.0185046848,
0.0184537955,
-0.0473941602,
0.044062674,
-0.0446798168,
0.0130797084,
0.0041584792,
0.0032518327,
0.0212532002,
0.007276196,
-0.0541780032,
0.1239001527,
-0.0259046797,
-0.0236572083,
0.0131138889,
0.0134802563,
-0.0497559719,
-0.0575337559,
0.0299752541,
0.0433980301,
-0.0432096161,
-0.0081303092,
0.0212526359,
0.0434427522,
-0.0297874846,
0.0366835818,
0.0285114888,
0.0365875587,
0.0315703824,
0.0620207042,
0.0327283442,
0.0312620699,
-0.0773806125,
-0.0263021365,
-0.0391666032,
-0.0059912247,
-0.0660250112,
-0.0137458453,
-0.0628431886,
0.0770577267,
0.0024232997,
-0.0699175298,
0.0363508537,
0.0372097678,
0.0085518677,
-0.014989539,
-0.0403033681,
0.0171332266,
-0.0058365399,
-0.0013103896,
-0.0438611805,
-0.0454275049,
-0.101625368,
-0.0098297223,
0.0118429642,
0.0223125387,
-0.0208267905,
0.0025872344,
-0.0093359072,
-0.0055922791,
0.0190290473,
0.0404995903,
0.0446019806,
-0.031501323,
-0.0237371512,
0.0021680589,
0.04349529,
-0.0800391212,
0.057443995,
-0.039746616,
-0.0009176667,
-0.0210157763,
-0.0241673496,
-0.0241386704,
0.0036483812,
0.0244377591,
0.042369768,
0.0435191058,
0.0027688681,
0.0247241799,
0.0012797904,
-0.0027496472,
-0.040938329,
0.0628213808,
0.0212955344,
0.0203512944,
0.0150766643,
0.0317708179,
0.0397077948,
0.0280676913,
0.0114841517,
0.0041610412,
0.0578641742,
-0.0127501488,
0.0023561523,
0.0852112994,
-0.015709186,
0.025812719,
-0.0484425388,
0.0018123694,
0.0189496391,
0.0638033375,
0.0043586679,
0.0469921753,
0.0613082163,
-0.0039988589,
-0.0194866322,
-0.042622596,
-0.0211061873,
0.0167157669,
0.0040734224,
0.0129813217,
-0.0549448952,
0.020962324,
0.0040953849,
-0.0259983875,
-0.0071490966,
0.0368236452,
0.0431232154,
-0.006990015,
0.0475988425,
0.0034655805,
-0.0631709918,
-0.0615398064,
-0.0369132869,
-0.0072799409,
-0.009564626,
-0.0654242188,
0.0815007016,
0.0037144795,
0.0815026984,
-0.0799647123,
-0.0363871865,
0.0208914895,
0.025640687,
0.0420990549,
0.0045916862,
0.0054837009,
0.0163001288,
-0.0340229236,
-0.0010304304,
-0.0294592902,
0.000341657,
0.0549847558,
0.0430668406,
0.011669266,
-0.0494684987,
-0.04189533,
0.0283925589,
0.0155926235,
0.0031824484,
0.0067007602,
0.0031566578,
-0.0125769954,
0.0092555527,
0.0062327175,
0.0160777252,
0.0185537841,
0.0424484313,
0.0128683336,
-0.0539148524,
0.0033748855,
-0.0046699573,
0.0002506228,
0.034017317,
-0.0594230145,
-0.0004076184,
0.0160703603,
0.0103610884,
-0.0006997201,
0.0411103927,
-0.0430443995,
0.030488858,
0.0120060919,
0.0160426013,
-0.0211285781,
-0.0186609644,
-0.007881335,
0.0949900746,
-0.0104208272,
-0.0256777685,
0.0094885351,
-0.0161920916,
0.0241799559,
-0.0753906816,
0.0247447863,
0.0304759555,
-0.008273324,
-0.0257686321,
0.01966401,
0.0070034182,
-0.0303573459,
-0.0071579372,
-0.1001768112,
0.0310659315,
0.0207046047,
0.0083231945,
0.0042739776,
-0.0397113562,
-0.0083269849,
0.0151898963,
0.0182311684,
0.045105841,
0.0157103408,
-0.004925542,
0.0248484127,
0.035773363,
-0.0260896888,
-0.0112229399,
-0.0091395061,
-0.0335361101,
0.0295274351,
0.0104763526,
0.0326433033,
-0.0303294621,
-0.0130988993,
-0.0547959171,
0.0310479589,
-0.0196954254,
-0.0116738882,
0.0606475584,
-0.0717330128,
-0.0500717126,
0.0800665766,
-0.0143828318,
0.0669473782,
0.0366174281,
-0.0676487759,
-0.0025791528,
0.0439956263,
-0.0364773311,
0.0303746033,
0.022104891,
0.0051593534,
0.0269185547,
0.0503180213,
-0.0280030053,
-0.0097148381,
-0.0531267412,
0.031674128,
-0.0030876023,
-0.038299419,
0.0004696572,
0.027579302,
-0.0186212566,
0.0267300401,
-0.0484345108,
0.0113652665,
0.0966845751,
0.0211721249,
-0.0163592268,
-0.0316173583,
0.0183623377,
-0.0645286292,
0.0882663727,
-0.0176076461,
0.0254097935,
0.0064703887,
-0.0051546455,
0.0170744974,
-0.008331717,
0.0180152263,
-0.0079782782,
0.0264149886,
-0.0536393523,
0.003022294,
0.0499409214,
-0.0825617239,
-0.0225372873,
0.0112643978,
-0.0277034044,
-0.0138113741,
0.0067636729,
-0.0739273056,
-0.0656972155,
-0.0169432536,
0.0464368016,
0.023877969,
-0.0134670464,
-0.008606662,
-0.0159094762,
-0.0095407125,
0.0200096928,
0.0281883385,
0.0396426097,
-0.0264430698,
0.0419257507,
0.042287834,
0.0311028771,
-0.0435879305,
-0.0047978247,
0.0380596854,
-0.0134556675,
-0.000837958,
-0.0431038998,
0.021217484,
0.0174111016,
-0.0391961075,
-0.0915859938,
0.0430009551,
0.0179249682,
0.0023774358,
0.008238419,
-0.0764178708,
0.0278722271,
0.0094097294,
0.0388700627,
-0.0087279323,
-0.0230669528,
0.0190129299,
-0.0068857581,
-0.0246151313,
-0.0615742579,
0.0131129427,
-0.1075626984,
0.0274384506,
0.0264322273,
-0.0661021173,
0.0066973274,
-0.0247428305,
-0.0130913714,
-0.0028868462,
-0.0933086053,
0.0264067128,
0.011830993,
0.0135923894,
-0.0198341142,
0.0314894132,
0.0071999314,
0.0158740208,
-0.0182300862,
-0.0116532063,
0.0379087478,
0.0136565799,
-0.0168036781,
-0.0148837557,
-0.0226656497,
0.0391983017,
0.0371785499,
-0.0160931442,
0.0041133151,
0.0088844299,
0.011689269,
-0.0215935297,
0.03825216,
-0.0195998494,
-0.0730825141,
0.0209821388,
0.0975545719,
-0.0164423212,
-0.0255743153,
-0.0304719601,
-0.0015196636,
0.0445947051,
0.0532056503,
-0.0289140958,
-0.0312663987,
-0.0356230102,
-0.0468282402,
0.0145447431,
-0.0420881733,
-0.032656353,
-0.0168469734,
0.0753102452,
-0.0234996211,
-0.052179303,
-0.0251867417,
0.0187995937,
0.056748081,
0.009314579,
-0.0127234226,
-0.0045821471,
0.0266393125,
-0.0177072752,
0.0108997701,
-0.0281897485,
0.0094172843,
-0.06590233,
0.0236345939,
0.0262218621,
-0.0739787966,
-0.0235639811,
0.057275556,
0.0012074453,
-0.0257495772,
0.0125549901,
0.050532572,
0.029420292,
0.0065287733,
-0.0066581019,
0.0146100633,
0.0091470964,
0.0069967164,
0.008293651,
0.0222047195,
0.0287053958,
-0.0202592779,
0.0319446176,
-0.0074637276,
-0.0224982016,
-0.0215514544,
-0.0208201427,
0.0476616025,
-0.0260198042,
0.0403913148,
0.0300852116,
-0.0239895154,
0.0485784411,
0.0186523553,
0.0035899787,
-0.0105126696,
-0.0231336541,
-0.0283194687,
0.0119231762,
-0.0385349095,
0.0221729018,
0.0386622623,
0.0074395593,
0.0233400501,
0.0332178921,
0.0153249912,
0.0310971979,
-0.0050779004,
-0.019056309,
0.0162973404,
0.0454149693,
0.0236866362,
-0.0061862813,
0.0093459878,
0.0155140702,
-0.0224841814,
0.0515988879,
0.0139932707,
0.0797677413,
0.05760663,
-0.0498491339,
-0.0405113958,
0.0162660088,
-0.0810683295,
-0.0798587725,
0.0133145386,
0.0033756301,
0.0890150145,
0.0492755249,
-0.0524788424,
-0.047955811,
0.0020581705,
0.0159157105,
-0.0286286436,
0.0266061649,
-0.087047793,
0.0265585706,
0.0494278297,
0.0103707258,
-0.0135767264,
0.0184823722,
0.0164118428,
0.0692380443,
0.0039800126,
0.0036187875,
-0.0235420689,
-0.0091740256,
-0.0263687763,
0.0689790174,
-0.0055468255,
0.0170561057,
-0.0649069548,
-0.0101897074,
0.0183041915,
0.0297016911,
-0.0276292637,
-0.014439147,
0.0901747048,
0.0097151082,
-0.0264067277,
-0.0506506823,
-0.0296150241,
-0.0036573589,
-0.0540241823,
0.005689126,
-0.0253628884,
-5.953122897e-33,
-0.000534571,
-0.0243988577,
0.0524928533,
0.00399367,
-0.0264181998,
-0.0428310223,
-0.0521754175,
0.0037227985,
-0.015632866,
0.0152326692,
-0.0425098836,
0.0282857455,
-0.0055768639,
0.0007927624,
-0.0172047671,
-0.0018656477,
0.0111495359,
0.03067871,
-0.0623819493,
-0.0124487299,
0.0172207747,
0.0105917733,
0.0244379546,
-0.0517267101,
-0.0608823746,
-0.0035734114,
-0.0315126516,
0.0485773496,
-0.0524974503,
0.0548469387,
-0.0183459744,
-0.0382863656,
-0.0142685296,
-0.0372031555,
-0.0057947044,
-0.0067200079,
-0.021562878,
-0.0527076498,
-0.0253747981,
-0.0105583258,
-0.0203186199,
-0.0409064889,
-0.0346260145,
0.0284361169,
-0.0245821904,
0.002736981,
-0.0107617993,
-0.0105425725,
0.0131198373,
0.0191268027,
0.0042429329,
0.0104491683,
0.0187992528,
-0.0304651223,
0.0328106061,
-0.0089276014,
-0.0128595792,
-0.0004055036,
0.043321304,
0.0610181876,
0.0045824023,
-0.0518350266,
0.0096383793,
0.008504346,
0.0534351952,
0.0451704934,
-0.0028756575,
-0.0182240959,
-0.039492406,
-0.0178673044,
-0.0037399281,
0.0472661518,
-0.002771517,
0.0375413336,
0.0281677619,
0.055914484,
-0.047716517,
0.0174563564,
-0.0025207775,
0.0481243655,
0.0332121849,
0.0374735892,
-0.0622561239,
-0.0021980968,
0.0255309995,
-0.0139793232,
-0.0185691565,
-0.0136696948,
0.0267204046,
-0.0073077162,
-0.0357005522,
-0.0217989422,
0.0301771741,
-0.0474525467,
0.0153034804,
0.0918006897,
-0.026163023,
0.0570908636,
0.0020277584,
0.0113970079,
-0.0252402928,
-0.0773549601,
-0.0060222503,
0.0177978594,
0.0214115959,
0.023154771,
-0.0814243704,
0.0636989698,
0.0053447625,
-0.0094386488,
-0.0092723295,
0.0108152954,
0.0330513567,
-0.0421361625,
0.0327321626,
-0.0261565465,
0.0010538418,
0.0474103577,
0.0280520897,
0.0170887727,
-0.0222608242,
0.0019036056,
-0.0133428695,
0.0256695263,
0.0136007015,
-0.016536884,
-0.0708534494,
-0.0178554431,
-0.0127455397,
-0.0784107447,
-0.0202646516,
0.0720225349,
2.984e-7,
-0.0228300225,
0.045281101,
0.0289499853,
0.0187905133,
0.0140463803,
0.0444524921,
-0.0122041944,
-0.0203427617,
-0.0032818487,
-0.0140452711,
0.0541406982,
0.0579436161,
-0.0151696391,
0.0036109171,
-0.0865777582,
-0.0035750596,
0.0190778226,
-0.0170528106,
-0.0721326098,
-0.011414797,
0.005164165,
0.0631230697,
0.0443130955,
0.0246192366,
0.046060998,
-0.0653834939,
0.0423851945,
-0.000481543,
0.0677080825,
0.0060732812,
-0.05970186,
0.0291114226,
-0.0368395001,
-0.0114336144,
0.0050330767,
0.0559657365,
0.0196054447,
0.0275903195,
0.0096322913,
-0.0043385867,
-0.0046655824,
-0.0139842676,
-0.0015810567,
-0.0569563396,
0.0585159697,
-0.0473311208,
-0.0170536935,
0.0545801148,
0.0057438673,
-0.0496478863,
-0.0384220295,
-0.0246919636,
-0.0176002402,
-0.046715878,
-0.0385780185,
0.0187365394,
0.0020506484,
-0.0556246489,
0.021740539,
0.0100555234,
-0.0510342531,
0.0116614439,
-0.0456568189,
0.0214773044,
0.0170573592,
-0.0740258917,
-0.0293039419,
2.890695351e-34,
0.0283931438,
-0.0360866003,
-0.0047884434,
0.0030586144,
0.0526269674,
0.0292073712,
0.0055959043,
-0.0185349099,
0.0412012264,
-0.0404582843,
-0.0365812927
] |
http://arxiv.org/abs/2405.15134v2
|
http://arxiv.org/pdf/2405.15134v2
|
Efficient Biomedical Entity Linking: Clinical Text Standardization with Low-Resource Techniques
| 2024-05-24T01:14:33 | 2024-05-27T01:36:38 |
Akshit Achara, Sanand Sasidharan, Gagan N
|
cs.CL
|
cs.CL
|
Clinical text is rich in information, with mentions of treatment, medication
and anatomy among many other clinical terms. Multiple terms can refer to the
same core concepts which can be referred as a clinical entity. Ontologies like
the Unified Medical Language System (UMLS) are developed and maintained to
store millions of clinical entities including the definitions, relations and
other corresponding information. These ontologies are used for standardization
of clinical text by normalizing varying surface forms of a clinical term
through Biomedical entity linking. With the introduction of transformer-based
language models, there has been significant progress in Biomedical entity
linking. In this work, we focus on learning through synonym pairs associated
with the entities. As compared to the existing approaches, our approach
significantly reduces the training data and resource consumption. Moreover, we
propose a suite of context-based and context-less reranking techniques for
performing the entity disambiguation. Overall, we achieve similar performance
to the state-of-the-art zero-shot and distant supervised entity linking
techniques on the Medmentions dataset, the largest annotated dataset on UMLS,
without any domain-based training. Finally, we show that retrieval performance
alone might not be sufficient as an evaluation metric and introduce an article
level quantitative and qualitative analysis to reveal further insights on the
performance of entity linking methods.
|
Efficient Biomedical Entity Linking: Clinical Text Standardization with Low-Resource Techniques: Clinical text is rich in information, with mentions of treatment, medication
and anatomy among many other clinical terms. Multiple terms can refer to the
same core concepts which can be referred as a clinical entity. Ontologies like
the Unified Medical Language System (UMLS) are developed and maintained to
store millions of clinical entities including the definitions, relations and
other corresponding information. These ontologies are used for standardization
of clinical text by normalizing varying surface forms of a clinical term
through Biomedical entity linking. With the introduction of transformer-based
language models, there has been significant progress in Biomedical entity
linking. In this work, we focus on learning through synonym pairs associated
with the entities. As compared to the existing approaches, our approach
significantly reduces the training data and resource consumption. Moreover, we
propose a suite of context-based and context-less reranking techniques for
performing the entity disambiguation. Overall, we achieve similar performance
to the state-of-the-art zero-shot and distant supervised entity linking
techniques on the Medmentions dataset, the largest annotated dataset on UMLS,
without any domain-based training. Finally, we show that retrieval performance
alone might not be sufficient as an evaluation metric and introduce an article
level quantitative and qualitative analysis to reveal further insights on the
performance of entity linking methods.
|
[
0.0143668037,
0.0043290658,
0.0205409825,
-0.0426675268,
0.0073828013,
0.0256761536,
0.0055245077,
0.035172604,
0.0108664548,
-0.0464673638,
-0.0221376885,
0.0109711681,
0.010949322,
-0.0323304199,
0.0053983661,
-0.0279905517,
0.0361247472,
0.0448911674,
-0.0605658293,
-0.015874723,
-0.0357449464,
0.0221970957,
0.0084208064,
0.0032651478,
0.0089079915,
0.0217512418,
0.0552958436,
0.0087165553,
0.0461133234,
-0.0366193801,
-0.0385777727,
0.0388874449,
-0.0296726562,
0.0381517485,
0.0000019353,
-0.0675871223,
-0.0854445621,
0.0387530848,
-0.0238645989,
-0.0548896827,
0.0650232434,
-0.0959949568,
0.0098081008,
-0.0032404503,
-0.0191428587,
-0.0641804188,
0.0186116844,
0.0268258173,
0.0203267317,
0.0415904708,
-0.0155736571,
-0.0563379191,
0.0244505312,
-0.0066045355,
-0.0164898802,
-0.0406648926,
0.0484227501,
0.009651755,
0.0034699806,
0.0167164207,
-0.029684227,
0.0619386286,
-0.0214672796,
0.0335044302,
0.0320963822,
0.0095896851,
-0.0228427686,
0.004568506,
0.0471250266,
0.078438431,
-0.0225770734,
-0.0528793223,
0.0504139587,
0.0364183336,
-0.0108221592,
0.0515894294,
-0.0008725569,
-0.0181532186,
-0.0261950269,
0.0102736689,
0.0160245206,
0.005047285,
0.0396100022,
0.007114769,
-0.0378812961,
-0.00994479,
0.0075694858,
0.0002476136,
0.0048714704,
-0.0156208463,
-0.0028370717,
-0.0247803088,
0.0326231867,
0.0213230662,
-0.0510422252,
0.0314010978,
-0.045178622,
-0.0610764585,
0.039370846,
-0.0674150065,
-0.0737880766,
-0.0062880469,
0.0459905975,
0.0274035763,
-0.0367756113,
-0.0156854168,
-0.0071007139,
0.0229647215,
-0.0289952029,
-0.0548982099,
-0.0541032031,
0.0316295065,
-0.0422987267,
0.0403198302,
-0.0001716822,
-0.0011568223,
-0.0912383273,
-0.0371863805,
-0.0208432525,
0.0560121126,
0.0181596838,
-0.003037225,
-0.0240576193,
-0.0243532863,
-0.0088082971,
0.0221911557,
0.0186224505,
-0.0250272118,
-0.0379090607,
-0.0348743126,
0.0445633121,
-0.0181809254,
-0.0044377069,
-0.0288804378,
0.0180387069,
0.0624243543,
0.0391791761,
-0.0192054939,
-0.0084934477,
-0.0485486053,
0.0183459334,
0.0261104796,
-0.0372994617,
-0.0877523273,
0.0187279396,
0.0286981482,
0.0212007556,
-0.0766359344,
0.0132056512,
0.0030571984,
-0.0973697007,
0.0989814773,
-0.056742277,
-0.0245098975,
0.0372357555,
0.0570955016,
0.0232448429,
-0.0072033643,
0.0255379435,
-0.0482871868,
0.0328501649,
0.0090217143,
-0.0155808274,
-0.0281195845,
0.0306459814,
-0.0471140482,
-0.0080200387,
-0.0051015834,
0.0087371003,
0.0673221424,
-0.0019654483,
0.105167605,
-0.0271077771,
-0.0572748445,
-0.0097907716,
0.0249512549,
0.0649966225,
0.0041453992,
0.022085974,
-0.0145101035,
-0.0280467533,
-0.0323734134,
-0.0254096687,
-0.0061460757,
0.0501376018,
0.0064483746,
0.0088779479,
0.0324205868,
-0.0070784953,
0.0022689421,
-0.003481609,
-0.042909909,
-0.0463551246,
0.0011393385,
0.049852293,
0.0208275374,
-0.0108744958,
0.1144884154,
0.0440798737,
-0.0044860686,
0.011114045,
-0.0242759176,
-0.0623906665,
0.05445702,
0.061768759,
-0.0147941513,
-0.0495510809,
-0.0297806226,
-0.017535504,
0.01099212,
-0.0112662101,
-0.056597285,
0.0041974457,
-0.0370560139,
0.0321192481,
0.0627993718,
0.0095249498,
-0.0095463442,
0.0344017558,
0.0653514713,
0.0259728823,
-0.066944778,
0.0096963523,
0.0281626973,
-0.0468793362,
-0.0025250376,
-0.0109897824,
0.0026880035,
-0.0145615293,
0.0313660875,
-0.0162772331,
0.0086269313,
0.0287896041,
-0.0340109207,
-0.0357666947,
-0.0017638194,
0.0056081186,
-0.0168853588,
0.0287328269,
-0.0557996705,
0.024555428,
-0.0113700451,
-0.0277187545,
-0.0418909714,
-0.0249463301,
0.0051470706,
0.0491271615,
0.010258751,
0.0467479751,
0.0205965564,
-0.0142009379,
-0.0549677461,
0.0375085399,
-0.017637141,
0.044278834,
0.0216843449,
-0.0094954809,
0.0127891628,
0.0155245373,
-0.0600171238,
0.0192069132,
0.0034115589,
0.0077538858,
-0.0013289211,
-0.0518324636,
-0.0324892551,
-0.0290231984,
-0.0827604532,
0.0251209773,
0.0223957375,
0.016613394,
0.042283427,
0.0557703264,
0.0199918058,
-0.0241225734,
0.0003135543,
-0.0306408703,
0.0098841265,
-0.0054986328,
0.0224158373,
0.0149362097,
-0.0066596922,
-0.0409455635,
-0.011977287,
-0.0137186758,
0.0215756409,
-0.0105672106,
-0.0319116563,
-0.0113564925,
0.0410606861,
-0.0583059043,
0.0137472833,
-0.0009467392,
0.0232387725,
-0.0140345059,
0.03646034,
0.0001879252,
0.0678645819,
0.004246335,
0.0327526033,
0.0606424585,
0.0290477667,
-0.0539759472,
-0.017883854,
-0.0219750702,
-0.0189679042,
-0.0080503607,
-0.0442015156,
0.0174759254,
0.0343539789,
-0.0171702504,
-0.0229870658,
-0.0022843669,
-0.0066593732,
-0.01746553,
0.0238398183,
-0.0194858443,
0.0526390485,
-0.0089257322,
-0.0058436594,
0.0269430894,
-0.0244972836,
0.0479053669,
0.0047937203,
0.0600215159,
-0.0551927462,
-0.0067548542,
-0.012406569,
-0.0072557093,
-0.0089969048,
0.0404052995,
-0.0180537961,
-0.020507263,
-0.0045080455,
0.0373850018,
0.0386085026,
-0.0590341017,
0.0621797964,
-0.0086869551,
0.0485574827,
0.0025195626,
-0.0190537032,
-0.0313222334,
0.0357649848,
-0.0231733397,
0.0350383297,
0.0189788584,
-0.0065391907,
-0.0457321368,
0.0444873907,
0.0175151378,
0.0582442768,
0.0598323569,
-0.0448805206,
0.0134477513,
0.0322484411,
0.0369675048,
-0.0345704779,
0.0137602873,
0.0365595333,
0.0327980556,
0.0570857078,
0.0310466196,
-0.0236467849,
0.0058053331,
-0.045411177,
0.0214465987,
-0.091273509,
-0.0315797701,
-0.0059737931,
-0.0145302154,
-0.0136001837,
-0.0009171136,
0.1212935224,
0.0567169189,
-0.0556163713,
0.00996459,
0.0188963898,
-0.0359712578,
-0.0216124058,
0.000793484,
-0.0919968858,
-0.0176082868,
-0.0147590432,
-0.0074075377,
0.0146464175,
-0.0572011545,
0.0343313627,
-0.0310963057,
-0.0229510199,
-0.0169018935,
-0.0515383519,
-0.0271322187,
-0.0373093262,
0.0060835644,
-0.0417115092,
-0.0188656971,
0.0006666246,
-0.0112853535,
0.0024010348,
-0.0098485472,
0.0034384856,
-0.0881184489,
-0.0026182807,
-0.0834983215,
0.0779790133,
-0.0098053729,
0.0306293909,
0.0041702059,
-0.0232021343,
-0.0025652414,
-0.0065692761,
-0.0377006195,
-0.000520462,
0.0552663505,
-0.048878774,
0.0155210514,
-0.0417601094,
-0.0011272415,
-0.002617111,
-0.0332247727,
-0.0595746227,
0.0179622956,
-0.007019321,
-0.0015947853,
-0.0212542377,
-0.0275152437,
-0.0103881462,
0.0601520874,
-0.0506159477,
0.0223767795,
-0.0017874256,
0.0276745874,
0.061664179,
0.0323750079,
-0.0043046395,
0.0315199904,
-0.1201517656,
-0.0387698524,
-0.07452894,
0.0057011005,
0.0330466218,
0.0136798974,
-0.0025426955,
0.0325773805,
0.0254134871,
-0.0142444773,
-0.080142647,
0.0318434797,
0.0063147284,
0.0227822028,
-0.025944015,
-0.0276862718,
0.0868392438,
0.0423620343,
-0.0222329218,
-0.0013259946,
0.01754448,
-0.0188874509,
-0.0029478336,
-0.0194105413,
0.023143027,
0.0087909587,
-0.0131989978,
-0.0072269253,
0.0174897611,
0.0672032908,
-0.0183986742,
-0.0071385941,
-0.0254107602,
0.0195587426,
-0.0054613505,
0.0485581718,
-0.0277875271,
0.0214439929,
-0.0131027242,
0.0283598546,
-0.0536013655,
-0.0329273753,
0.0147542721,
0.0074314955,
0.0204843488,
-0.0419483669,
0.0203359965,
0.0046570743,
0.0113954162,
-0.0000526787,
0.0187403616,
0.003236156,
0.020850068,
0.0465889871,
0.0384269208,
-0.0199462138,
0.0540537722,
-0.0183764547,
-0.0334351771,
0.0108610913,
0.0206947755,
-0.0144960368,
-0.035692092,
0.0222771578,
-0.0111172246,
0.0045550615,
-0.0237366576,
0.0442645922,
0.0383302756,
-0.0533210039,
-0.0038080581,
0.0256796423,
0.0361727886,
-0.035143353,
0.0617753938,
-0.0246828757,
-0.0015643683,
-0.01456994,
0.0086271362,
0.0542339459,
0.0000682495,
0.0698696747,
-0.0104743382,
-0.014020713,
0.0023973978,
0.0034006992,
-0.0205380004,
0.0211963672,
0.0117009617,
-0.0000144005,
-0.003142583,
0.0216905028,
0.1060690433,
-0.0147188408,
0.0296183955,
-0.0081401234,
-0.0421401113,
-0.0211257543,
0.0062421421,
0.0132121844,
0.0135030234,
-0.0353590287,
-0.0371842794,
-0.0498619415,
-0.069906339,
0.0519764759,
-0.0175043829,
-0.0336647183,
-0.0018742651,
0.0507501625,
0.0193402823,
0.0650598854,
-0.0085747121,
-0.0156082856,
0.013070208,
-0.0228545703,
-0.0599771142,
0.0001160646,
-5.436076931e-33,
-0.0074690534,
-0.0246317014,
-0.0068263155,
-0.0406599268,
-0.0147959339,
0.0260536913,
0.0109493863,
-0.0419810489,
-0.0242349822,
0.012305255,
-0.0074477079,
0.026214689,
-0.0019888626,
0.0208935346,
-0.011874455,
-0.0205884706,
0.011602385,
0.0255369451,
-0.0216922443,
-0.0124398554,
0.0290549677,
0.0463879034,
0.0461840034,
-0.0686615109,
0.0006667727,
0.0125797112,
0.0354264341,
0.0037183701,
0.05147646,
0.0699075758,
0.0115238698,
0.0148898093,
0.03453172,
-0.0430502519,
0.0199843831,
-0.010653289,
-0.1316494048,
-0.0063074888,
0.0031319731,
-0.0259921961,
-0.0124766855,
-0.0155751975,
0.0744480714,
-0.0042775664,
-0.0114822183,
-0.0163720883,
0.013718307,
-0.0110246316,
-0.056876149,
-0.0737328008,
-0.0257881768,
0.0575080402,
0.0443012081,
0.0271811616,
0.022343412,
-0.0695828646,
-0.0310765095,
0.0331838168,
-0.0482863486,
-0.0152162658,
-0.0039216462,
0.0349942371,
0.008984196,
-0.0387804843,
-0.0105431834,
0.0278626904,
0.0192420911,
-0.0354112312,
-0.0005914108,
0.0165033508,
-0.0234772805,
0.0082501359,
0.0244632736,
0.0475983508,
0.0351662897,
-0.0401683673,
-0.0792459473,
0.0644739419,
-0.0440069139,
0.0433750339,
-0.0159588438,
-0.0068911677,
-0.0285775904,
-0.0142971678,
-0.0523509309,
0.0450586975,
-0.0016333542,
-0.0372487456,
0.0074300808,
-0.0049280967,
-0.0195417535,
0.0426413119,
-0.0037876926,
-0.0363181718,
-0.0236151051,
-0.0240180511,
0.0587782785,
0.0314761177,
-0.0303929262,
-0.0085933795,
-0.0394539945,
-0.0133980121,
0.0407360978,
-0.0055015478,
0.0282921735,
0.0198899359,
0.0082526524,
-0.0521649309,
-0.0627209321,
0.015875008,
0.0407923907,
-0.067881003,
-0.0447032675,
0.0108091291,
0.0105404379,
0.0491763987,
0.0368254669,
0.0016434757,
0.0461145751,
0.0006727577,
-0.0339576118,
0.0200932082,
0.0389362797,
-0.0070410324,
-0.0202781316,
0.0085788332,
-0.0337951444,
0.0531463362,
0.0355908796,
-0.0494887829,
-0.0151265217,
-0.0335098095,
2.831e-7,
0.0871219411,
0.0490957163,
-0.0137986699,
-0.0125065269,
0.0227283332,
-0.0262827091,
0.0799444169,
0.0345194265,
-0.0223990921,
-0.0142391156,
0.0055328929,
-0.0247594155,
-0.0017432701,
0.0473049171,
-0.0682071075,
-0.0487309955,
-0.0546114445,
-0.0503784753,
0.0073860516,
-0.0330209509,
0.020975953,
0.0611395836,
-0.0285043698,
0.0038951442,
-0.0071819769,
-0.0025944647,
0.0285694562,
-0.0338889323,
0.033628583,
0.0405845828,
0.0228917543,
0.0154399155,
0.0131920762,
0.0567516536,
-0.0104411952,
0.0045797215,
0.029009711,
0.0670188218,
-0.0467760712,
0.0146336714,
-0.005590776,
-0.0818100423,
0.0120283403,
-0.0439353734,
0.0482364818,
-0.0670601875,
-0.0031507099,
0.009330322,
-0.0035762,
0.0382187404,
0.0437122025,
0.0087244306,
-0.0580204166,
0.0596927032,
0.0336490385,
0.0445265807,
-0.0187738705,
-0.0866940394,
0.0213751774,
0.0605729297,
0.0119541911,
-0.0352030657,
0.0481478721,
-0.0525610894,
-0.0053066337,
-0.0477268621,
0.0139333615,
2.388437839e-34,
-0.0013992036,
0.0214724969,
-0.0052668285,
0.0208879523,
0.0169872995,
-0.0568801053,
-0.035031382,
0.0013083465,
-0.0066692624,
-0.0717297122,
0.0201260969
] |
http://arxiv.org/abs/2405.15130v1
|
http://arxiv.org/pdf/2405.15130v1
|
OptLLM: Optimal Assignment of Queries to Large Language Models
| 2024-05-24T01:05:37 | 2024-05-24T01:05:37 |
Yueyue Liu, Hongyu Zhang, Yuantian Miao, Van-Hoang Le, Zhiqiang Li
|
cs.SE
|
cs.SE, cs.CL, cs.LG
|
Large Language Models (LLMs) have garnered considerable attention owing to
their remarkable capabilities, leading to an increasing number of companies
offering LLMs as services. Different LLMs achieve different performance at
different costs. A challenge for users lies in choosing the LLMs that best fit
their needs, balancing cost and performance. In this paper, we propose a
framework for addressing the cost-effective query allocation problem for LLMs.
Given a set of input queries and candidate LLMs, our framework, named OptLLM,
provides users with a range of optimal solutions to choose from, aligning with
their budget constraints and performance preferences, including options for
maximizing accuracy and minimizing cost. OptLLM predicts the performance of
candidate LLMs on each query using a multi-label classification model with
uncertainty estimation and then iteratively generates a set of non-dominated
solutions by destructing and reconstructing the current solution. To evaluate
the effectiveness of OptLLM, we conduct extensive experiments on various types
of tasks, including text classification, question answering, sentiment
analysis, reasoning, and log parsing. Our experimental results demonstrate that
OptLLM substantially reduces costs by 2.40% to 49.18% while achieving the same
accuracy as the best LLM. Compared to other multi-objective optimization
algorithms, OptLLM improves accuracy by 2.94% to 69.05% at the same cost or
saves costs by 8.79% and 95.87% while maintaining the highest attainable
accuracy.
|
OptLLM: Optimal Assignment of Queries to Large Language Models: Large Language Models (LLMs) have garnered considerable attention owing to
their remarkable capabilities, leading to an increasing number of companies
offering LLMs as services. Different LLMs achieve different performance at
different costs. A challenge for users lies in choosing the LLMs that best fit
their needs, balancing cost and performance. In this paper, we propose a
framework for addressing the cost-effective query allocation problem for LLMs.
Given a set of input queries and candidate LLMs, our framework, named OptLLM,
provides users with a range of optimal solutions to choose from, aligning with
their budget constraints and performance preferences, including options for
maximizing accuracy and minimizing cost. OptLLM predicts the performance of
candidate LLMs on each query using a multi-label classification model with
uncertainty estimation and then iteratively generates a set of non-dominated
solutions by destructing and reconstructing the current solution. To evaluate
the effectiveness of OptLLM, we conduct extensive experiments on various types
of tasks, including text classification, question answering, sentiment
analysis, reasoning, and log parsing. Our experimental results demonstrate that
OptLLM substantially reduces costs by 2.40% to 49.18% while achieving the same
accuracy as the best LLM. Compared to other multi-objective optimization
algorithms, OptLLM improves accuracy by 2.94% to 69.05% at the same cost or
saves costs by 8.79% and 95.87% while maintaining the highest attainable
accuracy.
|
[
0.0613445379,
0.0775525048,
0.0211224668,
0.0265486855,
-0.0444329455,
-0.0038468975,
0.0112956883,
-0.0112061976,
0.0163667779,
-0.0994414315,
-0.0359356031,
-0.0304132253,
-0.0421311893,
-0.0127435708,
-0.0186112579,
-0.0395333171,
0.031407129,
-0.0395396836,
-0.006121506,
-0.01519306,
0.0093888873,
0.0307221115,
0.0099576535,
0.0002878276,
-0.0110302856,
0.0087885149,
0.0093647875,
0.0107646948,
-0.0616463162,
-0.0050941599,
0.0208840724,
0.0491246395,
-0.0097668134,
0.0059511485,
0.000001949,
-0.0775789768,
-0.0817498937,
-0.0081698578,
-0.0208118651,
0.005708822,
0.0944323391,
0.0126359947,
0.0424866378,
0.0651292726,
-0.0211047344,
0.0378494002,
0.0466000661,
0.0076784962,
0.0673221052,
0.0635901839,
-0.0022386627,
-0.0327209979,
-0.0416505449,
-0.0042286371,
0.0107268356,
-0.0922383294,
0.0165361557,
-0.0077107251,
0.0129620871,
0.0140911713,
0.0009632878,
0.0426294021,
-0.0175690353,
-0.015006111,
0.013675347,
-0.0235062744,
-0.0375474431,
-0.0040736939,
-0.0073444401,
0.038933441,
-0.0203170851,
-0.0031525036,
-0.0050259549,
0.008201275,
-0.0183654707,
-0.0055015008,
-0.0397482701,
0.0063892314,
0.0171216968,
0.0102623012,
-0.0045599863,
-0.013807361,
0.017004285,
0.0361936428,
-0.0264137238,
0.1011077017,
-0.0168793984,
-0.0091880662,
0.0139495805,
0.0217827111,
-0.0103177195,
-0.0381868146,
0.0204384029,
0.0528987572,
-0.080634594,
0.0560551398,
-0.0348822325,
-0.0113043115,
0.0152022811,
-0.0535472855,
0.0107433908,
-0.0010561674,
0.0446884409,
0.083500281,
0.0064308611,
0.0160853267,
-0.0021214155,
-0.0082077542,
-0.0838082805,
-0.0750553906,
-0.0554555804,
-0.0219205432,
-0.0229648035,
0.0658608973,
-0.0168355834,
-0.0214706361,
-0.0506619401,
0.0353892297,
0.017095061,
-0.0218501631,
0.0455731191,
0.0339913145,
0.0170297362,
-0.0120455306,
-0.0556829795,
0.0171017181,
0.0045863278,
0.0317634158,
-0.0402252823,
-0.0321933702,
0.0132197356,
0.0449697189,
-0.0036502792,
-0.0250780303,
0.0230270699,
0.0529053546,
-0.0202341732,
-0.0260260068,
-0.0543006286,
-0.0469600298,
-0.0048522507,
-0.0085055484,
-0.0484584495,
-0.05828977,
0.0236007776,
0.0209006406,
0.003239349,
-0.0279747415,
0.0072338777,
0.0125469621,
-0.0376896635,
0.0515523516,
0.0220438577,
0.0479750335,
0.0576590151,
-0.0057310546,
0.0145309251,
0.110946238,
0.0030701293,
0.0290501062,
0.0287662018,
0.0170992147,
-0.0066045504,
0.0187085997,
-0.0111297416,
0.0034962923,
-0.0293709226,
0.0269098561,
-0.0034288017,
0.030968463,
0.0056310981,
0.0540420189,
0.0089112893,
0.0022746022,
0.0133185778,
0.024134323,
0.0159241483,
-0.0225891303,
-0.0339861922,
0.0062020947,
-0.0261150803,
-0.074468188,
-0.0181404613,
0.0417922698,
-0.0278890338,
-0.0202679895,
0.0476335511,
-0.0162351541,
-0.0333669335,
-0.0487667248,
-0.0413958505,
-0.0132725947,
-0.0629340932,
0.0536423959,
0.0533927791,
-0.0500613973,
-0.0379708409,
0.0776836425,
0.0076160878,
0.0023468772,
0.0246958416,
-0.0528916307,
-0.0177871101,
0.0634447336,
0.0177511834,
-0.0726705641,
-0.003930009,
-0.0053764028,
0.0196000505,
0.0084947199,
0.0787604675,
-0.0100375125,
-0.0135221044,
-0.0347055495,
-0.029447984,
0.0200782344,
0.0147788087,
0.0260575376,
0.0573064424,
0.0224970635,
0.0060697058,
-0.049370341,
-0.0224011671,
0.024841208,
0.0048155049,
-0.0529383756,
0.03846699,
-0.0120643815,
0.0417962074,
-0.0032016381,
0.0045312759,
0.017237125,
0.0195688047,
-0.0203045532,
-0.0608479269,
-0.0803448781,
-0.0022629565,
0.0074422536,
0.043844644,
0.0109146666,
-0.0101895304,
0.018853616,
-0.0521213859,
-0.0496730544,
-0.0086010573,
-0.03212272,
0.029737791,
-0.0461173654,
0.0103466334,
-0.0097572915,
-0.0220028348,
-0.0835604891,
0.0495316386,
-0.0014841136,
0.0340291783,
-0.0050940588,
-0.0420199372,
0.002763564,
-0.0213195533,
-0.0180623624,
0.0067065675,
-0.0140549354,
0.0663617104,
0.0115541788,
0.0132407919,
-0.0366327688,
-0.0350881331,
-0.0916348845,
0.0304698944,
0.0116165737,
0.0277911816,
0.0111741796,
0.0112263421,
0.0176527295,
0.0027312844,
-0.0522877611,
-0.0212734379,
0.0178803168,
0.0290326979,
0.0209564678,
0.0171616189,
-0.0028194652,
-0.0097665517,
-0.0166094359,
-0.020992713,
0.0325380601,
-0.0051629329,
-0.0121749137,
-0.0307247732,
0.0192111358,
-0.0451585762,
0.0099438289,
0.0222370885,
0.0066814092,
-0.0601962879,
-0.0525271222,
-0.0144815454,
0.055870261,
0.0213334057,
-0.0109589593,
0.0378084667,
0.0055556386,
-0.0676263422,
-0.0004066961,
-0.0339488797,
-0.0492259711,
-0.0230231024,
-0.0479588658,
0.0457251668,
0.0316998288,
-0.0202136394,
-0.0048592989,
-0.034133561,
0.0084028998,
-0.0195516404,
-0.0220286809,
-0.0878721923,
0.04084602,
0.015518235,
0.0429145396,
0.020591585,
-0.0237907283,
0.0636872873,
0.0079574846,
0.0100815371,
-0.0591061115,
-0.0379556976,
-0.0759988949,
0.0300993472,
-0.0227228347,
0.0613193028,
-0.0163653884,
-0.0202501155,
-0.0467213914,
-0.0322577767,
0.0437315814,
-0.0544442162,
0.0486645922,
-0.0159957204,
0.0271297079,
-0.0025497994,
-0.0290403813,
-0.0119433217,
-0.0506758615,
-0.0590303876,
0.0004380326,
0.0227410533,
0.0111701265,
-0.0745429248,
0.0288908258,
-0.0103860274,
0.0085500339,
-0.0157520659,
-0.0135154659,
0.0166137107,
0.0003067631,
0.0573161468,
0.0336257666,
0.0146226324,
-0.0092389584,
0.0006855056,
0.0391375385,
0.0582954809,
-0.0263291579,
0.0209237523,
0.0031651095,
0.0576818772,
-0.056358207,
-0.0113322334,
-0.0005994138,
0.0209320243,
-0.0134316292,
0.0181362424,
0.1248853281,
0.0486011244,
-0.0421285778,
-0.0293789823,
-0.0058873584,
0.021742044,
0.0558329672,
0.0560233779,
-0.0227893386,
-0.049582161,
0.0065147313,
0.0681569576,
-0.0018418209,
-0.0270104408,
-0.0216521341,
-0.1361536086,
-0.0046475474,
-0.0322204791,
-0.0181330368,
0.0174792539,
-0.046870131,
-0.0169045608,
-0.0567859001,
-0.0029038077,
-0.0041213441,
0.0034373365,
-0.0152291674,
0.0379018113,
-0.0102579854,
-0.020836683,
0.0179862436,
0.0005341574,
0.0221616849,
0.0179061629,
0.007730545,
-0.0092799431,
0.0050792042,
0.0681109652,
-0.0086670183,
-0.0315935202,
-0.0390478075,
0.0351210572,
-0.0087385867,
0.011358941,
0.0237180367,
-0.0052973418,
-0.0691909418,
0.0307298284,
0.0408381671,
0.0938322023,
0.002902207,
0.0292291418,
-0.0134847574,
-0.0462806486,
-0.0136472415,
0.091340594,
-0.0288229529,
-0.0652282014,
-0.02659945,
-0.0385171361,
-0.0358627811,
-0.0072093438,
0.0011784,
0.0349183232,
-0.024698237,
-0.0282785222,
-0.0408289023,
0.0164354071,
0.0363159403,
0.042432826,
-0.0077633508,
0.0399151817,
0.0096831918,
0.0048920568,
-0.0337620266,
-0.0116701331,
-0.0492790118,
0.0292049851,
-0.0798295289,
0.0180510674,
0.0631442741,
-0.0320616141,
-0.0355610549,
-0.0161940977,
-0.0570018291,
-0.0632955432,
-0.0227393098,
0.0438620634,
0.0239104033,
0.0110764503,
-0.0252487138,
-0.066701144,
0.0139926989,
0.0887935013,
-0.0200129952,
-0.0029136373,
0.0003511322,
0.0261800103,
-0.014777218,
0.0347023681,
-0.0672711283,
-0.0090177907,
-0.0301649533,
-0.0037160257,
-0.0408349559,
-0.0706968904,
0.0185432434,
0.0652712882,
0.0103344237,
0.0255318061,
-0.0117386803,
0.0144517599,
0.0326982997,
-0.0035411806,
-0.0700845122,
-0.0152774798,
0.0051154108,
0.0461051166,
-0.0237367414,
-0.0054900274,
0.0250818878,
0.0227703694,
-0.0092032114,
-0.0059009306,
0.0009724975,
0.0092614433,
-0.0052262121,
0.0330465697,
0.0071138088,
-0.0020997117,
0.0035635605,
0.0400218889,
0.0433341078,
-0.0176288188,
-0.0231472757,
0.0252691153,
0.0098732822,
-0.0300028231,
-0.011899502,
-0.0482120179,
-0.091357708,
-0.0180772524,
-0.017337773,
0.0213160664,
-0.0051368908,
0.0379179567,
-0.0242186859,
-0.0328922719,
-0.0259252489,
0.002032747,
0.026344806,
0.0120311175,
0.046873793,
0.0183327813,
-0.0162958093,
0.0064106896,
0.0806276575,
0.0203940626,
0.0289042797,
-0.0275294818,
-0.0773638636,
-0.0496863201,
-0.0201934315,
0.0071039936,
0.0218499359,
-0.016372893,
0.0035077471,
-0.0254870523,
-0.0337379575,
0.0331846587,
-0.0557466708,
-0.0167069901,
-0.0152989486,
0.0729017183,
0.0922802761,
0.0255289003,
0.0104266424,
-0.0642289072,
-0.0017600192,
-0.0367403105,
-0.0423230715,
0.034034308,
-6.36667283e-33,
0.0096669449,
-0.0253930539,
0.0147872847,
-0.0270409547,
0.0339132547,
0.0016911653,
0.0105230222,
0.0292843189,
-0.0196182299,
-0.0417801812,
-0.0003750829,
0.0085871117,
0.0079154279,
0.0453714542,
0.0277118478,
-0.0163510554,
-0.0006030509,
-0.0070987055,
-0.0225568898,
-0.0177975614,
0.0301006772,
0.0259212125,
0.0233148504,
-0.055763036,
-0.0170448944,
0.0378742367,
-0.0201846324,
0.0637187138,
0.0160881411,
0.057759475,
0.0083792303,
0.0586960763,
0.0171037782,
-0.0472275317,
0.0103577217,
-0.02503011,
-0.0194918364,
-0.0351066031,
0.0041426681,
0.0135940555,
0.0087365098,
0.0016373616,
0.0266608726,
0.0243350882,
0.0149708949,
0.0045815487,
0.0382489897,
-0.009419661,
-0.016992718,
-0.0170228854,
0.016163148,
0.0163321774,
0.0015461827,
-0.0060014427,
0.047016874,
0.0520037711,
-0.0209349692,
0.0581129715,
-0.0040581734,
0.0196512192,
-0.0302118454,
0.0650751665,
0.0639759675,
0.0522954166,
-0.0268889144,
0.0209209584,
-0.0197977182,
-0.0116838207,
0.0007642932,
-0.0191432983,
0.0081572896,
0.0281228982,
0.0283434112,
0.0354403853,
-0.0163569096,
0.0021249687,
-0.0564894304,
0.0261492096,
0.0236626118,
-0.0099456254,
0.0503470376,
0.0387414247,
-0.0032162617,
-0.0595307425,
-0.054990191,
0.0241035745,
-0.0631095394,
0.0198921449,
0.0256145578,
0.0376489609,
-0.0324645564,
-0.0175002459,
0.0376564078,
-0.0477078073,
-0.0650967285,
0.0168818012,
0.00421086,
0.0245226193,
0.0137343565,
0.056571465,
-0.0540928207,
-0.0180129111,
-0.0263439529,
-0.0119613186,
-0.0282052383,
-0.0015222452,
0.0125661911,
0.0041038934,
-0.0677011311,
0.0097339544,
-0.0055419547,
0.0061380048,
-0.0321946628,
0.0168604013,
0.0118577639,
-0.0033206481,
0.0118126571,
0.0794361904,
0.0512846969,
0.0123303728,
-0.0167447105,
-0.0586120524,
0.005131965,
-0.0170806292,
-0.0207694471,
0.0124187069,
-0.0087504741,
-0.031932611,
0.0019946394,
-0.0192510337,
-0.0321701765,
0.0051681339,
2.701e-7,
0.1055466384,
0.0734678507,
-0.0621045046,
0.0627530962,
0.0028010849,
0.0502372235,
0.0158025529,
0.0006390117,
0.0028983636,
-0.0276441667,
0.0013414879,
0.0182471778,
0.045551762,
0.0474247932,
-0.0776729137,
-0.0323116109,
-0.0190528538,
0.0298069064,
-0.0394490883,
0.0205862317,
0.0723673999,
0.0994435251,
0.0176606402,
-0.0113688167,
-0.0450739786,
-0.0416765735,
0.029172916,
-0.0344645232,
-0.0182074886,
0.0315746926,
0.0194974244,
0.028623648,
0.0238047503,
-0.0004939557,
-0.0106281238,
-0.0015380656,
0.0274248961,
0.0757852793,
-0.0126841851,
-0.0459932312,
0.0174950734,
-0.0161048397,
0.0161142293,
-0.0456340164,
0.0911643505,
-0.0017770107,
-0.0081971651,
-0.0202368535,
0.020190658,
-0.0235341098,
0.0254604742,
-0.0198371224,
-0.0206344202,
0.0116724083,
0.0108533911,
-0.0107966531,
-0.0280909184,
-0.0648364201,
-0.0075310087,
0.0428646244,
-0.0479901955,
-0.0140267676,
-0.0198259838,
0.0591984428,
0.0150422072,
-0.0438852832,
0.0177746117,
2.541888525e-34,
0.044151701,
0.0229608882,
0.0078787431,
0.0409753136,
0.0233455636,
0.0359093621,
-0.0560138375,
0.0060572182,
0.0199427791,
-0.0728410184,
0.0210968461
] |
http://arxiv.org/abs/2405.15122v1
|
http://arxiv.org/pdf/2405.15122v1
|
Generalizable and Scalable Multistage Biomedical Concept Normalization Leveraging Large Language Models
| 2024-05-24T00:31:04 | 2024-05-24T00:31:04 |
Nicholas J Dobbins
|
cs.CL
|
cs.CL
|
Background: Biomedical entity normalization is critical to biomedical
research because the richness of free-text clinical data, such as progress
notes, can often be fully leveraged only after translating words and phrases
into structured and coded representations suitable for analysis. Large Language
Models (LLMs), in turn, have shown great potential and high performance in a
variety of natural language processing (NLP) tasks, but their application for
normalization remains understudied.
Methods: We applied both proprietary and open-source LLMs in combination with
several rule-based normalization systems commonly used in biomedical research.
We used a two-step LLM integration approach, (1) using an LLM to generate
alternative phrasings of a source utterance, and (2) to prune candidate UMLS
concepts, using a variety of prompting methods. We measure results by
$F_{\beta}$, where we favor recall over precision, and F1.
Results: We evaluated a total of 5,523 concept terms and text contexts from a
publicly available dataset of human-annotated biomedical abstracts.
Incorporating GPT-3.5-turbo increased overall $F_{\beta}$ and F1 in
normalization systems +9.5 and +7.3 (MetaMapLite), +13.9 and +10.9 (QuickUMLS),
and +10.5 and +10.3 (BM25), while the open-source Vicuna model achieved +10.8
and +12.2 (MetaMapLite), +14.7 and +15 (QuickUMLS), and +15.6 and +18.7 (BM25).
Conclusions: Existing general-purpose LLMs, both propriety and open-source,
can be leveraged at scale to greatly improve normalization performance using
existing tools, with no fine-tuning.
|
Generalizable and Scalable Multistage Biomedical Concept Normalization Leveraging Large Language Models: Background: Biomedical entity normalization is critical to biomedical
research because the richness of free-text clinical data, such as progress
notes, can often be fully leveraged only after translating words and phrases
into structured and coded representations suitable for analysis. Large Language
Models (LLMs), in turn, have shown great potential and high performance in a
variety of natural language processing (NLP) tasks, but their application for
normalization remains understudied.
Methods: We applied both proprietary and open-source LLMs in combination with
several rule-based normalization systems commonly used in biomedical research.
We used a two-step LLM integration approach, (1) using an LLM to generate
alternative phrasings of a source utterance, and (2) to prune candidate UMLS
concepts, using a variety of prompting methods. We measure results by
$F_{\beta}$, where we favor recall over precision, and F1.
Results: We evaluated a total of 5,523 concept terms and text contexts from a
publicly available dataset of human-annotated biomedical abstracts.
Incorporating GPT-3.5-turbo increased overall $F_{\beta}$ and F1 in
normalization systems +9.5 and +7.3 (MetaMapLite), +13.9 and +10.9 (QuickUMLS),
and +10.5 and +10.3 (BM25), while the open-source Vicuna model achieved +10.8
and +12.2 (MetaMapLite), +14.7 and +15 (QuickUMLS), and +15.6 and +18.7 (BM25).
Conclusions: Existing general-purpose LLMs, both propriety and open-source,
can be leveraged at scale to greatly improve normalization performance using
existing tools, with no fine-tuning.
|
[
0.0226822104,
0.0350596085,
0.0256759021,
-0.0552649945,
-0.0249489285,
0.0179535262,
-0.005421286,
0.0029934577,
-0.0407980494,
-0.0175692737,
-0.0132042291,
-0.0610987544,
0.0276005417,
0.051248841,
-0.0180491358,
-0.0363919139,
0.0346962437,
0.0188189168,
0.0077722282,
0.0120045282,
0.0069440017,
0.0072352248,
0.0292206164,
0.0573628843,
0.0434592552,
-0.0156289116,
0.0479149446,
-0.0161510985,
0.011131471,
-0.0604679845,
-0.0197789948,
0.0585634746,
0.0055548842,
0.0102828145,
0.0000022248,
-0.0683523864,
-0.0568094142,
0.0272776578,
0.0044149286,
0.0293460302,
0.0748107657,
-0.0778900236,
0.0043143639,
0.0197799485,
-0.0265268832,
-0.0401820913,
0.0341698267,
0.0445464179,
0.0195748042,
0.0456616022,
-0.0256774127,
-0.0390767381,
0.0375555195,
-0.015716644,
-0.0000497911,
-0.0426693074,
0.0489745513,
0.0146702165,
0.0051200679,
-0.0074006319,
0.0219811723,
0.0530382022,
-0.0328627937,
0.0244074482,
0.0472702123,
0.0544833094,
-0.0044009835,
-0.0110595049,
0.0125067327,
0.0408555679,
0.0034754882,
0.0006515904,
0.0366874114,
0.0072764703,
-0.0234111231,
0.0068432917,
-0.047357589,
0.0219910108,
-0.0306186415,
0.0132707348,
0.0263146516,
-0.0163041353,
0.0175205134,
-0.0039911652,
-0.0311877262,
0.0085186493,
0.0256436747,
-0.0173732098,
0.0193532538,
-0.0244543459,
-0.0429383069,
-0.0302506089,
0.0750936717,
0.0044334708,
-0.0203122031,
0.0054829302,
0.0086689489,
-0.0493208803,
0.0767273754,
-0.0698349252,
0.0027095207,
0.0045923484,
-0.0172838606,
0.0266991779,
0.0014145001,
-0.0016209038,
-0.0146771576,
0.0249592364,
-0.0573602132,
-0.0009501064,
-0.0409091525,
0.003100557,
-0.0130499257,
0.0540962778,
-0.021124037,
-0.0014402305,
-0.1005476266,
-0.0173450746,
0.0044617243,
-0.0112967184,
0.0348313972,
0.0277607776,
-0.0080218399,
-0.017778486,
0.0007721529,
0.0529186055,
0.0041681859,
-0.0028497595,
-0.0492822081,
-0.0647110716,
0.0175666343,
-0.0119364718,
-0.0036540844,
-0.0057732109,
0.0406938642,
0.0271732919,
0.0364972241,
-0.0513865575,
-0.0259257369,
-0.0293031149,
-0.0416711606,
-0.0191744287,
-0.0409048609,
-0.0708350092,
0.0192201193,
0.0164259952,
0.0099910926,
-0.0424662046,
0.0013471787,
0.0318739712,
-0.0549585074,
0.0750192627,
-0.027683299,
0.0186308045,
0.0333340354,
0.0368725434,
0.0132053755,
0.0090043591,
0.0201465432,
-0.0556511134,
-0.0116986316,
-0.0297970995,
-0.0159802716,
0.02243164,
0.0003646126,
-0.0604038425,
-0.0118726771,
-0.0019214465,
0.0133904349,
0.0764198676,
0.0345243141,
0.0729308873,
0.0043341867,
-0.0289912801,
0.0182966758,
0.0966439769,
-0.0243193097,
0.013907155,
0.0011209757,
-0.0213158522,
0.0091884891,
-0.0495409556,
-0.0080927527,
-0.013597996,
-0.0010045316,
-0.0128681213,
0.0156310685,
0.0362254642,
-0.0266913287,
-0.0131982751,
-0.0104472796,
-0.0199805312,
-0.0150484964,
-0.0374997891,
0.0352286212,
-0.0350186676,
-0.01487194,
0.0523454845,
0.0316451862,
-0.0076642539,
0.0240731221,
-0.0145578161,
-0.050754033,
0.0463813096,
0.06550318,
-0.0253499169,
-0.0250616278,
-0.0442791767,
-0.0133384308,
0.0226004068,
-0.0229108837,
-0.0093780356,
-0.0095408885,
-0.0561853088,
0.0016371728,
0.0515630692,
0.0030833113,
-0.0193959754,
-0.00639235,
0.0050099557,
-0.0067904363,
-0.0372213535,
-0.0224994738,
0.0471508801,
-0.0431030914,
-0.019548215,
-0.0001586354,
-0.0446802042,
0.0273281317,
0.0445169359,
-0.0102164969,
0.035924986,
0.0430443361,
-0.0412998758,
-0.0319757611,
-0.0144756362,
-0.0327141359,
-0.000540905,
0.0626158118,
-0.0149142742,
0.0074372021,
-0.0127767455,
-0.0247901361,
-0.0404068604,
-0.0348798633,
-0.0115827639,
0.0074574407,
-0.0303773265,
0.0444336124,
0.0147409448,
0.0268710386,
-0.0548683926,
-0.0020149308,
-0.0438948944,
0.0103690466,
0.0203686133,
-0.1103814319,
0.0146768168,
-0.0145691521,
-0.0236789621,
0.0139481304,
-0.0045048543,
0.0090167262,
0.0003572221,
0.0021484005,
-0.0437854454,
0.0057008909,
-0.103406325,
0.0412067473,
0.044966612,
0.0455572493,
0.0407309979,
0.0578330234,
0.0038099522,
-0.0005279732,
-0.0346272066,
-0.0477731042,
0.0444386341,
0.0162331201,
0.0343179479,
0.0088047311,
-0.0166982058,
-0.0169317406,
0.0196473505,
-0.0313624963,
0.0360737778,
-0.0165407788,
-0.0347091518,
-0.0047723525,
0.0454872213,
-0.0207021013,
0.0005684764,
0.0273250956,
-0.0053996951,
-0.0045079761,
0.0187861994,
0.0448870026,
0.0292250682,
0.0160255637,
0.0068580885,
0.050310716,
0.0050414354,
-0.0372354873,
-0.0217590891,
-0.011152721,
-0.0420463234,
-0.0102670724,
-0.0379748717,
0.0109095331,
0.0658594072,
0.0398547165,
0.0172074661,
-0.0365946852,
0.0028177632,
-0.0330546722,
0.0127761466,
-0.0429382138,
0.0151261799,
-0.0480769239,
0.0134130288,
-0.0025880972,
-0.0230544787,
0.0873880163,
-0.0208970513,
0.0099836392,
-0.0544913039,
0.0007371656,
-0.0191143416,
-0.0093903793,
-0.0159950946,
0.0340356417,
0.0003142911,
-0.0110188676,
0.0258220788,
0.0089286063,
-0.0177210737,
-0.0760689974,
0.0836108848,
-0.0497625247,
0.024962537,
-0.001031454,
-0.0150442496,
-0.0303657111,
0.0302916523,
-0.025907496,
0.029254552,
0.0213317852,
0.0029673446,
-0.0510452092,
0.0611934736,
0.0083377538,
0.0534464084,
0.013512318,
-0.0281150583,
0.0093525462,
0.0087811993,
0.0170127898,
-0.0086525353,
-0.0042571938,
0.0126609392,
0.0576221719,
0.0602175593,
0.0128484834,
0.0126345474,
0.0124118896,
-0.0116002532,
0.0113669625,
-0.0829990804,
-0.0076181977,
0.006095978,
0.0312370006,
-0.033028245,
0.0023868182,
0.1439451575,
0.0493245497,
-0.0868481472,
-0.0017546532,
-0.0158988498,
-0.0183254238,
-0.0269528218,
-0.0161459837,
-0.0755100772,
-0.0212850999,
-0.0231640637,
0.0367644653,
-0.0080249263,
-0.0558212698,
0.0140361404,
-0.0361111388,
0.0121591445,
0.0295432452,
-0.0756120011,
-0.0355359241,
-0.0599301718,
0.0077227741,
-0.0337893367,
0.0379786901,
0.0190261584,
-0.0127333552,
0.0030681603,
0.0155095533,
-0.0696816966,
-0.1096645072,
0.0379838571,
-0.0603999086,
0.0961559862,
0.0490962826,
0.0106276553,
0.0075582117,
0.0017341024,
0.0114209838,
-0.0200259257,
-0.0230056718,
-0.0244549364,
0.0366946831,
-0.0004904532,
0.0065272423,
-0.077157028,
0.0275773313,
-0.0311799347,
-0.0321203247,
-0.0256802123,
0.0940976217,
-0.0269200336,
-0.0130165666,
-0.0555506088,
-0.0181994289,
-0.0239364598,
0.0891479403,
-0.004030705,
-0.0312051848,
-0.0105019845,
0.0209296178,
0.0301885754,
0.0045167571,
-0.0285455678,
0.0352541469,
-0.0618179701,
-0.016934216,
-0.0421099775,
-0.0216988269,
0.0640164465,
-0.0495829731,
0.0100446073,
0.0879641697,
0.0337266214,
-0.0072527844,
-0.0823008195,
0.0489136837,
-0.0062193442,
0.0036583082,
-0.0380407721,
-0.0365694948,
0.0444754586,
0.021823721,
-0.0218629129,
0.0342692025,
0.0048220404,
-0.0186344758,
-0.0077132988,
-0.001401726,
0.0191701446,
-0.025984766,
0.0102900732,
-0.0349540859,
0.0592680834,
0.041670531,
-0.0317486562,
0.0010392253,
-0.0286702309,
0.0435811467,
-0.0139208399,
0.0185376052,
0.0039173057,
0.046447482,
-0.0323645696,
0.0741489232,
0.0013355691,
-0.0550922081,
0.0423076078,
0.0665604547,
0.0416913554,
-0.0443518348,
-0.0186164416,
0.0032051967,
-0.0398693644,
-0.0306164008,
0.0347682983,
-0.0189778451,
0.0280364044,
0.076135695,
0.0181320943,
-0.0460529104,
-0.0076450417,
-0.0864790156,
-0.0332736,
-0.0113537991,
0.0327348448,
0.0030179592,
0.013668214,
0.0461351573,
-0.0260049962,
0.0007921269,
-0.0014962761,
-0.0302991383,
-0.0133644855,
-0.0426012874,
0.0181012452,
-0.0185124949,
0.0578313135,
-0.0129114725,
0.0542320944,
-0.0270669796,
-0.0009576985,
-0.0130197313,
0.0238222107,
0.040610224,
0.0294023678,
0.0789202005,
-0.0015044237,
0.0142056793,
-0.0521057434,
0.013427726,
-0.0351085328,
0.0349087901,
0.034792725,
-0.0044729449,
-0.0167127755,
-0.0359261595,
0.0831924453,
-0.033732973,
0.058189895,
-0.0174061991,
-0.064654924,
-0.0288471747,
-0.0175336096,
-0.0076441225,
0.016579628,
-0.0773153752,
-0.0160232298,
-0.0909066126,
-0.0110204909,
0.0305111222,
0.0016563244,
-0.0131682046,
-0.0151941925,
0.0357790813,
0.0136705916,
0.0006484844,
-0.0409638062,
-0.0005692704,
0.0294138361,
-0.0421532691,
-0.0709323436,
0.0211659409,
-6.277959741e-33,
-0.0206862278,
-0.0219440814,
0.0224131159,
-0.0039878921,
-0.0299048908,
0.0189022068,
-0.0135208918,
-0.0021778017,
0.0232009757,
0.0171774495,
-0.0173976291,
-0.0133454949,
0.004206453,
-0.0047777207,
0.0285456385,
-0.0011968578,
0.0281229801,
-0.0173893608,
-0.0115259606,
0.01824365,
-0.0029853405,
0.0706825182,
0.0827036425,
-0.0717682242,
0.027493082,
0.0440794975,
-0.0062702149,
0.005741871,
0.0381970704,
0.0365704596,
-0.0652182251,
0.0486913361,
0.0404092446,
-0.0217828825,
0.0014625216,
-0.0132787647,
-0.083436504,
0.0135517213,
0.0575016402,
-0.051454138,
0.0095725758,
-0.0363432877,
0.077013813,
-0.0048037921,
-0.0479296967,
0.0222972445,
0.0132696498,
-0.034356907,
-0.0144819589,
-0.029655708,
-0.035941001,
0.0515978858,
0.0128628779,
0.016246438,
0.0224357694,
-0.0360893682,
-0.0648096278,
0.0264517274,
-0.012037375,
-0.0061514219,
0.0422336608,
0.0269089695,
0.0051150951,
-0.0086555639,
0.0039935191,
0.039283257,
0.0287387427,
-0.0317055658,
-0.0306521636,
0.0083684148,
0.0076718037,
0.0098800845,
0.0321170352,
0.0238345806,
0.0291584153,
0.0154258199,
-0.0675149038,
0.0744633749,
-0.0334084369,
0.0241662096,
0.0228596572,
-0.0133029483,
-0.0134623069,
-0.0105248718,
-0.0447952822,
0.0036907594,
-0.0192792788,
-0.0382840075,
0.0366715342,
-0.0201672539,
-0.0028156724,
0.0129436105,
0.0168660022,
-0.0136804283,
-0.0261486899,
-0.0091729593,
0.0567372814,
0.0131669808,
-0.0252655819,
0.0012643186,
-0.0607073866,
-0.0191367026,
0.0308577958,
-0.0408929363,
0.0409895368,
0.0071908939,
-0.0333153121,
-0.0209746975,
-0.0609326735,
-0.0137718068,
0.0111340778,
-0.0466492586,
-0.0231532063,
0.0141802356,
0.0159434956,
0.0188723076,
0.0209473316,
0.0124507491,
0.034856569,
0.0043273596,
-0.0437899828,
-0.0304597672,
0.0428422205,
0.0363037027,
-0.013483922,
0.0394054167,
-0.0101501411,
0.0161334947,
0.058235392,
-0.0215186886,
-0.0082158437,
-0.0202241838,
2.989e-7,
0.1106169894,
0.0759658515,
0.0154222967,
-0.0114221685,
0.0057173739,
0.0028849067,
-0.0038934601,
0.0191227943,
0.0102735581,
-0.0031536184,
0.052126471,
0.0111577474,
0.0023738779,
0.0032419674,
-0.0608029924,
-0.0484662838,
-0.0599939823,
-0.0281052534,
-0.0391465537,
0.0180792212,
0.0273050778,
0.0613101758,
-0.0196350031,
-0.0389987789,
0.0244881418,
-0.0265784264,
0.0329627022,
-0.0572996959,
-0.0115447817,
0.0524267107,
0.0218806881,
-0.0055706017,
0.0336269736,
0.0411345698,
0.0383036099,
-0.0088374782,
0.0099020032,
0.1034619361,
-0.0579975545,
0.0343905054,
0.0266414378,
-0.0204481967,
0.0220350772,
-0.0471932106,
0.0186656676,
-0.0388269834,
-0.0392655954,
-0.0265742932,
-0.0101081897,
-0.0022565406,
0.0577212386,
-0.0001509416,
-0.024069747,
0.0322808176,
0.0316259898,
0.0359536856,
-0.0555293374,
-0.0660245419,
-0.0165895615,
0.0647904724,
-0.0167156961,
-0.0448947139,
0.0017350966,
-0.0118617248,
0.0000233688,
-0.0749164373,
0.0127570629,
3.008201165e-34,
0.0373007171,
-0.0273642354,
0.0060045854,
0.0419157557,
0.0129151214,
-0.0194783546,
-0.0043936716,
0.0181327835,
-0.0346621312,
-0.1290214062,
0.0313994512
] |
http://arxiv.org/abs/2405.15115v1
|
http://arxiv.org/pdf/2405.15115v1
|
Towards Better Understanding of In-Context Learning Ability from In-Context Uncertainty Quantification
| 2024-05-24T00:08:55 | 2024-05-24T00:08:55 |
Shang Liu, Zhongze Cai, Guanting Chen, Xiaocheng Li
|
cs.LG
|
cs.LG, cs.CL, stat.ML
|
Predicting simple function classes has been widely used as a testbed for
developing theory and understanding of the trained Transformer's in-context
learning (ICL) ability. In this paper, we revisit the training of Transformers
on linear regression tasks, and different from all the existing literature, we
consider a bi-objective prediction task of predicting both the conditional
expectation $\mathbb{E}[Y|X]$ and the conditional variance Var$(Y|X)$. This
additional uncertainty quantification objective provides a handle to (i) better
design out-of-distribution experiments to distinguish ICL from in-weight
learning (IWL) and (ii) make a better separation between the algorithms with
and without using the prior information of the training distribution.
Theoretically, we show that the trained Transformer reaches near Bayes-optimum,
suggesting the usage of the information of the training distribution. Our
method can be extended to other cases. Specifically, with the Transformer's
context window $S$, we prove a generalization bound of
$\tilde{\mathcal{O}}(\sqrt{\min\{S, T\}/(n T)})$ on $n$ tasks with sequences of
length $T$, providing sharper analysis compared to previous results of
$\tilde{\mathcal{O}}(\sqrt{1/n})$. Empirically, we illustrate that while the
trained Transformer behaves as the Bayes-optimal solution as a natural
consequence of supervised training in distribution, it does not necessarily
perform a Bayesian inference when facing task shifts, in contrast to the
\textit{equivalence} between these two proposed in many existing literature. We
also demonstrate the trained Transformer's ICL ability over covariates shift
and prompt-length shift and interpret them as a generalization over a meta
distribution.
|
Towards Better Understanding of In-Context Learning Ability from In-Context Uncertainty Quantification: Predicting simple function classes has been widely used as a testbed for
developing theory and understanding of the trained Transformer's in-context
learning (ICL) ability. In this paper, we revisit the training of Transformers
on linear regression tasks, and different from all the existing literature, we
consider a bi-objective prediction task of predicting both the conditional
expectation $\mathbb{E}[Y|X]$ and the conditional variance Var$(Y|X)$. This
additional uncertainty quantification objective provides a handle to (i) better
design out-of-distribution experiments to distinguish ICL from in-weight
learning (IWL) and (ii) make a better separation between the algorithms with
and without using the prior information of the training distribution.
Theoretically, we show that the trained Transformer reaches near Bayes-optimum,
suggesting the usage of the information of the training distribution. Our
method can be extended to other cases. Specifically, with the Transformer's
context window $S$, we prove a generalization bound of
$\tilde{\mathcal{O}}(\sqrt{\min\{S, T\}/(n T)})$ on $n$ tasks with sequences of
length $T$, providing sharper analysis compared to previous results of
$\tilde{\mathcal{O}}(\sqrt{1/n})$. Empirically, we illustrate that while the
trained Transformer behaves as the Bayes-optimal solution as a natural
consequence of supervised training in distribution, it does not necessarily
perform a Bayesian inference when facing task shifts, in contrast to the
\textit{equivalence} between these two proposed in many existing literature. We
also demonstrate the trained Transformer's ICL ability over covariates shift
and prompt-length shift and interpret them as a generalization over a meta
distribution.
|
[
-0.0401232876,
0.0695507377,
0.0132127274,
0.0049439063,
0.0181605164,
0.0263327658,
0.0206703953,
0.0247813091,
-0.0458191372,
0.0054735914,
-0.0222980715,
-0.0078863679,
0.0199157167,
-0.0380814299,
0.0401732884,
-0.0390405692,
-0.0403342023,
-0.0263817646,
-0.0158797335,
-0.003996795,
-0.0299930759,
-0.0693165958,
0.0128496187,
0.0273639075,
-0.0349504426,
0.0044305818,
0.0185515843,
0.0136594186,
0.0143528543,
-0.0176815316,
0.0330910012,
0.012451712,
0.0192698147,
0.0255483817,
0.0000021448,
-0.0163588766,
-0.0443891436,
-0.0193733256,
-0.0326825306,
-0.0745861158,
0.007894516,
-0.0422157049,
0.0061137755,
0.0294930581,
-0.0105709322,
-0.0095163006,
0.0790288746,
0.0399936847,
-0.0169299301,
0.0822689906,
-0.0162823051,
-0.0056122723,
0.0101534277,
-0.0556807481,
0.0459252931,
-0.040434774,
0.0045101894,
-0.0239306614,
-0.0514872447,
-0.0322695933,
-0.0417509675,
0.0333026648,
0.027485013,
0.0519667417,
0.0119554624,
0.0609967932,
-0.0145915849,
0.0354236178,
-0.0202430561,
-0.0007136312,
-0.084722124,
-0.0256452635,
-0.0274088811,
0.0024734384,
0.0368512645,
0.054441452,
-0.04971347,
0.0239718966,
-0.0211345181,
0.0366420522,
-0.0447386205,
0.0454949997,
0.0038855379,
-0.0413553976,
-0.0330727696,
0.043295633,
-0.0218119342,
-0.0228567328,
0.0092022745,
0.0257372987,
-0.0711098462,
-0.0017567758,
-0.0445506833,
0.0086573781,
-0.0325107388,
0.0105994008,
-0.0081574786,
-0.0375684835,
-0.0381473973,
-0.0112531753,
0.0265008193,
0.0196029786,
-0.0121092759,
-0.0328235179,
0.0106003024,
0.0680404902,
-0.0059648701,
-0.0015325807,
-0.035836719,
0.006199853,
-0.0543259047,
0.0221596379,
-0.019302221,
-0.0026883003,
0.0017635156,
0.0187465474,
-0.0727958456,
0.0185116902,
0.0109346844,
0.0395541638,
0.0091221817,
0.0126497867,
-0.0786067918,
0.0162316356,
-0.025922725,
0.0208751168,
0.0226973426,
0.0171992853,
-0.0014803953,
-0.0188768581,
-0.019976344,
0.0265156031,
-0.0029333381,
0.004887993,
0.0293789338,
0.085630998,
-0.0250233766,
-0.0181588922,
-0.0583999716,
-0.0153634083,
-0.0392783359,
-0.0235022027,
0.0278929695,
-0.079971537,
-0.0581756793,
0.0304829925,
0.0162730105,
-0.0271153525,
0.0141898943,
-0.0089305844,
0.0296962522,
0.0805982575,
0.057768859,
-0.0313131548,
0.0342180692,
-0.012406745,
-0.0568981394,
0.0752895549,
0.0038116907,
0.0096860193,
0.0279708151,
0.0004456227,
0.037466865,
0.0399026386,
0.0086281979,
-0.0001184284,
-0.0185886193,
-0.007143836,
0.0156746916,
0.003659352,
0.0265668519,
-0.0208488908,
-0.0111032752,
-0.0162053779,
0.0675186142,
0.0118651949,
0.0283562783,
-0.0086250454,
0.0037174374,
0.0155219091,
-0.0205323454,
0.034744624,
-0.0841494352,
0.0202874392,
-0.0264198948,
0.0126954345,
-0.0696569905,
0.0030647863,
-0.0228881128,
-0.0411761887,
0.0036805321,
-0.0482812524,
-0.040401008,
0.0166740753,
0.0783606693,
0.0148113836,
-0.0534595363,
0.0401454605,
0.0304033272,
-0.0018522041,
-0.0296874978,
-0.0594186001,
-0.0216518864,
-0.055461444,
-0.0091728866,
-0.0499594249,
0.0120549994,
-0.0397097841,
-0.0144250328,
-0.0417897664,
-0.0434825085,
-0.0012466222,
0.0199589059,
-0.0112602068,
-0.051094614,
0.0229632724,
0.0019410774,
-0.0010203761,
-0.001293518,
-0.0097075952,
0.0086438814,
-0.0681985468,
-0.0291717183,
0.0163262952,
-0.0093026897,
-0.0596084483,
0.0268864986,
0.026923446,
0.0461234748,
-0.0498053469,
-0.0227233078,
0.000366167,
0.0169389304,
-0.0003183268,
-0.0391479731,
0.0350890383,
-0.0659485459,
-0.0320934802,
-0.0009734477,
-0.0607162304,
0.0124707557,
0.0080739772,
-0.0258199852,
-0.0301844589,
0.001101486,
-0.03338743,
0.0182879586,
0.0039933673,
0.0320786983,
-0.0022480711,
0.0024692002,
-0.0532691702,
0.0428873375,
-0.0003572628,
-0.0086200209,
0.0063884272,
-0.0095770229,
0.0205063596,
0.017480297,
0.0107648689,
-0.0225398671,
-0.0278636254,
0.0085757999,
0.0321519673,
0.0224239379,
-0.0097622424,
-0.0349514075,
-0.0698724091,
-0.0077548013,
-0.0673062578,
0.0267183855,
0.0095178438,
-0.0105430642,
0.0001593469,
0.0093537206,
-0.0312042907,
0.0103964554,
-0.0171799622,
-0.0310147069,
0.0200782195,
-0.0007452858,
-0.0169984512,
0.0218172893,
-0.0013432573,
-0.0013875762,
0.0715235546,
0.0221011117,
-0.0120969098,
-0.0296617802,
0.0389374942,
0.0508470833,
0.0062782927,
-0.021314621,
0.010547136,
-0.022770457,
-0.0303489547,
0.060268838,
0.0896051675,
0.0261913016,
0.0248645134,
0.0365527645,
0.0237577725,
0.0005972984,
0.0004902211,
-0.0602327324,
-0.0688116625,
-0.0689695254,
-0.0012048417,
0.0340979658,
0.033917442,
-0.0001826173,
0.0048756083,
-0.0491327234,
-0.0010168443,
0.0121511286,
-0.0074037872,
-0.0227227937,
0.0245879833,
-0.0142084025,
0.0201621894,
0.0137176942,
-0.0340048112,
0.0203304198,
0.0012689988,
-0.0043070056,
-0.0486627892,
-0.040850468,
0.0034089291,
0.0060907621,
0.0117900763,
-0.0092353886,
0.0231588949,
0.0103291338,
-0.093772009,
0.0017956715,
0.0397804156,
-0.0021503535,
0.0341178328,
-0.0381989852,
-0.023830751,
0.0070621208,
-0.0403698571,
-0.0427254997,
-0.0398722142,
0.0133648301,
-0.0055414331,
-0.0025325723,
-0.0416502357,
-0.0407648757,
0.0162269641,
-0.0019312914,
0.0644448772,
-0.0224013384,
-0.0730431974,
0.0222731791,
0.0185653195,
0.0052240305,
-0.01120352,
0.019839704,
-0.051658418,
-0.0044868235,
0.0045769732,
0.0748978034,
-0.0071259406,
-0.0359840095,
-0.0003625971,
0.0224566255,
-0.0078960564,
-0.0127304588,
0.0766984001,
0.0632253215,
-0.0077804895,
-0.0385820344,
0.0824608132,
-0.0174893402,
-0.0587446466,
-0.0032594262,
-0.0040985215,
0.0060417382,
0.0291864146,
0.0155930333,
-0.0257632211,
-0.0396966189,
-0.0107809491,
0.0665886253,
-0.0138155287,
-0.0207479484,
0.0307047665,
-0.0606656596,
0.02569318,
0.0028454377,
-0.0426398776,
-0.0505950525,
-0.0162204839,
-0.0001942235,
-0.0660436004,
-0.0031757492,
-0.0345385186,
-0.0580563918,
0.0150958989,
0.0688779801,
-0.0264976043,
-0.0474054404,
-0.0247199591,
-0.0374420173,
-0.0303595271,
0.0655250326,
-0.0143556045,
-0.0260913782,
-0.024530597,
0.0032566364,
-0.0427789688,
0.0220621508,
0.0171141513,
0.0954138041,
-0.0210982673,
0.0096811811,
0.0186952334,
0.0792768747,
-0.0143254036,
0.0183469392,
0.042477008,
0.0578810051,
-0.0317417271,
-0.0142591251,
0.0195844732,
-0.0395188034,
0.0795279071,
0.0203906614,
-0.0224004835,
-0.0431023575,
-0.0180876646,
-0.0090381065,
0.0231737513,
0.0538131483,
0.0126802381,
0.0207628552,
-0.0118712913,
-0.0746317953,
-0.0746765956,
0.0397243388,
-0.0257156715,
0.0741486549,
0.0043266807,
-0.049212195,
0.017421728,
0.0160564501,
-0.0588309802,
-0.0043605962,
0.0397162661,
0.0110000279,
-0.0272175651,
0.0363601968,
0.0427008197,
-0.028840553,
-0.0432731397,
0.0018873251,
-0.1153362766,
-0.0241927188,
0.0207551811,
0.0296271779,
0.0235435404,
0.0749255866,
0.033595778,
0.0121462401,
0.114862442,
0.0691812262,
0.0261278208,
0.0104350746,
0.1016556993,
-0.0158247538,
0.0018026412,
0.0104553578,
-0.038752161,
0.0220909193,
0.0157612637,
-0.0708670393,
-0.0560362674,
-0.0034217173,
0.019199552,
-0.0333841853,
0.0993577018,
-0.0444020852,
-0.0402595177,
0.0171826575,
0.0247596167,
0.0439497791,
0.001690961,
-0.0100742001,
-0.0282271393,
0.0612696186,
0.0652579069,
0.0050638011,
0.0614355877,
0.0382960178,
-0.0037969421,
-0.0195741449,
0.014819406,
0.0048192465,
-0.0117453486,
0.0113041559,
0.0265932512,
0.0286537744,
-0.0329777226,
0.0754260719,
0.0230210833,
0.0240416173,
0.0754187182,
0.0403942838,
0.0360614918,
-0.0203105323,
-0.055363968,
-0.1026124954,
-0.0064660767,
-0.0335351042,
0.0029824334,
-0.0035621522,
-0.065479219,
-0.008142435,
-0.0700865239,
0.0057890662,
-0.0328777395,
0.0463250652,
-0.006487492,
0.0278474558,
0.0450859964,
-0.010958666,
0.0009109245,
-0.0238898769,
0.0696823597,
-0.005292187,
0.0333046988,
-0.0364423804,
-0.0198362507,
-0.0378734469,
-0.0068980595,
0.0549743921,
-0.0104077291,
0.0239761621,
-0.0025715879,
-0.0214198884,
-0.0194460601,
0.0423793793,
-0.0425856449,
0.0525881536,
-0.0586709827,
0.0310118999,
0.0780318305,
-0.0031205683,
0.0199495126,
0.0088907899,
-0.0415585749,
0.0199182834,
0.0111633996,
-0.0444219187,
-5.752288585e-33,
0.0142217819,
-0.0523017272,
-0.033804968,
0.0507875755,
0.045296073,
0.0224247929,
-0.0145160025,
0.0177625716,
0.031523969,
-0.0146919377,
-0.0016645278,
0.01604308,
0.021981962,
0.0373004675,
0.0216856487,
-0.0172015503,
0.0245997272,
-0.0151665211,
-0.0266278591,
-0.0091764703,
0.0583540276,
0.0141263278,
0.0426526293,
-0.049163077,
-0.0203151908,
0.0131686898,
-0.0225468073,
-0.0097537246,
0.0060305162,
0.0028186033,
0.031200774,
0.0508871824,
0.0119799217,
-0.0664315298,
-0.0002775601,
0.0000475233,
-0.0321963057,
-0.0133743826,
-0.0595864877,
0.0433245488,
0.0219272953,
0.0178189203,
0.0568284392,
-0.0031358313,
-0.0578317158,
-0.056218762,
0.0606516786,
-0.009501297,
0.0248402357,
-0.0649688616,
-0.0223962087,
0.0099139689,
-0.007562032,
-0.0327896364,
0.0281331539,
0.0975530595,
0.0100273462,
-0.0176079776,
-0.0266951639,
-0.0010354003,
-0.0269478094,
0.0198827684,
0.0359521657,
0.0682604834,
-0.0012492487,
0.0259590596,
0.0224148817,
0.0020029172,
-0.003909755,
-0.0684997961,
0.019532714,
0.042274747,
0.019030042,
0.0335248709,
0.0720034838,
-0.0370310396,
-0.0783360749,
0.0557723306,
-0.0054870211,
0.0172318388,
0.0030818807,
0.0100420294,
-0.0204875283,
-0.0382037573,
-0.0174141582,
-0.0242809542,
0.0383619852,
0.0141433328,
0.0246871263,
0.0049541155,
0.014958648,
0.0290398207,
0.016508406,
-0.0285484232,
0.0169353168,
-0.0020548841,
-0.048186034,
0.0263944454,
0.0207130276,
-0.0220534448,
-0.0479755849,
-0.0196418352,
0.013947309,
0.0513780229,
0.0298052281,
0.001844992,
0.0910783932,
0.0496743917,
-0.0376170427,
-0.016286904,
-0.0315851271,
0.0414232202,
0.0261994433,
0.0386245772,
0.0602483638,
-0.0309946332,
-0.0146207185,
0.0104702283,
-0.0322793126,
-0.0119036809,
-0.0157620981,
0.0498533063,
0.0050944821,
-0.0061511584,
-0.0414414555,
0.0039270967,
-0.0064259451,
-0.0369119011,
-0.0049883914,
-0.0262503605,
-0.027720429,
-0.0073366095,
2.844e-7,
0.0325233191,
0.0113378782,
0.0250034556,
0.0056644622,
0.0263852589,
0.042581398,
-0.0456998646,
0.0107512753,
0.0150086721,
-0.0219519921,
0.0464756377,
-0.0080899466,
0.041934289,
0.0024241728,
-0.0277396683,
-0.1108663902,
0.0020588594,
0.0410290584,
-0.0024640588,
0.0606161803,
0.0368057713,
0.0505779423,
0.0019294968,
-0.0330108516,
0.0302826688,
0.0024289591,
0.0608033985,
-0.0496752076,
0.0475149229,
-0.0141319921,
-0.0390985981,
0.0170300547,
0.021313386,
0.0568173826,
-0.0209222473,
-0.0042394348,
-0.020997677,
0.0160055067,
0.013238227,
0.034439873,
0.0186803788,
0.0534226894,
0.0340023823,
-0.0292130411,
0.0435721762,
-0.1203043833,
-0.0439839885,
-0.0238454193,
-0.0522246175,
0.0105267093,
-0.0079995636,
0.0079894634,
0.0250498559,
-0.0103291394,
-0.0036294574,
0.0315172225,
0.0000941314,
-0.0361128077,
-0.0322457552,
0.0676622614,
-0.0296714008,
-0.0495598055,
-0.0359305851,
-0.0280829035,
0.0433502421,
-0.0320789292,
-0.0103817927,
2.906590469e-34,
0.0203310642,
0.0267032105,
0.0176386163,
-0.0197106693,
-0.0000144695,
-0.046848055,
-0.004151287,
0.0154962186,
0.0639201775,
-0.0222193189,
0.0213985238
] |
http://arxiv.org/abs/2405.15110v1
|
http://arxiv.org/pdf/2405.15110v1
|
CHARP: Conversation History AwaReness Probing for Knowledge-grounded Dialogue Systems
| 2024-05-24T00:00:30 | 2024-05-24T00:00:30 |
Abbas Ghaddar, David Alfonso-Hermelo, Philippe Langlais, Mehdi Rezagholizadeh, Boxing Chen, Prasanna Parthasarathi
|
cs.CL
|
cs.CL
|
In this work, we dive deep into one of the popular knowledge-grounded
dialogue benchmarks that focus on faithfulness, FaithDial. We show that a
significant portion of the FaithDial data contains annotation artifacts, which
may bias models towards completely ignoring the conversation history. We
therefore introduce CHARP, a diagnostic test set, designed for an improved
evaluation of hallucinations in conversational model. CHARP not only measures
hallucination but also the compliance of the models to the conversation task.
Our extensive analysis reveals that models primarily exhibit poor performance
on CHARP due to their inability to effectively attend to and reason over the
conversation history. Furthermore, the evaluation methods of FaithDial fail to
capture these shortcomings, neglecting the conversational history. Our findings
indicate that there is substantial room for contribution in both dataset
creation and hallucination evaluation for knowledge-grounded dialogue, and that
CHARP can serve as a tool for monitoring the progress in this particular
research area. CHARP is publicly available at
https://huggingface.co/datasets/huawei-noah/CHARP
|
CHARP: Conversation History AwaReness Probing for Knowledge-grounded Dialogue Systems: In this work, we dive deep into one of the popular knowledge-grounded
dialogue benchmarks that focus on faithfulness, FaithDial. We show that a
significant portion of the FaithDial data contains annotation artifacts, which
may bias models towards completely ignoring the conversation history. We
therefore introduce CHARP, a diagnostic test set, designed for an improved
evaluation of hallucinations in conversational model. CHARP not only measures
hallucination but also the compliance of the models to the conversation task.
Our extensive analysis reveals that models primarily exhibit poor performance
on CHARP due to their inability to effectively attend to and reason over the
conversation history. Furthermore, the evaluation methods of FaithDial fail to
capture these shortcomings, neglecting the conversational history. Our findings
indicate that there is substantial room for contribution in both dataset
creation and hallucination evaluation for knowledge-grounded dialogue, and that
CHARP can serve as a tool for monitoring the progress in this particular
research area. CHARP is publicly available at
https://huggingface.co/datasets/huawei-noah/CHARP
|
[
0.0582313836,
0.0476454943,
-0.0239903368,
0.0239947885,
-0.0287782233,
0.0235279966,
-0.0060821017,
-0.0119720902,
0.0172335319,
-0.0289222375,
-0.0261035543,
-0.0362771079,
-0.0067709126,
0.0170969311,
-0.0317882299,
-0.031606406,
0.0434093028,
-0.0042465655,
0.0159635115,
0.0107491892,
0.0196761936,
-0.0028150424,
-0.0345378593,
0.0301660653,
-0.0516126677,
-0.0257024951,
-0.0606848001,
0.001840891,
0.0280619469,
-0.0226805098,
0.0102579491,
0.0805658996,
-0.0314762518,
0.0796589777,
0.0000019495,
-0.0441734307,
0.0232481584,
-0.054294657,
-0.0561399236,
0.0000121139,
-0.0241203979,
0.0485922731,
-0.0033606356,
-0.0071495329,
-0.0607422255,
0.0497475602,
0.0205345377,
0.1271717846,
0.0763488784,
0.0625701398,
-0.0036206597,
-0.0116201099,
0.0357411727,
0.0216574967,
0.0125060026,
-0.0223788414,
0.0052427119,
0.0700379387,
-0.0260316115,
-0.0111901071,
0.0125471326,
0.0248959642,
0.025313925,
-0.0432100669,
-0.0019289043,
-0.0015215727,
-0.0917173922,
-0.0492781959,
-0.0215950869,
0.0814456344,
0.0608761609,
-0.0302356817,
0.0272898711,
0.0054974859,
0.0187931005,
-0.0332785249,
-0.0410525165,
-0.0429728925,
-0.0313515849,
-0.021548396,
-0.0264946353,
0.0318540819,
0.0143298898,
0.0069672079,
-0.00147943,
0.0639184788,
-0.0100772837,
-0.0424207449,
0.0273403078,
-0.0065486273,
0.0191783197,
-0.0871219113,
0.0039892518,
0.011794406,
-0.0372742936,
-0.0026924452,
-0.0192818418,
0.0639414787,
0.0183388758,
-0.0824490041,
-0.0388634428,
0.0169976968,
-0.0316981301,
0.0438197218,
0.0231122952,
0.0544901341,
0.0340410657,
0.0139096919,
0.0188163221,
0.032705605,
-0.0416473709,
0.0110064438,
-0.0648494661,
-0.0181568954,
-0.0206712577,
-0.023666326,
0.0004358274,
0.0446551777,
-0.0173301473,
0.0377942398,
-0.0290848184,
0.0040638051,
0.0036570707,
-0.0013358976,
-0.0189681295,
0.0215734951,
-0.062834546,
-0.0104220472,
0.0558007769,
-0.1124028116,
-0.0017853946,
0.0042351931,
-0.0101428945,
-0.0006013915,
-0.0264436211,
0.000553599,
0.0267205816,
-0.0217632055,
-0.0036742066,
-0.0313564241,
-0.0491584726,
-0.0820424557,
-0.0217155553,
-0.0186074581,
0.0417582504,
-0.0093200617,
-0.0097880745,
-0.0053393538,
0.0382026359,
0.0170215182,
-0.0163125116,
0.0429566726,
0.0207949635,
-0.0000059453,
0.0267163329,
0.0386140794,
-0.0548777729,
0.0734909847,
-0.0032444748,
0.0059825294,
0.0500109009,
-0.0210994575,
0.0376101844,
-0.0107197426,
0.0068428554,
0.005007972,
-0.0230271723,
0.0263402164,
-0.0225295946,
0.0126579227,
-0.027763037,
0.0095116859,
-0.036193572,
0.006799798,
0.0552652888,
0.0315867849,
-0.0071723717,
-0.0037234412,
-0.0239503086,
-0.0219047666,
0.0221653823,
0.0268964637,
-0.0672101006,
0.0005708606,
-0.0021623641,
-0.0174026415,
0.0410210751,
-0.0298055727,
-0.0399827361,
-0.0392456464,
0.0017430653,
0.0064397235,
0.0171259455,
-0.0205708127,
-0.0253880527,
0.0238989685,
-0.0246719252,
0.054668162,
0.0125779202,
-0.0599703863,
-0.0358566791,
-0.0290024709,
0.039626915,
0.0730197504,
0.0485245101,
-0.011555451,
-0.0525806472,
-0.0295898225,
0.0623333231,
0.058262635,
0.0314148888,
-0.0015936253,
0.0270927958,
-0.0494777113,
-0.0116776219,
0.016458191,
0.0414292179,
-0.0170973055,
-0.0173519887,
0.0130923856,
0.0107408557,
-0.0181096587,
0.0013300111,
0.0284734555,
0.0069888895,
0.0306602474,
-0.0504516177,
-0.0173939764,
0.0253703427,
0.0204536505,
0.0071005626,
-0.0196455792,
0.034124244,
-0.003063987,
-0.0241899174,
0.0834256411,
-0.0700306967,
-0.0297276191,
0.034129642,
-0.0208044127,
0.0066129421,
-0.0245233923,
0.0025760718,
-0.0393769406,
-0.0030458996,
-0.0176948346,
0.0715137646,
-0.0428002067,
0.0549024194,
-0.0085403668,
0.029532088,
0.0053808261,
0.0593517721,
-0.0239960011,
0.0118719535,
-0.0098945871,
-0.0224280823,
0.0394581519,
-0.0219753366,
0.0005841423,
0.0060133766,
-0.0166834258,
0.0030958585,
0.0030779794,
-0.0331138298,
-0.0859554335,
-0.0030073782,
-0.1392411441,
0.0103462776,
-0.0437025316,
-0.0278731938,
0.0162493978,
0.0230175834,
0.0031952949,
-0.0234226938,
-0.0246207397,
0.0287963841,
0.0519157723,
0.0304298904,
0.0104466137,
0.0135438237,
-0.0152689107,
-0.0308798105,
0.0356370136,
0.0220196601,
0.0149535518,
0.0024153297,
-0.0422457419,
-0.0468421355,
-0.000750119,
-0.0693861917,
0.003684886,
-0.0194059238,
-0.0374446921,
0.0228474084,
0.0292096063,
0.0042391364,
0.034289699,
0.0591942295,
0.021236131,
-0.0504032224,
-0.0396781713,
-0.038076818,
0.0307791028,
0.0030572203,
0.0193349067,
-0.0200880188,
-0.0000208591,
-0.0016648726,
0.0339760371,
-0.0204429086,
0.0524393804,
-0.0554620288,
-0.001205917,
-0.0076843253,
-0.0636043474,
0.0101779895,
-0.011234005,
-0.0248202737,
0.0638083145,
-0.0431320742,
-0.0409358926,
0.0384998024,
-0.0130295753,
-0.0069070696,
-0.0024781476,
0.0337493308,
0.0199023429,
0.0347297564,
-0.0313424543,
-0.0058610006,
-0.0210741013,
-0.0341321342,
0.015915364,
-0.0119361402,
-0.0088621434,
-0.04768629,
0.0474660918,
-0.0002215839,
0.0037055302,
0.0346862301,
-0.0117631489,
-0.1004004851,
0.0005381437,
-0.0473149642,
-0.0429318324,
-0.0461355634,
0.0130872959,
-0.037860509,
0.037954621,
-0.0266878009,
0.0187256448,
-0.004913873,
-0.009612211,
0.0035818706,
-0.0182791874,
0.0189827848,
0.030508792,
0.0587234721,
0.0004530869,
0.1204767451,
0.0607833192,
0.0651567578,
0.0289672017,
0.0063979183,
-0.0251626726,
-0.0261378717,
-0.0007298224,
-0.013524862,
-0.0205493364,
0.0958601534,
-0.0217123758,
-0.0546294451,
0.1203402057,
0.0289670322,
-0.0207885373,
0.0008289003,
-0.0545363985,
0.0462408401,
0.004494953,
-0.0311660934,
-0.1249878556,
-0.0195522495,
0.0151915625,
-0.0154584385,
-0.0168402623,
-0.0694384277,
0.0024187407,
-0.0382581465,
0.0124254264,
0.0082290675,
-0.0043710074,
-0.0277547389,
-0.0166724492,
0.030762773,
0.0144035332,
0.0550341941,
0.0037839753,
0.061153885,
-0.0304736532,
-0.0234198906,
-0.0452696942,
-0.0202140305,
0.0550796576,
-0.022051964,
0.0754345506,
0.0519433208,
0.0630792007,
0.0250911377,
-0.0704232007,
0.0375504754,
0.0057274699,
0.0094230724,
0.0010461474,
0.0399906486,
0.0073023294,
0.0040672626,
-0.0085608819,
0.0214053392,
-0.0326517522,
-0.0004197033,
0.0455116965,
0.0509615652,
-0.0796840787,
0.0060558603,
0.0169984382,
0.0456770323,
0.0457010046,
0.0818631351,
0.0423233099,
-0.0240215119,
-0.0182416625,
-0.004849189,
0.033277005,
-0.0246118456,
0.0005767281,
0.0235114731,
0.0557599999,
0.0017850794,
-0.0873216838,
-0.0635167882,
-0.0216697417,
0.0388461724,
0.0083337193,
-0.0031009316,
-0.0353336558,
0.0234337933,
-0.0053156754,
0.0128135774,
-0.0228316579,
0.0158647094,
-0.0444338582,
0.0101796519,
-0.0374325924,
0.0057706521,
0.0131987091,
0.0303844493,
-0.0611102842,
-0.0348695293,
0.0422039628,
0.0130412877,
0.0328172185,
0.0740571842,
-0.0430477709,
0.0161155667,
-0.0413913466,
-0.023309378,
-0.0220891796,
0.0009735363,
0.0610795729,
0.0284217875,
-0.0057812943,
0.049906183,
-0.0073588579,
0.016482668,
-0.0388988778,
0.0316461287,
-0.0105723785,
-0.0573171489,
-0.0033321462,
-0.0430319272,
0.0145664942,
0.0242111571,
-0.0263765939,
-0.0162162166,
0.0242548324,
-0.0235028341,
-0.0036202294,
-0.0326665267,
0.0203839112,
-0.0190348178,
-0.0069733555,
0.0109911775,
-0.0297650602,
-0.0688719079,
0.0084178075,
-0.0568350293,
-0.0287897754,
0.0603166372,
0.0105721559,
-0.0184279382,
-0.036520794,
-0.0398099869,
0.0436111242,
0.0374621078,
0.052526366,
-0.0067075477,
0.0899618566,
0.0054849801,
-0.0327781066,
0.0386552289,
0.0167151112,
-0.0687556267,
-0.0382193699,
0.0002125165,
0.0318838432,
0.0192740094,
0.0268329363,
0.0224473383,
-0.0039546201,
0.0168691352,
-0.0057163504,
0.0472402684,
0.064104937,
-0.0383788422,
-0.0413310267,
0.0237264875,
-0.0360781327,
0.02736957,
0.0120329196,
-0.0486145541,
0.0662040412,
-0.0093836086,
0.0032454068,
-0.0550737269,
-0.030232979,
-0.039625451,
0.0370443575,
-0.0062029636,
0.0110364305,
-0.0684374571,
-0.0365798399,
0.0097826002,
-0.008562522,
-0.0032785651,
0.0014591462,
0.0209309161,
0.0552099459,
-0.0182681493,
0.00297852,
-0.0211387966,
-0.0085503859,
0.0108169178,
0.0050933738,
-0.0320314094,
-5.950394281e-33,
-0.0249565132,
-0.0036561887,
0.0030652962,
-0.0077973837,
-0.0685961172,
-0.0104069347,
-0.039088618,
-0.0008614933,
-0.0178333782,
0.0296488348,
-0.0188643243,
0.0563746653,
0.0153457085,
0.0080194455,
0.0095767099,
-0.0377470404,
0.034515962,
0.0335131474,
-0.0241786893,
0.0623831786,
0.036452774,
0.0076719718,
0.0288372897,
-0.0876783133,
0.0429977477,
-0.0160999242,
-0.0100117745,
0.020085711,
0.020422196,
0.0515878536,
-0.0329163782,
0.0250873193,
0.0230388455,
-0.0444184728,
0.0051162713,
0.0301854331,
-0.0337131582,
-0.0358645357,
0.0189792067,
0.044504609,
-0.0271763317,
0.0259701628,
-0.0145474635,
-0.024714401,
0.0214795861,
0.0209286641,
-0.0030100776,
-0.028334152,
-0.0236663297,
0.0352828987,
-0.0506259464,
0.0137308668,
-0.0230794027,
-0.0564116389,
-0.0274101552,
0.0599047951,
0.0003261185,
-0.0419032201,
-0.0493357889,
0.01319829,
-0.0090887491,
-0.0364323705,
0.0113214776,
0.0144270677,
0.0002048837,
0.0789317712,
0.0186022948,
0.0447188728,
0.0004242822,
-0.035814587,
0.0016079023,
0.0579893813,
0.0282735787,
0.0149977636,
-0.034788236,
-0.0163831804,
-0.0293919481,
-0.0216297694,
0.0307320394,
0.0084790587,
0.0158334617,
-0.0176714938,
-0.0861662626,
0.0113452729,
-0.0159545857,
-0.00815384,
-0.0105990218,
-0.0175883658,
0.0370127968,
0.0120567372,
-0.0261260252,
-0.001734687,
0.0467052981,
-0.018329028,
0.0059323492,
0.0656709597,
0.0013986331,
-0.0167449433,
-0.0127485441,
-0.0000321633,
0.0019127647,
-0.0118632447,
0.0352841392,
0.0280336626,
0.0409473218,
-0.0127909202,
0.0371284485,
-0.0318803415,
-0.0396912992,
-0.0126496442,
-0.0276688337,
0.0575610623,
-0.0256236959,
0.0263291746,
0.0320686139,
0.0102986693,
-0.0396015197,
0.036031384,
0.0433757342,
0.0041034594,
-0.0316280276,
0.0009737484,
0.0110940579,
-0.0005921444,
-0.0164605714,
-0.0418012999,
-0.0245607104,
-0.0345826223,
-0.0161937773,
-0.0459483042,
0.0058631944,
-0.0192866679,
2.667e-7,
-0.0383442752,
0.0001636067,
0.0317341909,
-0.042797178,
0.0450803898,
0.0236260034,
0.0170657951,
0.0144110406,
0.0197123457,
-0.0979598239,
0.0033217038,
-0.0047397134,
0.0880604014,
0.022556914,
-0.0844774544,
-0.0721920133,
-0.0842354596,
-0.0347757041,
-0.0546009392,
-0.0334107764,
0.0418965109,
0.0468613282,
0.0249343235,
-0.0125181181,
-0.0092205489,
0.0083611999,
0.0122302854,
-0.011677389,
0.0019788444,
0.0149463657,
0.0516960807,
-0.0040123947,
0.0093895504,
0.0675374418,
-0.0056357807,
0.0199524313,
-0.0091191297,
0.0276342314,
-0.0052576629,
-0.0742987692,
-0.0225890148,
-0.0082131308,
0.0169928391,
-0.0593074858,
0.0563245788,
-0.0030151051,
0.0078555383,
-0.0070808763,
-0.0196096729,
-0.043615263,
0.0419087298,
0.0374245308,
-0.0301422495,
0.0303355083,
0.004349892,
0.0164078604,
0.0049495129,
-0.0572577752,
0.0199408457,
-0.0301203392,
-0.0278831515,
-0.0638786107,
-0.0025572546,
-0.0462434664,
-0.022064602,
-0.0181648787,
0.0210678857,
2.081525218e-34,
0.0302754752,
-0.0010488125,
0.0501018427,
0.0503641441,
0.0376615785,
0.0009752318,
0.0656282678,
-0.0148547301,
0.0436833426,
-0.0353673324,
0.016618738
] |
http://arxiv.org/abs/2405.15097v1
|
http://arxiv.org/pdf/2405.15097v1
|
Contrastive and Consistency Learning for Neural Noisy-Channel Model in Spoken Language Understanding
| 2024-05-23T23:10:23 | 2024-05-23T23:10:23 |
Suyoung Kim, Jiyeon Hwang, Ho-Young Jung
|
cs.CL
|
cs.CL, cs.AI
|
Recently, deep end-to-end learning has been studied for intent classification
in Spoken Language Understanding (SLU). However, end-to-end models require a
large amount of speech data with intent labels, and highly optimized models are
generally sensitive to the inconsistency between the training and evaluation
conditions. Therefore, a natural language understanding approach based on
Automatic Speech Recognition (ASR) remains attractive because it can utilize a
pre-trained general language model and adapt to the mismatch of the speech
input environment. Using this module-based approach, we improve a noisy-channel
model to handle transcription inconsistencies caused by ASR errors. We propose
a two-stage method, Contrastive and Consistency Learning (CCL), that correlates
error patterns between clean and noisy ASR transcripts and emphasizes the
consistency of the latent features of the two transcripts. Experiments on four
benchmark datasets show that CCL outperforms existing methods and improves the
ASR robustness in various noisy environments. Code is available at
https://github.com/syoung7388/CCL.
|
Contrastive and Consistency Learning for Neural Noisy-Channel Model in Spoken Language Understanding: Recently, deep end-to-end learning has been studied for intent classification
in Spoken Language Understanding (SLU). However, end-to-end models require a
large amount of speech data with intent labels, and highly optimized models are
generally sensitive to the inconsistency between the training and evaluation
conditions. Therefore, a natural language understanding approach based on
Automatic Speech Recognition (ASR) remains attractive because it can utilize a
pre-trained general language model and adapt to the mismatch of the speech
input environment. Using this module-based approach, we improve a noisy-channel
model to handle transcription inconsistencies caused by ASR errors. We propose
a two-stage method, Contrastive and Consistency Learning (CCL), that correlates
error patterns between clean and noisy ASR transcripts and emphasizes the
consistency of the latent features of the two transcripts. Experiments on four
benchmark datasets show that CCL outperforms existing methods and improves the
ASR robustness in various noisy environments. Code is available at
https://github.com/syoung7388/CCL.
|
[
0.0209085718,
0.0824728534,
0.0245245732,
0.0107713528,
0.0074392366,
0.03473977,
0.0076054027,
-0.0496059693,
-0.0061029946,
-0.0095994435,
-0.0314517133,
-0.0619355924,
0.0232002512,
0.0199573282,
-0.0008026057,
-0.0013497249,
0.0625869334,
-0.0451151393,
0.0190186501,
-0.0124710342,
0.0183985848,
0.0072452626,
-0.0090458486,
0.0104200002,
0.002243547,
-0.012673567,
-0.0210508592,
0.022479184,
0.0113033848,
-0.014242474,
0.0153982276,
0.0379480459,
0.0474574044,
0.0284362137,
0.0000022355,
-0.053420078,
-0.039017152,
-0.0013678912,
-0.0803065524,
-0.0058088559,
0.0302908998,
0.0138819078,
0.0051810485,
0.0425620563,
0.0026228344,
-0.0200866349,
0.051401034,
0.0455369875,
0.0422258675,
0.0411817878,
0.0094706919,
-0.055930119,
0.0039218869,
-0.0187722873,
0.0282628462,
-0.0843617544,
0.0032100903,
0.0045068325,
-0.0142629473,
-0.0199678559,
-0.0061203237,
0.0335924961,
0.0170011092,
-0.0209194999,
0.0129002193,
-0.004126675,
0.0166570563,
-0.0587685779,
-0.0006711918,
0.087717399,
0.0049041728,
-0.0165316314,
0.0136721907,
-0.0431458689,
-0.040210288,
-0.0516953841,
0.0041654282,
0.0199623592,
0.0005717359,
0.0179010611,
0.0765541121,
-0.0044132299,
0.0300702229,
-0.0018511813,
-0.0560120605,
-0.0087875286,
0.0071153622,
-0.0283304397,
-0.0017085443,
-0.0271506235,
0.005389554,
-0.0373121165,
0.0208469387,
0.0001335871,
-0.0762519762,
0.0069210459,
-0.0216185674,
-0.0485409908,
0.0146324448,
-0.0607865937,
-0.0613678731,
0.0659498721,
-0.0760157853,
0.0475218296,
0.0367739797,
0.066488415,
0.0288003683,
0.0236249343,
-0.0307070222,
-0.0047858339,
-0.014901055,
0.0310518909,
-0.0201966949,
0.0403674282,
-0.0293489471,
-0.0278120823,
-0.0537086092,
0.0326264314,
-0.0258113798,
0.0336009003,
0.0348555483,
-0.0103082713,
-0.0580748357,
0.0141237816,
0.0219059102,
0.0279205292,
0.0217411499,
0.0309027955,
-0.0163643118,
-0.0615458749,
-0.014755616,
0.0280203391,
0.0172962341,
-0.0190137699,
0.0080899335,
0.0411982611,
0.0015770006,
-0.0498851761,
-0.0477385521,
0.050941512,
-0.0099086612,
-0.0427505672,
-0.0475613214,
-0.0326914862,
0.0130638955,
-0.0064783916,
-0.0008992968,
-0.0600305051,
0.0034195632,
-0.0449908562,
-0.0174636785,
0.0986575708,
-0.0146362996,
-0.01984464,
0.0276761055,
0.0015727401,
-0.0095686456,
0.0166879147,
0.0072284513,
-0.0426933542,
-0.0068090167,
-0.0413709581,
-0.0122726811,
0.0158726182,
-0.0002355031,
-0.0446512774,
0.0121824536,
0.0303581003,
0.0085766101,
0.0183969326,
0.0361585505,
0.0639817268,
-0.0346853957,
-0.025323702,
0.0421385914,
-0.0040173577,
0.0464941077,
0.0011073611,
0.0186764263,
0.0605093874,
-0.0292281657,
0.0113890553,
-0.0404413976,
-0.0002658241,
-0.0243161228,
0.0227446724,
0.0780283287,
0.0197258834,
-0.0418112576,
-0.0326608904,
-0.0222532768,
-0.0287102051,
0.0276955292,
0.057486888,
0.0529070608,
-0.0451050885,
-0.0556410998,
0.026105307,
0.0166926421,
-0.0110934339,
0.0286438689,
-0.0385364555,
0.0246673673,
0.0431573391,
-0.0012286111,
-0.0410102755,
-0.0408461913,
-0.0610506795,
-0.0476205871,
0.0263523571,
0.011198584,
-0.0044352203,
0.0891821608,
-0.0628363043,
-0.0180373415,
0.0043092202,
0.0026492104,
0.0195919983,
-0.0665911809,
-0.0166689493,
0.0016884418,
-0.0128984554,
-0.0223952234,
0.0050304485,
-0.0277722459,
-0.0241285991,
0.0295189135,
0.0044463244,
0.0873467773,
0.0221870802,
-0.0393729806,
-0.009388715,
0.0141796218,
0.0036780145,
0.0280403066,
0.0282077398,
-0.0743909031,
-0.0606056191,
0.0545734428,
-0.0053087757,
-0.0384746715,
0.046770595,
-0.0148164155,
-0.0679103509,
0.0180863366,
-0.0356924273,
0.0335878618,
-0.0710300654,
0.0385589786,
0.0248198342,
0.0193461385,
-0.002311277,
0.0171242859,
-0.0127864713,
-0.012282121,
0.0528803319,
0.0118455868,
0.0061946097,
-0.0436178483,
-0.0071949749,
0.0032422328,
-0.0362951458,
-0.0069641992,
-0.0051136352,
0.0356851667,
-0.0321769863,
0.0020636199,
-0.0077645117,
-0.0318885706,
-0.0443591736,
0.0022616438,
0.021539852,
0.0412728339,
0.0289644916,
-0.0018949673,
-0.0012873904,
0.013843054,
-0.0036138608,
-0.0222881846,
-0.0138407601,
0.0135088842,
-0.0385211334,
-0.0487926677,
0.0245881807,
0.0007254793,
0.0021097374,
0.0013026794,
-0.0249601342,
-0.0920433253,
0.0677321628,
0.0141624976,
0.0630198494,
-0.0152222039,
-0.0355905741,
-0.0004601429,
-0.0247088298,
-0.0134499986,
0.0068310555,
0.0084179454,
0.0116653582,
0.0224416405,
-0.028336525,
-0.0608917251,
0.0341442712,
-0.0011156698,
0.0281482246,
-0.0180036519,
-0.020573996,
-0.0294081382,
0.0647110194,
0.0305456575,
0.0242229346,
-0.0290714279,
-0.075133495,
0.0125620626,
-0.025904566,
-0.0022895623,
-0.0240778197,
-0.0331247896,
0.035023693,
-0.0073693432,
0.0070760171,
0.0812388361,
-0.0717336163,
-0.0442590304,
-0.0817593038,
-0.0649522766,
-0.0012600407,
-0.0045472062,
-0.0441589691,
-0.0008514355,
0.0226208493,
-0.0810385346,
-0.0273987949,
-0.0299124122,
0.0102948798,
-0.0320784897,
0.0729526356,
-0.0236791205,
0.0366399437,
-0.0014762937,
-0.0374538451,
-0.0655046552,
-0.0517057255,
-0.0231837537,
0.0252874941,
0.0250609815,
-0.0280892719,
-0.019045746,
0.0105737345,
-0.0150592886,
0.0673751459,
0.0370669141,
0.0125667276,
-0.0188583136,
0.0374716148,
0.0301539358,
0.0458585359,
0.0222294126,
-0.0021915855,
0.0407750569,
-0.0151386382,
0.0185676645,
0.0129039036,
-0.0484262966,
0.0266620554,
0.0699266866,
0.026648799,
-0.0203859862,
0.0377577171,
0.0657915324,
-0.0296032541,
-0.0383464433,
0.1254006922,
-0.0035737574,
-0.0520418398,
-0.0280317888,
-0.0307497978,
0.0137516707,
0.0265956521,
0.0087471902,
-0.0685321242,
-0.0564847514,
0.0402504094,
0.0421604887,
0.0160201825,
0.027257992,
0.0049004797,
-0.0704860538,
0.0258984007,
0.0089412713,
-0.0956910402,
-0.0485887863,
-0.0380596034,
0.008205655,
-0.0524576195,
0.0892784968,
0.0083129788,
0.0193417761,
-0.0333772786,
0.026988402,
-0.0429150201,
-0.0806734785,
0.0386531018,
-0.0212073736,
0.1297225058,
0.0478791669,
-0.0195523705,
-0.0079481117,
-0.0373357907,
0.0624214076,
-0.0455098301,
-0.0396586135,
-0.0382832289,
-0.0000662696,
0.0210005865,
-0.0252054632,
0.0188465882,
0.0152808512,
-0.0043308083,
0.0090385498,
-0.0487755351,
0.025644362,
-0.0569364764,
0.029176984,
0.0247395337,
0.0488082767,
0.0666778237,
0.0908959433,
0.0387466252,
-0.0199586432,
-0.02925721,
-0.0161740016,
0.0465678535,
-0.0125785489,
0.0208544936,
0.0075934427,
0.0070447768,
-0.0240808222,
0.0181894396,
-0.0473778471,
-0.0009247633,
-0.0248783212,
-0.0119948247,
0.0177176241,
-0.0009415142,
-0.0085196299,
-0.0063708788,
0.0192433633,
0.0297879074,
0.021694677,
-0.0397927389,
-0.010433387,
0.0029549673,
-0.087981835,
-0.028655624,
0.0346999802,
-0.0883947611,
-0.0628080741,
0.0396953821,
-0.0096996538,
0.0197243299,
0.0784680694,
-0.0541633032,
-0.0101619288,
-0.0642159656,
0.0061845738,
-0.0396634936,
-0.0110997409,
0.0504642427,
0.0187629368,
-0.0133335954,
0.0207003802,
-0.0492747687,
-0.0050045745,
0.0019085255,
0.0504614636,
-0.0641522557,
-0.058308132,
0.0324951224,
0.0207548216,
0.0589242354,
0.0195324067,
-0.0499205403,
-0.0751428083,
0.0363674387,
0.0372563936,
0.0144480933,
-0.0121516092,
-0.0035120833,
0.0278554019,
-0.0037503296,
-0.0145806652,
0.012156425,
-0.0456123017,
0.0184246711,
0.0037676923,
0.0310705192,
0.0427621715,
0.0128632039,
-0.0044784062,
-0.025088001,
-0.0338136144,
0.055049099,
0.0255287755,
0.0350652747,
-0.0011164276,
0.0334159657,
0.0070527876,
0.0061824438,
0.012598143,
-0.0118210986,
-0.0492044576,
-0.0479629822,
-0.0221060254,
0.0159716252,
0.0424838178,
0.0045886352,
0.0429530405,
0.0195387546,
0.0342621244,
-0.0499860384,
0.0650752336,
-0.028900167,
0.0144372676,
0.0510518961,
0.0118669029,
0.0270519983,
0.0148729999,
0.024859868,
-0.0039152056,
0.0474871919,
-0.0359382182,
0.0364536829,
-0.050159052,
-0.0029309897,
-0.001963055,
0.0353878401,
-0.0716674924,
0.0037471005,
-0.0052865022,
-0.0709753633,
0.0138058672,
0.0071723238,
0.0429096036,
-0.0004480257,
0.0071113575,
0.0607598089,
-0.0254918151,
0.0092870314,
0.0122696981,
-0.0409720689,
0.0372632891,
-0.0279173274,
-0.0165265668,
-7.466277266e-33,
-0.0058711842,
0.0363025293,
0.017731851,
-0.0480489396,
-0.0295862649,
-0.0001507604,
-0.0420213267,
-0.0196602978,
-0.0307783075,
-0.0281942189,
0.0015656459,
0.0381919667,
0.0322921462,
0.0264806785,
-0.000694253,
-0.0383484438,
0.0910032317,
0.0113305291,
0.0237804092,
-0.0052767927,
0.0245775748,
0.0604867451,
0.0420145243,
-0.0811821073,
0.0097187832,
0.0233256072,
-0.0010852207,
0.0273743868,
-0.0093125328,
0.0339382514,
-0.0155044133,
-0.0065599317,
-0.0054657385,
-0.0487004593,
-0.0259954613,
-0.0371560566,
-0.124098286,
-0.0202672761,
0.0193967894,
0.0330306552,
-0.0418671854,
-0.0058112349,
0.0507070199,
-0.0129836677,
-0.0014375102,
-0.0260648895,
0.0283135977,
0.001643814,
0.0005749907,
0.0142115913,
-0.0070825163,
0.0350452438,
-0.0572000295,
-0.0113497684,
0.041737929,
0.0249645188,
-0.0149514377,
0.0120015321,
-0.1113552824,
-0.0096422238,
-0.019923659,
0.0152851483,
0.0284041744,
0.0279250741,
-0.0200928766,
0.0307201594,
0.0271611307,
-0.0502146073,
0.0191534553,
-0.0531014279,
0.005801308,
0.0272179544,
-0.0148815922,
0.0326671042,
0.0456647091,
-0.0136492187,
-0.06293194,
-0.0042084027,
0.0158242434,
0.0366853848,
0.0053928713,
0.0386422239,
-0.0353377834,
-0.0142526384,
-0.0439438373,
-0.0202415828,
-0.0597403198,
0.025501065,
0.0260308031,
0.025607286,
-0.009422549,
-0.0064816345,
0.0648395494,
0.0056976881,
0.0414973423,
0.047447186,
0.0254834983,
-0.0292242095,
-0.0041329064,
0.0517944321,
-0.0522865243,
-0.0326050185,
0.025395209,
0.0393638052,
0.0197062455,
0.0057290788,
-0.0067003318,
0.0357874557,
-0.00577683,
0.0079365112,
-0.0198905598,
0.0228112098,
-0.0480219349,
0.0025436769,
0.0454044938,
-0.0133500565,
-0.001606854,
0.0591685288,
-0.0089595309,
0.0353948958,
-0.0201805532,
-0.0066604456,
0.0208355226,
0.038234774,
-0.0525160432,
-0.0069786399,
0.0221330803,
-0.0099813435,
0.0158771686,
-0.0351518951,
-0.0024116901,
0.0062167924,
2.991e-7,
-0.0095452769,
0.0033306018,
0.0198466312,
0.0320045389,
0.0155897848,
0.034984272,
-0.0376808681,
-0.0150601408,
-0.0024348102,
-0.005420045,
0.030060146,
0.0090441834,
0.0540840439,
0.018348502,
-0.0302960686,
-0.013828014,
-0.0482137874,
0.0118034342,
0.018162882,
0.006640438,
0.0715700164,
0.1123283505,
-0.0039798147,
-0.059454456,
0.0360458307,
-0.0101450142,
0.0008298568,
-0.0067107542,
0.0370271392,
-0.0141505683,
0.0008617034,
0.0009889374,
0.0069471113,
0.0169238765,
-0.0046886601,
-0.0016578747,
-0.0291443169,
0.0365183353,
-0.0076518357,
-0.0510541536,
0.0046975128,
0.0564161129,
0.0139996735,
-0.0445222743,
0.0468240455,
-0.0109190634,
-0.0159367416,
-0.0272714142,
-0.0525803044,
-0.0093160747,
0.0129516851,
0.0160622522,
0.0321452804,
-0.0089431163,
-0.0103328442,
-0.0290238634,
-0.0462453812,
-0.0688950121,
0.0128640234,
0.0105329594,
-0.0544441044,
0.0111866202,
0.0006634334,
0.0079450281,
0.0817651451,
-0.0341128744,
0.0157530531,
2.425962053e-34,
0.068573378,
-0.0012498278,
-0.0033835061,
0.0318034589,
0.0233013369,
-0.027365597,
-0.0274528768,
-0.0020115629,
0.0341068842,
-0.0759690553,
-0.0330389813
] |
http://arxiv.org/abs/2405.15092v1
|
http://arxiv.org/pdf/2405.15092v1
|
Dissociation of Faithful and Unfaithful Reasoning in LLMs
| 2024-05-23T22:38:58 | 2024-05-23T22:38:58 |
Evelyn Yee, Alice Li, Chenyu Tang, Yeon Ho Jung, Ramamohan Paturi, Leon Bergen
|
cs.AI
|
cs.AI, cs.CL
|
Large language models (LLMs) improve their performance in downstream tasks
when they generate Chain of Thought reasoning text before producing an answer.
Our research investigates how LLMs recover from errors in Chain of Thought,
reaching the correct final answer despite mistakes in the reasoning text.
Through analysis of these error recovery behaviors, we find evidence for
unfaithfulness in Chain of Thought, but we also identify many clear examples of
faithful error recovery behaviors. We identify factors that shift LLM recovery
behavior: LLMs recover more frequently from obvious errors and in contexts that
provide more evidence for the correct answer. However, unfaithful recoveries
show the opposite behavior, occurring more frequently for more difficult error
positions. Our results indicate that there are distinct mechanisms driving
faithful and unfaithful error recoveries. Our results challenge the view that
LLM reasoning is a uniform, coherent process.
|
Dissociation of Faithful and Unfaithful Reasoning in LLMs: Large language models (LLMs) improve their performance in downstream tasks
when they generate Chain of Thought reasoning text before producing an answer.
Our research investigates how LLMs recover from errors in Chain of Thought,
reaching the correct final answer despite mistakes in the reasoning text.
Through analysis of these error recovery behaviors, we find evidence for
unfaithfulness in Chain of Thought, but we also identify many clear examples of
faithful error recovery behaviors. We identify factors that shift LLM recovery
behavior: LLMs recover more frequently from obvious errors and in contexts that
provide more evidence for the correct answer. However, unfaithful recoveries
show the opposite behavior, occurring more frequently for more difficult error
positions. Our results indicate that there are distinct mechanisms driving
faithful and unfaithful error recoveries. Our results challenge the view that
LLM reasoning is a uniform, coherent process.
|
[
0.0372694731,
0.0339863263,
-0.045151934,
0.0171138942,
-0.0684988573,
0.0134776868,
-0.029318817,
0.0020822776,
0.0050262432,
-0.0550028346,
-0.0424481556,
-0.042834688,
0.0202985443,
-0.0206666403,
0.0405555554,
-0.0131753525,
0.0600372888,
-0.0717713907,
0.0087273959,
-0.0238662455,
-0.0223576501,
-0.0118638314,
-0.0121273268,
-0.0116250701,
0.0089836512,
-0.0146055268,
-0.0190850105,
-0.001594579,
-0.0062961127,
-0.0375598595,
-0.0147642912,
0.0239414964,
0.0165398512,
0.0534742549,
0.0000019234,
-0.0642964914,
-0.0026713773,
-0.0362994857,
-0.0426464044,
0.0371511579,
-0.0517850779,
0.0177192912,
-0.0071625095,
0.0289000627,
0.0117724519,
0.0320602991,
0.0056869271,
0.0579697527,
0.026226934,
0.0432584882,
0.0281802043,
-0.0104952687,
0.0160966497,
-0.0114692608,
0.0101208268,
-0.0430300161,
0.0193980895,
0.0461754501,
-0.0287267528,
0.0164826512,
-0.0264637936,
0.0047692079,
0.0171768442,
0.0160658415,
-0.0434760638,
0.0254727043,
0.0258272849,
-0.0461387821,
-0.0143041722,
-0.0061804284,
0.0278240032,
-0.043041531,
-0.0235704947,
-0.025437234,
0.0078765731,
0.019471379,
-0.0600996651,
0.0706995651,
-0.0362761319,
-0.0017460875,
0.0051197782,
-0.058613494,
0.0450348556,
-0.0413274132,
-0.0168053433,
0.0816906691,
-0.0122390585,
-0.0663361847,
0.0119606713,
0.0209383238,
-0.0531051531,
-0.0808712393,
-0.010146928,
0.030859204,
-0.0341821685,
0.0255629364,
0.0297099166,
0.0092388382,
0.0122958031,
-0.0096486975,
0.0449576117,
0.0274182353,
-0.00813937,
0.0177456345,
-0.0521515384,
0.0376913138,
-0.0513104387,
0.0345610194,
-0.0027549667,
0.010304668,
-0.0231117141,
-0.0553903468,
0.0183129702,
0.0730482265,
0.0300859325,
0.0216777027,
-0.0220145453,
0.0196391195,
-0.0342428237,
0.0087938206,
-0.0105077401,
0.026752267,
0.0276126433,
-0.0346823856,
-0.0905217826,
0.0202695597,
0.0007183953,
0.0872553736,
0.0327230915,
-0.0311201792,
-0.027303502,
-0.0219472591,
-0.0296716373,
-0.0048442311,
0.0071681715,
0.0522084162,
-0.0169748105,
-0.0116276145,
-0.0784209371,
-0.0192639008,
-0.0201863591,
-0.0819406733,
0.0119445659,
-0.0438939333,
0.0081996461,
0.0352040008,
-0.0139014488,
-0.0161012299,
0.0329303853,
0.0228463728,
-0.0491535999,
0.0392266624,
0.0063916119,
0.0240548998,
0.0334068723,
0.0186172351,
-0.0371948294,
0.0212503076,
0.0116333719,
0.0212295875,
-0.0028279352,
-0.0043326947,
0.0060716751,
0.0153374551,
0.0140694072,
-0.0350462086,
-0.0336871818,
0.0111723645,
-0.0049624327,
0.0245487336,
0.0394105464,
0.0530875288,
-0.0363432616,
0.0361205488,
0.0533149913,
0.0475926846,
0.0095163705,
0.009689101,
0.0108501213,
-0.0634689853,
0.0067426539,
0.0498390645,
-0.0979701653,
0.0499552786,
-0.0101806065,
0.0038705352,
0.0191510152,
-0.0059835482,
-0.0341056548,
-0.0141374171,
-0.0352645479,
0.0239449572,
-0.046014946,
0.0163608696,
0.0236051809,
0.0011485834,
-0.0445910059,
0.0637671202,
-0.0010843234,
-0.0111193443,
-0.0040787747,
-0.0708207041,
0.048869431,
0.0256929733,
0.0233342387,
0.0037200041,
-0.0213564578,
-0.0109248944,
0.0198452044,
0.0179258659,
0.0201172475,
0.0238235965,
-0.0307014529,
-0.0113236001,
-0.0064086276,
-0.0296881944,
0.0218925569,
-0.0159754436,
-0.015470447,
-0.0199543517,
-0.0026357751,
0.0059868242,
0.0182645731,
0.021423148,
-0.0356678814,
0.0015169085,
-0.0566502362,
-0.0118202465,
0.0163333509,
-0.0333148167,
-0.0160593875,
0.0413752869,
0.0102288201,
0.014878734,
-0.0379216373,
0.0264713969,
-0.0714770034,
0.0079327319,
0.0392984115,
0.0010010547,
-0.0397161357,
0.0014454307,
0.0097156381,
-0.034425173,
0.018525295,
-0.0394812375,
0.0107384166,
-0.0118496343,
0.0746088102,
-0.0190503448,
-0.023799818,
-0.0130767636,
0.0625131279,
-0.0429741964,
0.0378927663,
-0.0054698251,
0.0109432433,
-0.0103072375,
0.0020760128,
0.0503143407,
-0.0105753317,
-0.0005740225,
-0.0083036255,
-0.0070357807,
-0.026832426,
-0.063318491,
-0.0479783416,
-0.0128794145,
-0.0050992053,
0.0080931326,
0.0092717214,
0.0201807544,
-0.0382010378,
-0.016761804,
0.0218593813,
0.0187665336,
0.0521745868,
0.0626809672,
0.0317602418,
0.0091555836,
-0.0382104889,
-0.0138066839,
-0.0456558764,
0.0152255604,
-0.0363278836,
0.0060508586,
-0.015465199,
-0.0120736826,
-0.0814259499,
0.0751737505,
-0.0414679348,
0.0269398112,
-0.0112060215,
-0.017076293,
0.0592726506,
0.0187973101,
0.0272754319,
0.0659262538,
0.0031426309,
0.0085748266,
0.0634296164,
-0.0290657375,
-0.043421153,
0.0182168949,
-0.0031872629,
-0.0417180061,
-0.00388283,
-0.0397479534,
-0.0264284816,
0.0785263553,
0.0365277156,
0.0188859235,
-0.0133119281,
0.0179541148,
0.0010181791,
-0.0316623487,
-0.0087339813,
-0.0651582405,
-0.0238387845,
0.0948555693,
-0.010810975,
-0.0457963347,
0.0491498336,
0.0043711173,
0.0334043503,
0.0220421664,
0.0675710067,
0.0106695648,
0.0630990863,
-0.0102886111,
-0.0046479101,
-0.005233902,
-0.0288341492,
-0.0579168834,
-0.0402846858,
0.0207133628,
-0.0331354104,
0.0609557368,
-0.0436143428,
-0.0137584275,
-0.0099680172,
-0.047030881,
-0.0673960075,
0.0074106045,
-0.0407721326,
-0.0112575162,
-0.0081605641,
0.0103968894,
-0.0197866913,
-0.0457246862,
0.096737422,
-0.0365295485,
-0.0140861524,
-0.0317469873,
-0.016106762,
0.0048415046,
0.0365643948,
0.0202571955,
0.096762076,
-0.0591680631,
0.0281543657,
0.0680504292,
0.0537864976,
0.055049926,
0.0026179999,
-0.0135742594,
0.0289730411,
0.016047053,
-0.0524805263,
0.0358052552,
0.0347080082,
-0.018496925,
-0.0353552178,
0.1051320285,
0.0404038392,
-0.0187116116,
-0.006515237,
-0.0963926017,
0.0230251197,
0.0451317914,
-0.0238841977,
-0.0754979551,
-0.0033518781,
0.0529046729,
-0.0254865456,
-0.0245520566,
-0.0275311302,
0.0067561739,
-0.1085343286,
0.0287962984,
0.0249214619,
-0.0023156353,
-0.0251342542,
-0.0346438959,
0.0150548322,
0.0695890933,
0.0492315702,
0.0780178979,
0.000973618,
-0.0114257764,
0.0234098919,
-0.0025001019,
-0.0107619576,
0.0346796215,
-0.0272206347,
0.0418104306,
0.0335907601,
-0.0066922107,
-0.0052988799,
-0.0619359352,
0.0841236264,
-0.0004788425,
0.0207625534,
-0.0109758256,
-0.0250902269,
0.0281581264,
0.0202729795,
-0.0403197035,
-0.0049732844,
-0.0077502285,
-0.0449292399,
0.0182262305,
0.091900751,
-0.0293957945,
-0.0482413024,
-0.0067192605,
0.0208900422,
-0.0523251854,
0.0394857861,
0.0193755999,
-0.0153571507,
-0.0524137877,
0.0004483591,
-0.0141895525,
-0.0007577021,
-0.0327374712,
0.0359440744,
0.0514031872,
0.0744070783,
-0.0273934286,
-0.0406420343,
-0.0017905325,
0.0082275197,
0.0052052476,
0.0179076288,
-0.0011170654,
-0.0364102013,
0.0035808329,
-0.0091262311,
-0.0341999456,
0.0167909898,
-0.0217257664,
0.0470471308,
-0.0200515613,
-0.0316901319,
0.0188829284,
-0.0160060897,
-0.0596437454,
-0.0232949164,
0.0238245185,
0.0456157513,
0.0342356414,
-0.0109406793,
0.0332326069,
-0.0451960228,
-0.0502203442,
-0.0102149863,
-0.0219809636,
0.0749084726,
0.008464694,
0.0192069914,
0.0103745321,
0.0323341228,
-0.0095637813,
-0.0140172904,
0.0510719232,
0.0336194821,
0.0308165271,
-0.0312629417,
0.0385044403,
0.0077147107,
0.0238726251,
0.0048602219,
0.0241627488,
-0.0313301943,
0.0399494842,
0.010239793,
-0.021271307,
0.0050641117,
-0.0059720767,
0.0463598445,
-0.0081342896,
0.0070387945,
-0.0598165616,
0.0101286247,
0.0267223697,
-0.0007130566,
-0.0302361753,
0.0173579194,
0.0326558352,
0.0287077576,
0.0014115061,
-0.0221051984,
0.0177024882,
-0.0128757469,
0.061343357,
0.0053550997,
0.0990114957,
0.0167227257,
-0.0267321765,
-0.0606416538,
-0.0388981439,
-0.0685625821,
-0.0713816658,
0.0456977002,
-0.0268140472,
0.0545675047,
-0.0098200459,
0.0351424366,
-0.0428907685,
0.0167972967,
-0.0159213971,
-0.004448093,
0.0132733947,
-0.0066929869,
0.0374950916,
0.015830189,
-0.0025242234,
-0.0125601385,
0.0459989198,
-0.02128659,
0.0340148397,
-0.0418306179,
-0.0106193377,
0.0059043192,
-0.0511680432,
-0.0171800815,
-0.0099274358,
-0.0295016225,
-0.0069655948,
-0.0723031834,
-0.0821481794,
-0.002547933,
0.0164006073,
0.0168962386,
-0.0628473908,
-0.0387450382,
0.0900866464,
0.0254162867,
-0.0236364435,
0.0045664343,
0.0317658111,
0.0246919282,
0.0275765751,
-0.0396166258,
-6.210816238e-33,
-0.0295225475,
-0.0315274373,
0.0481105484,
-0.0019058615,
-0.0256067198,
-0.0276995171,
-0.0160357878,
0.0408294648,
-0.005887391,
0.0121856024,
-0.0292097777,
0.0392665677,
0.0099480832,
0.0049443818,
0.0607949272,
0.0021710559,
0.0109116007,
-0.0024707776,
-0.0143553801,
-0.0236554723,
0.0548799708,
0.0022153943,
0.0264969692,
-0.070840247,
0.038323801,
-0.0014170673,
0.0162847806,
0.0046906401,
-0.0979884937,
0.0262383968,
-0.0056910282,
-0.0003997821,
0.0330729783,
-0.0291814208,
-0.0261653122,
0.0132294493,
-0.0002287743,
-0.0206603371,
0.0697623417,
-0.0013242281,
-0.0395358689,
-0.0106504606,
0.0223746989,
0.0201108716,
-0.0260170233,
-0.0402817652,
-0.0195412878,
-0.012057215,
-0.0194754079,
0.047228355,
-0.0739795417,
-0.0141299879,
0.0300504304,
-0.0134154996,
-0.0334536657,
0.0051498697,
-0.0324434713,
-0.0082972348,
-0.0716274381,
-0.0411975384,
0.0019012056,
0.0173507035,
0.0364898406,
0.0194109324,
-0.004387842,
0.0718378127,
0.0597991496,
0.0416906625,
-0.0597705208,
-0.0340202637,
-0.0213577934,
0.0481940396,
0.0533340387,
0.0068513379,
0.0068535465,
-0.0326116197,
-0.0680498928,
-0.0478158668,
-0.0169815924,
0.0527269281,
0.0365015529,
0.0096444394,
0.0362669565,
-0.0225321222,
-0.0724162534,
0.0119905928,
-0.052367378,
-0.0119814835,
0.0255960915,
0.0159713365,
0.0399030708,
-0.0480840765,
0.0199115369,
0.0140843019,
-0.0075073955,
0.0023301209,
-0.0116789127,
0.0116290422,
-0.0200317297,
0.0444497131,
-0.0559574999,
-0.0259254426,
-0.050425306,
0.0178143196,
0.0026775824,
0.0330580659,
0.007689402,
0.0088184448,
-0.0414127633,
-0.0039503616,
-0.0803030878,
0.027109433,
0.0017031167,
-0.0623342767,
0.0771992281,
0.0250200927,
-0.0297423769,
-0.0036328093,
0.0296955332,
-0.0392764173,
-0.0378638729,
0.0740917102,
-0.0374476984,
0.0287329741,
0.0043020179,
-0.0011614069,
-0.0442586653,
0.0364340022,
-0.0142109524,
-0.0519487523,
-0.0114300139,
0.0037795233,
2.781e-7,
0.0059594205,
0.0371271335,
0.0073403846,
0.030028211,
0.0552456938,
0.0326610543,
-0.045436237,
-0.0309018232,
0.1053271815,
-0.0382909626,
-0.0409122258,
0.0302370284,
0.0368987806,
0.040989615,
0.0256293714,
-0.0132276295,
-0.0782505348,
0.0060521816,
0.0082368208,
-0.0008648381,
0.0242501106,
0.0160112064,
0.0066965348,
-0.0328104757,
-0.0221534129,
-0.0123202056,
0.0970188081,
-0.0114449095,
0.0052511282,
0.0024933806,
-0.004511687,
-0.0066698748,
0.0258876327,
0.0410698615,
0.0483309627,
-0.0006634026,
-0.0312631801,
0.0074876142,
0.0159180537,
-0.0081369933,
-0.0372402817,
-0.0110723525,
0.0213390626,
-0.0672081485,
0.0503829792,
-0.061780408,
-0.0142801786,
-0.0180486944,
0.0312557407,
-0.0167122707,
0.0090905828,
0.0182183012,
0.0067754504,
0.0100517711,
-0.005350227,
0.058908131,
-0.01088663,
-0.0994251668,
-0.0326047018,
0.0611720197,
-0.0404114723,
-0.043855913,
-0.0315619968,
0.0489254668,
0.03520016,
-0.047512237,
-0.0113717727,
2.317437292e-34,
-0.0031614334,
-0.0291331001,
0.0673740432,
0.0676402673,
0.0419663116,
-0.0266053509,
0.0159656201,
0.0250988137,
0.0031553265,
-0.045785062,
-0.0306856241
] |
http://arxiv.org/abs/2405.15077v2
|
http://arxiv.org/pdf/2405.15077v2
|
Eliciting Informative Text Evaluations with Large Language Models
| 2024-05-23T21:56:12 | 2024-05-28T17:55:40 |
Yuxuan Lu, Shengwei Xu, Yichi Zhang, Yuqing Kong, Grant Schoenebeck
|
cs.CL
|
cs.CL, cs.AI, cs.GT
|
Peer prediction mechanisms motivate high-quality feedback with provable
guarantees. However, current methods only apply to rather simple reports, like
multiple-choice or scalar numbers. We aim to broaden these techniques to the
larger domain of text-based reports, drawing on the recent developments in
large language models. This vastly increases the applicability of peer
prediction mechanisms as textual feedback is the norm in a large variety of
feedback channels: peer reviews, e-commerce customer reviews, and comments on
social media.
We introduce two mechanisms, the Generative Peer Prediction Mechanism (GPPM)
and the Generative Synopsis Peer Prediction Mechanism (GSPPM). These mechanisms
utilize LLMs as predictors, mapping from one agent's report to a prediction of
her peer's report. Theoretically, we show that when the LLM prediction is
sufficiently accurate, our mechanisms can incentivize high effort and
truth-telling as an (approximate) Bayesian Nash equilibrium. Empirically, we
confirm the efficacy of our mechanisms through experiments conducted on two
real datasets: the Yelp review dataset and the ICLR OpenReview dataset. We
highlight the results that on the ICLR dataset, our mechanisms can
differentiate three quality levels -- human-written reviews, GPT-4-generated
reviews, and GPT-3.5-generated reviews in terms of expected scores.
Additionally, GSPPM penalizes LLM-generated reviews more effectively than GPPM.
|
Eliciting Informative Text Evaluations with Large Language Models: Peer prediction mechanisms motivate high-quality feedback with provable
guarantees. However, current methods only apply to rather simple reports, like
multiple-choice or scalar numbers. We aim to broaden these techniques to the
larger domain of text-based reports, drawing on the recent developments in
large language models. This vastly increases the applicability of peer
prediction mechanisms as textual feedback is the norm in a large variety of
feedback channels: peer reviews, e-commerce customer reviews, and comments on
social media.
We introduce two mechanisms, the Generative Peer Prediction Mechanism (GPPM)
and the Generative Synopsis Peer Prediction Mechanism (GSPPM). These mechanisms
utilize LLMs as predictors, mapping from one agent's report to a prediction of
her peer's report. Theoretically, we show that when the LLM prediction is
sufficiently accurate, our mechanisms can incentivize high effort and
truth-telling as an (approximate) Bayesian Nash equilibrium. Empirically, we
confirm the efficacy of our mechanisms through experiments conducted on two
real datasets: the Yelp review dataset and the ICLR OpenReview dataset. We
highlight the results that on the ICLR dataset, our mechanisms can
differentiate three quality levels -- human-written reviews, GPT-4-generated
reviews, and GPT-3.5-generated reviews in terms of expected scores.
Additionally, GSPPM penalizes LLM-generated reviews more effectively than GPPM.
|
[
0.0824267343,
0.0570787154,
0.0109890504,
0.0296070427,
-0.0569043644,
-0.0243222658,
0.0198329147,
0.014328138,
0.0107275322,
-0.0319955349,
-0.0493772291,
-0.0272042211,
0.0130251646,
0.0257182941,
0.0039675399,
-0.0284311436,
-0.0176887978,
-0.0196289569,
0.0749105662,
-0.0387641937,
0.0498472005,
-0.0026492223,
0.0555067919,
0.0145426737,
0.0051534413,
0.0121944342,
0.0281871203,
-0.0224133246,
-0.0147346212,
-0.0191288106,
0.006996728,
0.0823964551,
-0.0137680732,
0.0179215278,
0.0000019617,
-0.0307674501,
-0.0071566189,
0.0007159663,
0.0149120139,
-0.0012918458,
0.0853243992,
-0.021843804,
0.0504342988,
0.0035281498,
-0.056655813,
-0.0188618284,
-0.0029541617,
0.0276202857,
0.057686951,
0.0156153413,
0.0044706799,
-0.0133080566,
-0.0206306726,
-0.0266707949,
0.0175218768,
-0.0584372878,
-0.0055426438,
0.014607843,
-0.0312433001,
0.0407425314,
0.0234292876,
0.0019220773,
-0.0023861756,
-0.0490269028,
0.0901126787,
0.0356154703,
0.0163980238,
0.002438969,
-0.0152030066,
0.0215765703,
0.0312103797,
-0.0194345526,
-0.0036734999,
-0.0265723001,
0.0193631444,
0.0507668555,
-0.05257551,
0.0348277129,
-0.0062208641,
0.0406068563,
0.0028899512,
0.0394819416,
0.0489008762,
0.0245501716,
0.0878942311,
0.0784132034,
-0.0147301592,
-0.0398270413,
0.0096964352,
0.0684921816,
-0.008918331,
-0.0414060615,
-0.0194881205,
-0.0065755295,
-0.0415562764,
-0.0028088586,
0.0275760256,
-0.0126301637,
0.0377794914,
-0.0720550865,
0.0231517386,
-0.0414151102,
0.052152887,
0.053888537,
-0.0173607189,
-0.01251915,
0.0276242848,
0.029384587,
-0.0260197688,
0.0071734232,
-0.0230929051,
0.0136860982,
0.0238494091,
0.0411074758,
0.0078909099,
0.0220542625,
-0.0407118425,
-0.0066413507,
-0.0436970629,
0.0273001585,
0.0439826697,
0.0457614921,
-0.0123475557,
-0.0049289255,
-0.0732556283,
0.0113181779,
-0.0448525809,
0.0295499191,
0.0029244362,
0.0217314791,
0.0108319372,
0.0017630201,
0.0264036283,
0.0103686946,
0.0093023824,
0.0504584983,
0.0294052158,
-0.0141999051,
-0.0106621385,
-0.0473555923,
0.0216078032,
-0.0314702354,
-0.0472616889,
-0.0722358674,
0.0209528562,
0.0107857659,
-0.0169062205,
-0.0258016419,
0.0003469062,
-0.0049154041,
-0.0220645536,
0.0550586879,
-0.0230516996,
0.014183674,
-0.0157212857,
0.021655526,
-0.0367070101,
0.0255462304,
0.0069416342,
0.0584976487,
0.0725378692,
-0.0006768671,
0.0501657277,
0.0115244035,
0.0059938664,
0.02589106,
0.038563408,
0.0036435388,
-0.036991179,
0.0162295289,
-0.0118655162,
0.0598732494,
-0.0119818933,
-0.0063998252,
0.1264476031,
0.0627990663,
-0.0136023676,
0.0410810523,
0.000094168,
0.0200009942,
-0.0372064635,
-0.0020621934,
-0.0064856922,
0.0490099043,
0.0071686055,
-0.0182529259,
0.0659270734,
-0.027982384,
-0.0302581023,
-0.0408438258,
-0.0043769279,
0.014770532,
0.0137994764,
0.0186491273,
0.0042590294,
-0.0086447494,
-0.0197163057,
0.0042960877,
0.0407673568,
-0.0466310158,
0.0574505106,
-0.0395619795,
0.0569407828,
0.0894694626,
0.0504821315,
0.0278617516,
0.0660279617,
-0.022805512,
-0.0418885015,
0.014859112,
0.0497444347,
-0.0261432342,
-0.0407426581,
-0.0167984162,
-0.0515544079,
0.0261224471,
0.0165660083,
-0.0038295744,
0.0019841883,
-0.0247936528,
-0.0166228972,
-0.0132278008,
-0.0853282437,
-0.011478072,
0.0467936546,
-0.0101413075,
0.0160461403,
0.0006529322,
0.0473921224,
0.0268129595,
0.0071421396,
0.0095947664,
0.0176120456,
-0.0028801931,
-0.1012399346,
-0.0024579277,
-0.0013752516,
-0.0011217535,
-0.0237334557,
-0.0378347151,
0.0240443032,
0.0131932842,
-0.0018834075,
-0.0065464587,
0.0232874248,
0.0165828578,
0.0499487482,
-0.0078469543,
-0.0118667865,
-0.0099939555,
-0.0567721948,
-0.023302421,
0.0481632873,
-0.0122559294,
0.0200339705,
-0.0221772809,
-0.0247241091,
-0.0450346544,
0.0201118961,
-0.0053885086,
0.0023589372,
0.0506266318,
0.0390884839,
0.0668252856,
-0.0091319885,
-0.0257037766,
-0.0665505379,
-0.0677675456,
-0.0376049057,
-0.0210771449,
0.0239414833,
0.0366478115,
-0.0523836873,
0.04375479,
0.0286225416,
-0.0834755749,
0.0074387784,
0.0420607179,
0.0196545627,
0.033324264,
0.0424156711,
0.0159021821,
-0.027933009,
-0.0220129006,
-0.0063342606,
0.0259956736,
0.0769443959,
0.0018661133,
-0.0626971275,
0.0249694325,
-0.0509376228,
-0.0014018933,
-0.0018488278,
0.0012104657,
-0.0007784697,
-0.0000725349,
0.0068985387,
0.01242611,
0.0018694395,
0.0049368972,
0.0411088094,
0.0038305423,
-0.0596372262,
-0.0296928417,
-0.0576830953,
-0.0254914854,
-0.0523616448,
-0.0659912825,
0.0201087911,
0.0201923791,
0.0261941589,
-0.0344016477,
-0.0642345175,
0.0409944467,
-0.0297569335,
-0.0100908214,
0.0053567402,
-0.0313269608,
0.0045926413,
0.0075561251,
-0.022364568,
-0.0548402853,
0.0714576393,
-0.0328497626,
0.0139384046,
-0.0274296366,
0.0102885226,
-0.0239055622,
0.0773638412,
-0.0071373582,
0.0113455737,
-0.0451982953,
0.0350060128,
0.011647569,
0.0125254542,
0.100798592,
-0.0206598416,
0.0211735237,
0.001258817,
0.0065240264,
0.0031931109,
-0.0488673858,
0.0230929423,
-0.0979748741,
-0.0204478651,
0.0009468438,
0.0444011502,
-0.0105961859,
0.003063305,
0.0080479588,
-0.0135937911,
0.0273119677,
-0.0477394722,
-0.0357790701,
0.0071769548,
-0.0043475968,
0.0363148004,
-0.0503733829,
0.0226290599,
-0.0762918964,
0.0006671228,
0.0623457842,
0.0209438615,
-0.0110097602,
0.0181932263,
-0.0519316532,
0.0234017931,
-0.0087675834,
-0.0056688315,
0.0069617499,
0.0091090146,
-0.0473609902,
0.017575277,
0.079817839,
0.0074181589,
-0.1259594858,
-0.0171220638,
0.005522477,
0.0479178391,
0.0074376781,
-0.0073070284,
-0.0745370165,
-0.0195920784,
-0.0277864467,
0.0309699401,
0.0057343054,
-0.0395832136,
-0.0030585735,
-0.1056899354,
0.0261531249,
-0.0137966769,
0.0324214138,
-0.0209753495,
-0.0350717641,
-0.0147468001,
-0.0189225674,
-0.0291679427,
0.0075513078,
0.0036236437,
0.0117436768,
-0.011392598,
-0.0287762769,
0.0230591092,
0.0598748401,
-0.0553172752,
0.0663766265,
0.045113489,
0.0091978097,
-0.0225892998,
-0.0353984311,
-0.0087274471,
-0.0003963973,
-0.0232813824,
0.0047255978,
0.0716382191,
-0.0275339261,
0.0284414552,
-0.0127163408,
0.0163138695,
0.007606735,
0.055228956,
0.000899963,
0.0143181011,
0.0286308266,
-0.0731964409,
0.0191073343,
0.0244928449,
0.0117171286,
0.0770807415,
-0.0033693826,
-0.0609159842,
-0.0156842582,
-0.0270952992,
-0.0141912075,
-0.0336232707,
-0.0613211691,
-0.0153955976,
0.0027468742,
-0.0242142901,
-0.0818626359,
-0.0324712023,
-0.0140137924,
-0.0310211144,
0.0449401736,
0.0227258671,
0.0161146969,
-0.0031022632,
0.0454250686,
0.0073528872,
-0.0267325249,
-0.0007113534,
-0.0157698933,
-0.0089672785,
0.0212455131,
-0.0250943471,
0.0119997812,
-0.0133252963,
-0.0794895366,
-0.0298917796,
-0.0437556207,
0.0886889845,
0.0141014019,
0.0233933497,
-0.0198381972,
-0.0766419098,
0.0057491153,
-0.0083046686,
-0.0253614318,
0.0041430537,
0.0255188812,
0.0211620834,
0.0109004611,
0.0383256078,
-0.0536178537,
0.0173094943,
0.0239820648,
-0.0003298704,
-0.021391863,
-0.0283646602,
0.0198108386,
0.0135714067,
0.022363035,
-0.0168711003,
0.0229529534,
-0.0294606928,
-0.0033615842,
-0.0301454663,
-0.0402825624,
-0.0368716493,
0.000958602,
0.0004021129,
0.0190676004,
-0.0163880289,
0.01928683,
-0.006643686,
0.0606928878,
-0.0826339796,
0.0371240191,
0.0113645745,
-0.0170816835,
0.0030946112,
-0.0039582946,
-0.0294263847,
-0.0223765112,
0.0125902221,
0.0368949249,
0.0053032297,
0.0337548032,
0.0087793414,
-0.0160660855,
0.0241333097,
0.0201886799,
-0.0639699548,
-0.0225748233,
-0.0233136714,
0.0154121444,
0.0029975639,
0.0484522097,
0.0564863048,
-0.0522289611,
0.0071150749,
0.0122719975,
0.0056384425,
-0.0147456918,
-0.0349114612,
0.0109045701,
0.0048086667,
-0.0164475031,
0.011330151,
0.0713806003,
-0.0034309777,
0.0010251143,
-0.0267114304,
-0.0565332249,
-0.0088970466,
-0.0172978248,
-0.0037664035,
-0.0002326524,
-0.0004550237,
-0.0201837718,
-0.0618220754,
-0.0172935892,
-0.0163644459,
0.0356753357,
0.0338398814,
-0.0302703101,
0.0053002709,
0.0635751858,
-0.0179158412,
-0.0004535255,
-0.0823394433,
-0.0104414485,
-0.0093082814,
0.0040788613,
-0.0282713212,
-6.531020165e-33,
0.0290441401,
-0.0581235588,
-0.0164172836,
0.0284393262,
-0.0029929834,
0.0148107111,
-0.0227240101,
0.007342319,
0.0299855508,
0.0295025855,
-0.0246304125,
-0.0051959311,
-0.0003560189,
-0.0041954438,
0.0441695862,
-0.0295562111,
0.0508731157,
-0.0373763368,
-0.0370623469,
0.0438853242,
0.0096992645,
-0.0228072479,
0.0407025069,
-0.0274536982,
0.0901844949,
-0.0462864228,
-0.0644346103,
0.0488916487,
-0.0044415481,
0.0171236489,
0.0178259723,
0.0593129583,
0.028856881,
-0.0527364276,
0.0108893355,
-0.0055993884,
-0.0353815556,
0.0300492868,
0.0351367369,
-0.0416255221,
-0.0424866565,
0.0123501476,
0.0496520661,
0.0185675975,
-0.0747566372,
0.0032839477,
0.0362615064,
0.0047208173,
0.0137236631,
-0.0404845215,
-0.0116602061,
0.0060416302,
0.0165691432,
0.0186330844,
0.013104151,
0.0585422292,
-0.0072519314,
0.0307461303,
-0.0264445208,
0.0653227046,
-0.0035305789,
0.0608641021,
0.0151912048,
-0.0143926553,
0.0047739525,
0.0728384182,
-0.0256697647,
0.0674491376,
-0.0176882148,
-0.0346185006,
-0.0128038973,
0.0163914096,
0.0632387251,
0.0858195275,
0.0056510367,
-0.0442920513,
-0.049602855,
0.0397114605,
0.0687025338,
-0.045527909,
0.0626412481,
-0.0301156398,
-0.0533859581,
-0.0539480969,
-0.0230581388,
-0.0568544269,
-0.0359796211,
-0.0275726039,
-0.0176104438,
0.0465545617,
-0.0208223443,
-0.0222816747,
-0.0032308616,
-0.0339123458,
-0.055979792,
0.0114036305,
-0.0129892118,
0.0155147938,
-0.0103897825,
-0.023556035,
-0.0699061602,
0.031406749,
-0.069656305,
0.0164339114,
-0.0151720075,
0.0165723097,
0.0443763249,
0.0183381625,
-0.0507876128,
0.0106331529,
-0.020957483,
0.0291975625,
-0.0132183852,
-0.0336506404,
-0.0037464644,
-0.0086955605,
0.0038396167,
0.0407794192,
-0.005117076,
0.0681636184,
-0.0388279781,
0.0163937192,
-0.0154197598,
-0.0335471369,
-0.0106493691,
-0.0027251388,
-0.0214667302,
-0.0224724151,
0.033001665,
-0.0240244176,
-0.0167508032,
-0.0200526714,
2.758e-7,
0.0014434515,
0.0683762431,
-0.0495389439,
-0.0305420682,
0.0344496742,
0.0716488734,
0.0264627766,
0.0359026529,
0.0335913599,
-0.0429781675,
0.023450464,
0.0088351108,
0.0126519911,
0.0827429816,
-0.0964356065,
-0.0135216741,
-0.104585357,
-0.0084255142,
-0.007738722,
-0.0529442616,
0.0127586685,
0.0191676617,
0.0281439684,
0.0072042416,
-0.0133621776,
-0.0424823016,
0.0580734126,
-0.0500389189,
-0.062620841,
0.0219750144,
0.0346787199,
-0.0587635487,
-0.01103942,
0.0129369702,
-0.024741035,
-0.0076411245,
-0.0057975668,
0.0578411557,
0.0291567519,
-0.0230809972,
-0.0042414172,
-0.0127669526,
0.0202399809,
-0.0256068837,
0.0299384668,
0.0024989219,
-0.0393244252,
-0.0424924307,
0.0172526762,
-0.0467114784,
0.0205430146,
-0.0203078371,
-0.0482937507,
0.019572679,
-0.02185487,
0.0156171769,
0.0017806656,
-0.0289519709,
-0.0342705511,
-0.0243993681,
-0.0107782139,
-0.1140248254,
0.0158806909,
0.0101076942,
-0.0413791351,
-0.0479733609,
0.0394780152,
2.600957116e-34,
0.0537882447,
-0.0205544066,
0.0228105802,
-0.0171519481,
0.0577934422,
0.0438680835,
-0.0389525183,
-0.0218142178,
0.0578500181,
-0.0460083298,
-0.0229568165
] |
http://arxiv.org/abs/2405.15071v2
|
http://arxiv.org/pdf/2405.15071v2
|
Grokked Transformers are Implicit Reasoners: A Mechanistic Journey to the Edge of Generalization
| 2024-05-23T21:42:19 | 2024-05-27T03:55:35 |
Boshi Wang, Xiang Yue, Yu Su, Huan Sun
|
cs.CL
|
cs.CL
|
We study whether transformers can learn to implicitly reason over parametric
knowledge, a skill that even the most capable language models struggle with.
Focusing on two representative reasoning types, composition and comparison, we
consistently find that transformers can learn implicit reasoning, but only
through grokking, i.e., extended training far beyond overfitting. The levels of
generalization also vary across reasoning types: when faced with
out-of-distribution examples, transformers fail to systematically generalize
for composition but succeed for comparison. We delve into the model's internals
throughout training, conducting analytical experiments that reveal: 1) the
mechanism behind grokking, such as the formation of the generalizing circuit
and its relation to the relative efficiency of generalizing and memorizing
circuits, and 2) the connection between systematicity and the configuration of
the generalizing circuit. Our findings guide data and training setup to better
induce implicit reasoning and suggest potential improvements to the transformer
architecture, such as encouraging cross-layer knowledge sharing. Furthermore,
we demonstrate that for a challenging reasoning task with a large search space,
GPT-4-Turbo and Gemini-1.5-Pro based on non-parametric memory fail badly
regardless of prompting styles or retrieval augmentation, while a fully grokked
transformer can achieve near-perfect accuracy, showcasing the power of
parametric memory for complex reasoning.
|
Grokked Transformers are Implicit Reasoners: A Mechanistic Journey to the Edge of Generalization: We study whether transformers can learn to implicitly reason over parametric
knowledge, a skill that even the most capable language models struggle with.
Focusing on two representative reasoning types, composition and comparison, we
consistently find that transformers can learn implicit reasoning, but only
through grokking, i.e., extended training far beyond overfitting. The levels of
generalization also vary across reasoning types: when faced with
out-of-distribution examples, transformers fail to systematically generalize
for composition but succeed for comparison. We delve into the model's internals
throughout training, conducting analytical experiments that reveal: 1) the
mechanism behind grokking, such as the formation of the generalizing circuit
and its relation to the relative efficiency of generalizing and memorizing
circuits, and 2) the connection between systematicity and the configuration of
the generalizing circuit. Our findings guide data and training setup to better
induce implicit reasoning and suggest potential improvements to the transformer
architecture, such as encouraging cross-layer knowledge sharing. Furthermore,
we demonstrate that for a challenging reasoning task with a large search space,
GPT-4-Turbo and Gemini-1.5-Pro based on non-parametric memory fail badly
regardless of prompting styles or retrieval augmentation, while a fully grokked
transformer can achieve near-perfect accuracy, showcasing the power of
parametric memory for complex reasoning.
|
[
0.041346807,
0.0437870324,
-0.0235408545,
0.0035405513,
-0.0109466789,
0.0162214413,
0.0063323001,
-0.0092845587,
-0.0522276349,
-0.0447430983,
-0.0535324588,
-0.0051906481,
0.0080510592,
-0.0011890203,
0.0378054008,
-0.0804713219,
0.0496778004,
-0.0538504906,
0.0050850064,
-0.0016552013,
-0.0332013629,
0.0077987537,
-0.0283340421,
0.0465868823,
0.0042384551,
-0.0059176665,
-0.0271029547,
-0.0297433138,
0.0088776601,
0.0029614256,
-0.0039548804,
0.0241794065,
0.0123057691,
0.0512357168,
0.000002021,
-0.0082654376,
0.0108883185,
-0.0218602419,
-0.0336949341,
-0.0235956721,
-0.0322398096,
-0.0002251777,
-0.0108604375,
-0.0056269844,
-0.038847331,
-0.0440423712,
0.054636728,
0.0819215626,
0.0472132899,
0.0633136556,
-0.0001400631,
-0.0440270081,
0.0293565057,
-0.0178472623,
0.0272038486,
-0.0039418354,
-0.0249233786,
-0.0193761643,
0.004931089,
-0.0084167225,
-0.0121904779,
0.0060768891,
0.0585378073,
0.0022595387,
-0.0206705462,
0.0559403263,
-0.0451128669,
-0.0244878344,
0.0077795093,
0.0263908152,
-0.0010955429,
-0.0420608446,
-0.002064612,
0.0121883936,
0.0074077146,
0.0040410277,
-0.0339365788,
0.0407758728,
-0.0274867751,
-0.0064396113,
-0.0184719283,
-0.0156722721,
0.036831256,
-0.0074644941,
-0.0369415954,
0.0895707458,
0.0067999652,
-0.022572251,
0.0126246484,
-0.0029990857,
-0.0323446393,
-0.0482804924,
-0.0026629474,
0.0528539456,
-0.0340457745,
0.0179772843,
-0.0149626592,
0.0076622018,
-0.0144464141,
0.0196871627,
0.0186381768,
0.0568047613,
0.0058095194,
0.0114945453,
-0.0529292896,
0.0756360516,
-0.0383461006,
0.0939646736,
-0.0337891318,
-0.0158231445,
-0.0656608343,
-0.0594673418,
-0.0402885787,
0.0443512201,
0.0262057334,
-0.0118907262,
-0.0355617478,
0.0319377817,
-0.0052996334,
0.0196906514,
0.0115043828,
0.0243400149,
-0.0395518057,
-0.0215592925,
-0.0440479666,
0.0230290405,
-0.0426601693,
0.027734451,
0.0491757616,
-0.0362885483,
-0.0274936054,
-0.033875715,
0.0046107732,
-0.0465785153,
-0.0049766786,
0.0735877305,
-0.0374981128,
-0.0417397246,
-0.049895592,
-0.0146703525,
-0.0001708888,
-0.0707149357,
0.0282036047,
-0.0254691411,
0.0087759485,
0.0079468628,
-0.0072261621,
-0.0004739264,
0.0004558729,
0.0436984114,
-0.0067957975,
0.0776484534,
0.0495476685,
0.0027088199,
0.0132752163,
0.0213232245,
-0.0464496166,
0.0512427166,
0.0154053578,
-0.0042206589,
-0.0396483839,
-0.0166225359,
0.0184539072,
-0.0289765,
0.0313688181,
-0.0597608499,
-0.0101616737,
-0.0120126428,
-0.0197550356,
0.0325417481,
0.0434775017,
0.0558192693,
-0.0454068184,
-0.0318848714,
0.0643090159,
0.0559476092,
0.0334034413,
0.0456453599,
0.0500922017,
-0.0271062516,
0.0141726024,
0.0509413071,
-0.0876165107,
0.0623718798,
-0.0276047513,
0.0269952789,
-0.0228369683,
0.0094646886,
-0.0628658459,
-0.0483858958,
-0.0099572912,
0.0075022294,
-0.0012963479,
0.0065551321,
0.0527288355,
0.062919043,
-0.0554226451,
0.0708119422,
-0.005814068,
-0.0406000428,
-0.0360486358,
-0.0640147328,
0.0425168425,
0.020947177,
0.0276506282,
-0.0053569698,
-0.0080409823,
0.0046102004,
-0.0271336436,
0.026722407,
0.0546825379,
-0.0035867791,
0.0219280347,
-0.0476386361,
-0.0163675752,
-0.046686992,
0.0244717449,
0.0177637301,
0.0256395191,
-0.0240268409,
0.0258361008,
-0.0303741042,
0.0347224399,
0.0073469426,
-0.0259111244,
-0.0351085737,
-0.0366616175,
0.0136095388,
-0.0194699336,
-0.0245455615,
0.0235150084,
0.0360938683,
0.0153803593,
-0.0263555273,
-0.0523771346,
0.0419223309,
-0.0928872749,
-0.0225729719,
-0.0132140452,
-0.0562578626,
-0.0163638666,
0.0239202902,
-0.004740878,
-0.0612889268,
0.0522255339,
-0.0457367338,
0.0947092026,
-0.0353479646,
0.011362398,
0.0144449407,
-0.019973157,
-0.0089693237,
0.0802160278,
0.0102425739,
-0.0188476518,
0.0008578095,
-0.0042423294,
-0.0084497761,
-0.0313823521,
-0.0034914499,
0.0403633192,
-0.0145279327,
-0.0033500739,
-0.0179901421,
-0.0612409785,
-0.0594016537,
-0.044251509,
-0.059397351,
-0.0090369191,
0.0057873442,
-0.006300943,
0.0122077875,
0.0093299542,
0.0233645216,
0.0014077777,
-0.0151062496,
0.0016840195,
0.0062606018,
0.0450745709,
0.0331416652,
0.0122697558,
-0.0020803565,
-0.0152684469,
-0.0249283109,
-0.0250728503,
0.0592244156,
0.0001476189,
0.0772458613,
-0.0336505249,
-0.0001693581,
-0.0299717598,
0.0147983357,
-0.0283847563,
-0.0491961427,
0.0549480207,
-0.0169128515,
0.0238533318,
0.0638218075,
0.0083340108,
-0.0154074663,
0.0398332663,
-0.0402795523,
-0.0339568369,
0.0453389548,
-0.023763651,
-0.0523396358,
-0.0091716107,
0.001774626,
-0.0357788391,
0.0967861563,
-0.0148352748,
0.0030008378,
-0.0878474638,
0.021613162,
-0.0083799511,
0.0054832171,
-0.0207201652,
-0.0358591676,
-0.0074587059,
0.0409828573,
-0.0184705574,
-0.0309953019,
0.1003790349,
0.0046593184,
-0.0203582253,
-0.0317946151,
0.0282077771,
-0.0133282999,
0.0792367384,
-0.0235777795,
-0.0014669584,
0.0313532911,
-0.0179230236,
-0.0675479919,
-0.0235131308,
0.0397929549,
-0.0027488484,
0.083037287,
-0.0176851414,
0.0088262083,
-0.0075554815,
-0.0612607412,
-0.0619177967,
0.0225700624,
-0.0334496759,
0.0102505572,
0.0011036947,
-0.0538377389,
-0.0444125868,
0.0067675286,
0.0340944417,
0.0514833331,
-0.0056227162,
-0.0456169397,
-0.0105822803,
-0.007358666,
-0.0016456957,
0.0378041752,
0.0535289608,
-0.0350413285,
0.02779961,
0.0439990759,
0.054014422,
-0.0097852303,
-0.0312885232,
0.029236434,
-0.0092744473,
0.0035279309,
-0.0259102937,
-0.0106002558,
0.036417082,
-0.0799702108,
-0.0256980844,
0.0811156482,
0.0254191421,
-0.0516323559,
0.0094089676,
-0.0441061445,
0.0480049327,
0.035006538,
0.0419329107,
-0.0400754325,
-0.0267151948,
0.0347983316,
-0.0070692394,
-0.0212202147,
-0.0119655579,
-0.0125275441,
-0.0999607965,
0.0025320693,
0.0057567493,
-0.0243325997,
-0.0124703925,
-0.024307929,
-0.0208195429,
-0.0114062577,
0.0119161429,
0.0193142444,
-0.0254253466,
-0.0177792441,
0.0032704116,
0.0084671089,
-0.0238660034,
0.0502802283,
0.0093648769,
0.0055737728,
0.0310204141,
0.0176283922,
0.0227999669,
-0.0687514171,
0.0472451784,
-0.016352227,
0.0495179631,
0.0122092841,
0.0499996282,
0.0277944524,
-0.0150153581,
-0.0361708067,
0.0153071722,
-0.0331458636,
-0.0391772464,
-0.0435505211,
0.0749974549,
-0.0323083624,
-0.0559802577,
0.0125857256,
0.0117838606,
-0.04099112,
0.056501057,
0.0193209704,
0.0254959408,
-0.0687955916,
-0.0429696776,
0.0036045955,
-0.0493701473,
-0.0509019084,
0.0466869324,
0.0885008574,
-0.0246756282,
-0.057516057,
-0.0360017642,
-0.0015282041,
0.0496247411,
-0.0219030287,
-0.0448241718,
-0.0154944276,
-0.0056427587,
-0.0387706645,
0.0374484248,
0.0016292931,
-0.003818993,
-0.0466792472,
-0.0114578186,
-0.0101481695,
-0.0544901863,
-0.0272934195,
0.0152420849,
-0.0544546396,
-0.070662275,
0.0252164733,
0.0560242943,
0.0351532698,
0.0392542928,
0.0081320424,
-0.0083870078,
0.0395707227,
0.0012852633,
-0.0258168187,
0.0119205862,
0.069076851,
-0.0349254683,
0.0155821638,
0.0460693911,
-0.0140095968,
0.0192033686,
0.0232815072,
0.0083157904,
-0.0076363208,
0.0760805234,
0.0620432384,
-0.0279269069,
0.032450784,
0.0371853597,
0.0242779106,
-0.0207304358,
0.0404364057,
-0.0128895929,
0.0100857131,
-0.0290923323,
0.0007990415,
0.0189256836,
0.0085730739,
0.0152569767,
-0.0272147935,
0.0211101882,
0.0432544537,
0.0192382373,
0.0040319008,
0.053438589,
0.0086553395,
0.0090164421,
0.0130013442,
-0.0157017317,
0.0289078299,
-0.0117065506,
0.0556659698,
-0.0109942174,
0.0627248883,
0.0381760597,
-0.0058430675,
-0.0534382425,
-0.0074155056,
-0.1255329996,
-0.0828967765,
0.0404407643,
-0.0226978455,
0.0156499036,
-0.0266478956,
0.0125402426,
-0.0433443524,
0.004411147,
0.0256041735,
0.0188651122,
0.0118808318,
-0.0027791501,
0.0294270068,
0.0409375094,
0.0219967235,
-0.0273816474,
0.0600692779,
-0.0423668511,
0.0572196916,
-0.0374306105,
0.0171555299,
-0.0247187223,
-0.0367827229,
0.001080754,
0.0332294665,
-0.0016149771,
-0.0075548328,
-0.0438907221,
-0.0626474246,
-0.0287120603,
0.0065412149,
-0.0016457816,
-0.0294344835,
0.0365687944,
0.0640314668,
-0.0088110128,
-0.0597815253,
0.0133149009,
-0.0069230734,
0.0120686935,
0.0387990996,
-0.0202823747,
-6.524813554e-33,
-0.026780542,
-0.0065713096,
0.0482336804,
-0.0123920012,
-0.0180385802,
0.011753845,
-0.05219464,
-0.019477075,
-0.0428892858,
0.0172261819,
-0.0330334492,
0.0397035629,
0.0117318863,
0.0127671035,
0.0351848751,
-0.0277239401,
0.0338863246,
0.0436705835,
-0.0511084273,
0.0277754515,
0.05629557,
0.0136755016,
0.0849521905,
-0.0169567578,
0.0006418311,
0.0297512617,
-0.0195069797,
0.0110895243,
-0.083119534,
0.0807727948,
-0.0114679122,
0.0057704244,
0.0021518101,
-0.0205699354,
-0.0193531476,
0.0182257351,
-0.0850578547,
-0.0363291167,
0.0113065755,
-0.0016133045,
-0.0480738021,
-0.03830662,
0.0298926421,
-0.0241235755,
0.0016705754,
-0.0616332963,
0.0092158513,
-0.0049530934,
0.019334916,
0.015343775,
-0.0198415872,
-0.0052914973,
-0.0074025611,
-0.0220654048,
0.0050141448,
0.0476122722,
-0.0193040036,
0.0262882169,
-0.0528627448,
0.0307127368,
-0.0079064341,
0.0180596113,
0.0521607846,
0.0184320584,
-0.0105803506,
0.0498460568,
0.0191773362,
0.0090004681,
-0.0286066625,
-0.0365891978,
0.0032960325,
0.0659386292,
0.0249151532,
-0.0203963891,
0.0699556693,
-0.015484415,
0.0052929749,
-0.0247774106,
-0.0083613042,
0.0746689513,
0.0397544354,
0.0161075331,
-0.0185382236,
-0.0254104938,
-0.0519700088,
0.0059531811,
-0.0540346466,
-0.0247525685,
0.0131999087,
0.0353410728,
0.0194282029,
-0.0086443899,
-0.0025548029,
-0.0109434705,
0.0339587145,
0.0167900976,
-0.0089684073,
0.0059048845,
-0.0211100206,
-0.0025247687,
-0.0411152132,
-0.0154638803,
0.0136266174,
0.0355153158,
-0.0041704834,
0.0445495285,
0.0541114919,
0.0696066469,
-0.0375362448,
-0.033984676,
-0.0613618046,
0.010390183,
-0.0227583535,
-0.0112848394,
0.0278499033,
0.0714441761,
-0.0189998485,
0.0183161851,
-0.0300685558,
0.0046047852,
-0.0437850393,
0.051163163,
0.00963837,
0.0368058756,
0.0308595039,
-0.025297191,
-0.0394269638,
-0.030521106,
0.0362327434,
-0.0651028752,
-0.01583427,
0.043722596,
2.808e-7,
0.0222881101,
0.0299227946,
0.0289943125,
0.0142542738,
-0.0055527799,
0.0254835375,
-0.0008485733,
-0.0215914249,
0.005624224,
-0.0193721931,
-0.0063006068,
0.047256764,
0.0461003371,
0.0564005822,
-0.066797182,
-0.0090038329,
-0.061011076,
-0.0403996594,
-0.0172120817,
-0.0095483279,
0.0511592217,
0.0817061588,
0.0293732323,
-0.0196128748,
0.0289197154,
-0.0326506868,
0.0741777569,
-0.002786576,
0.0395182334,
0.0062134154,
0.0228511859,
-0.004488524,
0.0293848608,
0.0638289601,
0.0113964779,
-0.0066885478,
-0.0129337506,
0.0394018665,
-0.0015961468,
-0.0556508303,
-0.0239626151,
-0.0050591072,
-0.006297092,
-0.0752712563,
0.0508261621,
-0.0821716562,
-0.0230632033,
0.0216993392,
0.021607304,
-0.010313265,
0.0109437164,
-0.0033421898,
-0.0034771559,
0.0194658842,
-0.0095946873,
0.0215771571,
0.0238883458,
-0.1039294079,
-0.040742036,
0.0424492657,
-0.0455431715,
-0.0133206677,
-0.0513125025,
-0.029978212,
0.0498218089,
-0.0129062347,
-0.0309848264,
2.489570532e-34,
0.0224655699,
-0.0008221262,
0.0167342424,
0.0274249036,
0.0396184176,
-0.0128287496,
0.0369440317,
-0.0136183733,
0.0410397016,
-0.0722721964,
-0.0281772539
] |
http://arxiv.org/abs/2405.15070v1
|
http://arxiv.org/pdf/2405.15070v1
|
Optimizing example selection for retrieval-augmented machine translation with translation memories
| 2024-05-23T21:42:03 | 2024-05-23T21:42:03 |
Maxime Bouthors, Josep Crego, François Yvon
|
cs.CL
|
cs.CL
|
Retrieval-augmented machine translation leverages examples from a translation
memory by retrieving similar instances. These examples are used to condition
the predictions of a neural decoder. We aim to improve the upstream retrieval
step and consider a fixed downstream edit-based model: the multi-Levenshtein
Transformer. The task consists of finding a set of examples that maximizes the
overall coverage of the source sentence. To this end, we rely on the theory of
submodular functions and explore new algorithms to optimize this coverage. We
evaluate the resulting performance gains for the machine translation task.
|
Optimizing example selection for retrieval-augmented machine translation with translation memories: Retrieval-augmented machine translation leverages examples from a translation
memory by retrieving similar instances. These examples are used to condition
the predictions of a neural decoder. We aim to improve the upstream retrieval
step and consider a fixed downstream edit-based model: the multi-Levenshtein
Transformer. The task consists of finding a set of examples that maximizes the
overall coverage of the source sentence. To this end, we rely on the theory of
submodular functions and explore new algorithms to optimize this coverage. We
evaluate the resulting performance gains for the machine translation task.
|
[
0.0584691428,
0.0482795313,
-0.0223036036,
0.016177848,
-0.0414704531,
0.0065534287,
0.0082140667,
0.0282116309,
-0.0218857396,
-0.0528393388,
0.0009701248,
-0.0515398085,
-0.0165153109,
-0.0438426845,
0.0655490011,
-0.0359700359,
0.0495081916,
0.0084118219,
-0.0673033372,
0.0144028217,
-0.0158211831,
-0.0044435859,
0.0185520202,
-0.0049548307,
-0.0157174207,
-0.0158482231,
0.0163840614,
0.0165916365,
0.0477424972,
0.0220753625,
-0.0051892926,
-0.0221864246,
0.0178815369,
0.0514822006,
0.000001773,
-0.0134609984,
-0.0895226747,
-0.0000301855,
-0.0191422123,
-0.0334334411,
0.0313989557,
-0.0249960888,
-0.0167025402,
0.0317643881,
-0.0047819656,
0.0088383993,
0.0685068071,
0.000333971,
0.0666072071,
0.1197663993,
-0.0166709423,
-0.0282441564,
0.0017715654,
-0.0220546182,
0.0554935336,
-0.0101749375,
0.0092542572,
0.0130115235,
-0.0409502536,
-0.0365702473,
0.0111625735,
0.0613244101,
-0.0010305392,
0.0016841713,
-0.00907165,
-0.0362505615,
-0.0549548641,
-0.0238316115,
0.0009626242,
0.0144534688,
-0.0479452908,
-0.0222721118,
-0.0337127149,
0.0086448705,
-0.0122756092,
0.0269416403,
-0.0455373488,
-0.0165490136,
-0.0126683833,
0.0107619483,
0.0334944651,
0.0172536392,
0.0297553781,
0.0060537015,
-0.0782758072,
-0.0300096292,
0.0390071943,
-0.0520524867,
-0.0268275011,
0.0149066197,
-0.0270128902,
0.0252354387,
0.0824407339,
-0.0012807605,
0.0147881946,
0.0481689163,
0.0065900767,
-0.106514737,
-0.0200478788,
-0.00417729,
0.053544227,
0.0741185918,
0.0338568203,
0.0118558845,
-0.0321911573,
0.0469440036,
-0.0414214395,
-0.017016476,
-0.055342,
-0.0276985485,
-0.0558745973,
0.0088680303,
-0.0757103413,
0.0408001505,
0.0106623992,
0.0465285294,
-0.0316551179,
0.0488752462,
0.0393928401,
-0.0014345181,
0.0008494482,
-0.0365099385,
-0.0972011387,
-0.0086743282,
-0.0039121364,
0.0522333495,
0.0092388252,
-0.0500768162,
0.0231506582,
-0.068863757,
-0.0536798574,
0.0061184876,
0.0164647419,
-0.086841628,
-0.0318171643,
0.0858236104,
-0.0233150683,
-0.0813774094,
-0.0202951673,
-0.0206109006,
-0.005274049,
-0.0244965851,
0.038904801,
-0.0393044427,
0.0626826882,
-0.0096398406,
0.0035125294,
-0.0134623814,
-0.0027152672,
0.0049034748,
-0.0192132723,
0.0363144986,
0.0552186444,
-0.038441848,
-0.0022042836,
-0.0075823218,
-0.0047310875,
0.0393253714,
0.0105846711,
0.0407864749,
0.0450427234,
0.0245235953,
0.023707753,
0.0575635731,
0.0220113583,
-0.0147972126,
0.0191564467,
0.0029310128,
0.0157744624,
0.0390168987,
0.028243918,
0.0174726769,
-0.0591484979,
0.0005159363,
0.0340878479,
0.025537204,
0.0216071978,
0.0397396721,
0.0343042351,
0.0037143782,
-0.0119129401,
0.0719240382,
0.0279444307,
0.0269053429,
-0.0315303802,
-0.0005143115,
0.0408119895,
0.0244720131,
-0.0127476379,
-0.0098063722,
-0.0163873546,
-0.0203101393,
0.0056980196,
-0.0157288481,
0.0202093758,
0.0359408893,
-0.08799164,
0.0939463153,
-0.0386130586,
-0.054028973,
0.0025923895,
-0.0277592819,
-0.0465132147,
0.0467273742,
0.0619844794,
-0.0543913394,
-0.033937905,
0.0030907241,
-0.0438572578,
0.0339672528,
0.0581534281,
-0.0010374347,
0.0408566594,
-0.0709118694,
-0.0031863176,
0.0130223921,
-0.0109381676,
0.0236570109,
0.0096697444,
-0.0174139533,
0.0311213359,
-0.0218389276,
-0.0206110943,
0.0181897189,
0.0047568134,
-0.0350316912,
0.0039037652,
0.0325852633,
0.03940529,
-0.0185964033,
-0.0073057804,
0.0220961496,
0.0128117548,
-0.0770019963,
-0.0165174846,
0.0158152375,
-0.035624139,
-0.1034721732,
0.0006764405,
-0.0580973625,
-0.0370794274,
0.0051823179,
-0.0270818397,
-0.0399641842,
0.0140401563,
-0.072989516,
0.0583455004,
-0.038804166,
0.0353134461,
0.0385327078,
-0.032676369,
-0.0539726131,
0.029753888,
-0.0132935438,
0.0134828938,
-0.0228166655,
-0.0751319826,
0.0249836706,
-0.0563907921,
-0.006818695,
0.0249283426,
0.0114709008,
0.0436455533,
0.0403175093,
-0.0014123081,
0.010363928,
-0.0682598278,
-0.0010587188,
0.0003434158,
-0.0191055797,
0.0385992043,
-0.0015802549,
0.0726646259,
0.0358386822,
-0.0038659084,
-0.004795522,
0.0241511073,
-0.009730706,
0.0424019508,
0.0312695801,
0.0030003588,
0.0286407415,
-0.0501674116,
-0.0371741503,
-0.0005402467,
0.0979550183,
-0.0436075926,
-0.0239256807,
-0.0283329897,
0.010438798,
-0.0385443829,
0.010645465,
0.0009710974,
-0.0558336116,
-0.0150995599,
-0.0137778725,
0.055315461,
0.0331479013,
-0.0146006634,
-0.0149569716,
0.0137697691,
0.0096364236,
-0.0628805459,
-0.000847091,
-0.0210883617,
-0.0500063077,
-0.0056874151,
-0.0305720903,
0.0284943338,
0.0404983871,
0.0103158411,
-0.0256885532,
-0.0280449241,
0.0205430109,
0.0452418365,
-0.0091826711,
0.0027485653,
0.0023272666,
-0.0228850376,
0.0259160679,
-0.0297314953,
-0.0553825423,
0.0614971444,
0.0219927654,
0.002448593,
-0.0668206662,
0.0179187097,
-0.0083911698,
0.0313530304,
0.035248246,
0.0422422588,
-0.0187955201,
-0.0183463655,
-0.0391681381,
0.0314753018,
0.0037073167,
-0.0455607213,
0.0558590479,
-0.0300294589,
-0.0056153205,
-0.0024661766,
-0.0346726477,
-0.0407385007,
0.0169751681,
0.0416893885,
0.0043295692,
0.0186442174,
0.0231207162,
-0.0887612179,
-0.0254635084,
0.0512261726,
0.0561506897,
0.0116448747,
-0.0114546875,
-0.0081248591,
0.0375480689,
-0.0309023857,
0.0030444355,
0.0050488496,
-0.0240864083,
0.0081823692,
0.021538673,
0.0218460672,
-0.0228693113,
0.0071741287,
0.019659346,
0.0255653113,
-0.0378007852,
-0.0271714013,
-0.0151141342,
0.0234619174,
-0.0573765412,
-0.0072450438,
0.0585197769,
0.0384414494,
-0.0495853126,
0.0015186286,
-0.0199684892,
0.0002628979,
0.0254251882,
0.0564940758,
-0.0966745466,
-0.0000784108,
-0.0095621878,
0.0143185211,
-0.0169278495,
0.0091644786,
0.0025943222,
-0.0893943086,
0.027425345,
0.0500555113,
-0.018001359,
-0.0156226102,
-0.0055061122,
-0.0215726271,
-0.0129195889,
-0.008127505,
0.0235296469,
-0.0207312312,
-0.0038372402,
-0.0005888005,
0.0177817103,
-0.072587654,
-0.0152926343,
-0.0467106327,
-0.0506367609,
0.0703285411,
0.0340622775,
0.0314957239,
-0.0241033342,
-0.0421572216,
-0.0635013282,
0.0096779121,
-0.0010314333,
0.0460710377,
0.002316762,
-0.0152235059,
0.0186402611,
-0.0386251397,
-0.0082069132,
-0.0151488427,
-0.0070366156,
0.076840654,
-0.0161335953,
0.045564346,
0.0246431269,
-0.0378932618,
-0.0088624051,
0.051123593,
0.0386846736,
-0.0091643054,
0.0132715609,
-0.0086561358,
0.0183553342,
-0.0940535367,
-0.0048611755,
-0.0042388113,
-0.05209627,
0.0382164568,
-0.0970836207,
0.0050800666,
0.0094152503,
0.0602783151,
0.0022194074,
-0.0074035171,
0.0062270979,
-0.0338422284,
-0.0563579798,
0.0138993962,
-0.0211089719,
-0.0070609846,
-0.0063535543,
-0.0029122296,
0.0128097599,
-0.0586232431,
0.0001842924,
0.0525329411,
-0.0403020792,
-0.0564042106,
0.0019611903,
0.0617277212,
0.0261254162,
0.0734897554,
0.0106531633,
-0.0049801851,
0.0117430389,
0.0036905843,
-0.0020431983,
-0.0117332637,
0.0656163692,
-0.0220083129,
-0.0368069597,
0.0483966991,
-0.0113822976,
-0.017267853,
0.0113070849,
0.0386166982,
0.0038088288,
0.0000123273,
0.0537332557,
0.0220791157,
0.0182081461,
-0.0036642328,
0.0265376456,
-0.030032536,
0.0585144497,
0.0085636675,
0.0048576104,
-0.0372659117,
0.002449827,
0.0273385718,
-0.0248737466,
-0.0175578427,
0.0036139125,
-0.0190617349,
0.0414991789,
0.0013704096,
-0.0000704679,
-0.0470355041,
0.0027803404,
-0.0101396041,
0.033185076,
-0.0011764293,
0.0342305526,
0.0138189467,
0.000308307,
-0.0544517338,
0.0139129665,
-0.0069157323,
0.0242579337,
-0.0533351116,
0.0068054334,
-0.0441175848,
0.0133080445,
-0.0198939145,
-0.0056691482,
0.0657846779,
-0.0570718721,
0.0857995674,
-0.0109745665,
-0.0352367498,
0.0110148918,
0.0182047691,
-0.0003658379,
-0.0043757646,
-0.0121856127,
0.0234061759,
0.0026304843,
0.0106335813,
0.0447979011,
-0.0590867661,
0.0660577342,
-0.0235707778,
-0.0235495418,
-0.0372493938,
-0.0175828896,
0.0334278159,
-0.0039187404,
0.0064677596,
-0.0558535084,
-0.0573968478,
-0.0362077504,
0.0315070935,
0.0163940862,
0.0104402434,
0.0005174787,
0.0356782861,
0.0611322634,
-0.006517183,
-0.0026407505,
-0.0383835025,
0.012936553,
-0.0235029869,
-0.0261811055,
-0.0080200564,
-5.906488097e-33,
-0.0487726592,
0.0108214663,
-0.0053860215,
0.001053163,
-0.049117405,
0.0164488796,
-0.0339952484,
-0.0409980305,
0.0009667343,
-0.0340429284,
-0.011552278,
-0.0168157425,
0.024255842,
-0.0082295416,
0.0091572301,
0.0517224483,
0.069533132,
0.0037633071,
-0.0158136766,
-0.0012971997,
0.0182159152,
0.022321621,
0.1010575071,
-0.0236251373,
0.0032003138,
0.008935404,
0.0124846771,
0.0382171758,
-0.0293827094,
0.0619195476,
-0.0153023545,
0.0208781324,
0.0123027144,
-0.038147904,
-0.0277561322,
-0.0599338524,
-0.0745070875,
-0.0199266728,
0.0270878114,
0.0318128988,
-0.0417020321,
-0.0880835056,
0.0550881214,
-0.0579415895,
-0.0078864805,
0.013321287,
-0.0229031518,
-0.0369221196,
-0.0599604174,
-0.0291866828,
-0.0531653725,
0.0113651743,
0.0124841118,
0.0052986816,
0.0454050563,
0.0507577807,
0.01549548,
-0.0022716576,
-0.0398308076,
0.0094086984,
-0.0139719443,
0.0657943711,
0.0212362502,
-0.0392162018,
-0.0158316493,
0.0508721247,
0.0238667894,
0.0006084523,
0.0471129157,
-0.0176675897,
0.0013283837,
0.0154320467,
0.0117127188,
0.0456563085,
0.0614654571,
-0.0043390519,
-0.0201579574,
0.0358779579,
-0.0152417496,
0.0452549309,
0.0137098245,
0.053368751,
-0.0061757714,
0.0063557941,
-0.0292194113,
-0.0295470022,
-0.0383201763,
-0.0528638065,
-0.0184518043,
0.0131104589,
-0.033895921,
0.0305605866,
0.0925660878,
0.0093511967,
0.0316210911,
-0.008094579,
-0.0206975341,
0.0059614233,
-0.0161004253,
0.0316281095,
0.0118637709,
-0.0512744524,
0.0217101369,
-0.0101078646,
0.0243257787,
0.0073485877,
0.0410419218,
0.0594591573,
-0.0321719907,
0.0005130035,
-0.0528375767,
-0.0193508454,
-0.0503508039,
0.0234018881,
0.0105464272,
0.0339813121,
0.003438693,
0.0212899055,
0.0295027252,
-0.0661986098,
-0.0048398944,
0.0153173627,
0.0438562036,
0.0556296147,
-0.0495112576,
0.0314071923,
-0.0207270049,
0.0262747239,
0.037387196,
-0.0409576856,
-0.0169139132,
0.0162083413,
2.73e-7,
0.0632746294,
0.0772850737,
0.0540274605,
0.0214156397,
0.0241527986,
-0.0267869532,
-0.0245362408,
0.0349382237,
-0.0002711066,
0.0401149131,
0.0431447513,
-0.0229332,
0.0534566566,
0.0260392111,
-0.0340845324,
0.0176368747,
0.001329327,
-0.0103626614,
-0.049530942,
-0.0181243531,
0.0602469258,
0.0876279995,
0.0125354119,
-0.0144285485,
-0.0233089384,
-0.0057222755,
0.0731020123,
0.0025617089,
-0.0610569231,
-0.0038702975,
0.0487926342,
0.0860531181,
-0.0016828545,
-0.0121317571,
-0.030844124,
0.0002536386,
0.047353927,
0.0491091236,
-0.0040329578,
-0.0164003987,
0.0290192328,
-0.0055863438,
-0.0244970322,
0.026949428,
0.0036201556,
-0.0594746768,
0.0059244609,
0.042373728,
-0.0252688825,
0.0358814783,
0.0070848763,
-0.0090974765,
-0.0337327309,
0.0555485971,
0.0073509137,
0.0450880639,
-0.0272671673,
-0.0363521539,
-0.0267106146,
0.0122450041,
-0.022407759,
-0.0131549807,
-0.0099958042,
-0.0062144194,
0.0112497862,
-0.0882183015,
-0.0539604686,
2.630505646e-34,
0.048258014,
0.0036660125,
-0.0127515476,
0.0171565339,
0.0473686419,
-0.0449439846,
-0.0409742668,
0.0292542819,
-0.0298629552,
-0.05887888,
-0.0486684442
] |
http://arxiv.org/abs/2405.15067v2
|
http://arxiv.org/pdf/2405.15067v2
|
Promoting Constructive Deliberation: Reframing for Receptiveness
| 2024-05-23T21:35:22 | 2024-06-21T21:03:27 |
Gauri Kambhatla, Matthew Lease, Ashwin Rajadesingan
|
cs.CL
|
cs.CL
|
To promote constructive discussion of controversial topics online, we propose
automatic reframing of disagreeing responses to signal receptiveness to a
preceding comment. Drawing on research from psychology, communications, and
linguistics, we identify six strategies for reframing. We automatically reframe
replies to comments according to each strategy, using a Reddit dataset. Through
human-centered experiments, we find that the replies generated with our
framework are perceived to be significantly more receptive than the original
replies and a generic receptiveness baseline. We illustrate how transforming
receptiveness, a particular social science construct, into a computational
framework, can make LLM generations more aligned with human perceptions. We
analyze and discuss the implications of our results, and highlight how a tool
based on our framework might be used for more teachable and creative content
moderation.
|
Promoting Constructive Deliberation: Reframing for Receptiveness: To promote constructive discussion of controversial topics online, we propose
automatic reframing of disagreeing responses to signal receptiveness to a
preceding comment. Drawing on research from psychology, communications, and
linguistics, we identify six strategies for reframing. We automatically reframe
replies to comments according to each strategy, using a Reddit dataset. Through
human-centered experiments, we find that the replies generated with our
framework are perceived to be significantly more receptive than the original
replies and a generic receptiveness baseline. We illustrate how transforming
receptiveness, a particular social science construct, into a computational
framework, can make LLM generations more aligned with human perceptions. We
analyze and discuss the implications of our results, and highlight how a tool
based on our framework might be used for more teachable and creative content
moderation.
|
[
0.0838829353,
0.0636355653,
-0.0270991884,
0.0044738217,
-0.0209475402,
0.0025897759,
-0.0104388706,
0.0222553015,
-0.0185092576,
-0.0313769467,
-0.0105527341,
0.006176726,
-0.0418998376,
-0.0086915083,
-0.0364353023,
-0.0559885241,
0.0045994371,
0.0097987102,
0.0594282337,
-0.0018695175,
0.0428183377,
-0.0112691307,
-0.0102983033,
0.031788446,
-0.0100184083,
0.0107521107,
0.0251830369,
0.0240998957,
0.0141126774,
-0.0345506966,
0.0381939337,
0.0832803845,
-0.0023968876,
0.0372338332,
0.0000017682,
-0.0247584302,
-0.0073011983,
-0.0386801176,
0.0124172857,
-0.0133406064,
0.1219657883,
0.0109641701,
-0.0029081022,
0.0413919054,
-0.0697704405,
0.0282663628,
0.0309649352,
0.033313863,
0.0630498827,
0.051835373,
0.0022419179,
-0.043413274,
-0.0366935544,
-0.0020687229,
0.0010011268,
0.0005657913,
0.0189708844,
-0.0054180659,
-0.0451137796,
0.0038955845,
-0.012146621,
-0.0102103641,
-0.0143857328,
-0.0645954832,
0.044088915,
-0.0333486162,
0.0075749112,
-0.0270184819,
0.0020562608,
0.0512614883,
0.0866497606,
-0.0272289161,
0.0198103357,
-0.0124124121,
0.0275164675,
0.0320755541,
-0.069932811,
0.0004738593,
-0.0280464645,
0.0132070584,
-0.0053533157,
-0.01120965,
0.0189977791,
0.0357432067,
0.0685949922,
0.0174943544,
-0.0326356404,
-0.0073970738,
-0.0357842445,
0.0436438061,
0.0524273813,
-0.0351244397,
0.0195610393,
0.0301960148,
0.0015598466,
0.0028740107,
-0.0014997973,
-0.0325907543,
0.0598723516,
-0.0524931923,
0.0291657969,
-0.000652826,
0.020135425,
0.0679755062,
0.0231895205,
-0.0411230326,
-0.0317636468,
0.0144916698,
0.0288643334,
-0.007133788,
-0.0704830587,
0.0037550635,
0.0101733431,
0.0021138326,
0.0158239175,
0.0480979159,
0.0031722297,
0.0006775309,
-0.0771114901,
-0.0554483235,
-0.0292930976,
-0.0022016696,
-0.0069148354,
-0.0009300055,
-0.093961753,
0.0644059852,
-0.0435140915,
0.0286036506,
-0.0146380421,
-0.0343861245,
0.0021441879,
0.0130635137,
0.0244465899,
0.0139444573,
0.0339037441,
-0.005072182,
0.0056933495,
0.0236646198,
-0.0443787612,
-0.0367765054,
-0.0171149187,
0.0008140588,
0.0444543809,
-0.0749254003,
0.0278043095,
-0.01456689,
-0.0339340642,
-0.0300008357,
0.0019140696,
0.0132427085,
-0.0186502691,
0.0234755073,
-0.0398325659,
0.0038344334,
0.0773370788,
0.0495945141,
-0.0344811678,
0.0394184478,
-0.0231632516,
-0.0153405983,
0.0499945581,
0.001074857,
0.0210951436,
-0.0182814188,
0.0028928372,
-0.0093349656,
-0.0073754219,
0.0132932486,
-0.0761187598,
-0.0156053388,
0.0169502813,
0.0203013029,
-0.0496361405,
-0.0299085099,
0.0451943949,
0.0617442653,
-0.0663391724,
0.064365983,
-0.0385680646,
0.0150878765,
-0.0033679269,
0.0016242127,
0.0063234782,
0.0711274371,
-0.0067406995,
-0.0350053981,
0.0142469984,
0.021099139,
-0.0120504592,
-0.0688301548,
0.0089080576,
0.0116813704,
0.0735953823,
0.009427405,
-0.0068677268,
0.0328482874,
0.0093268715,
-0.0044708448,
0.0497316606,
-0.0486151464,
0.0025218255,
-0.0596954077,
0.0574839041,
0.0125421062,
0.0157636218,
-0.0340619832,
-0.0219775904,
-0.0487710014,
0.0731234252,
0.0739750341,
0.0870121866,
-0.0049012443,
-0.0594274066,
-0.0328617953,
-0.0325495303,
-0.003009646,
0.0124977184,
0.0344646648,
0.0111818518,
-0.0067564379,
-0.0222685579,
0.0529973134,
0.0032909876,
0.0286584739,
0.0704357326,
0.0167998038,
-0.0118122594,
0.0056951391,
0.0685019791,
0.0380599946,
-0.0066488204,
0.0007538812,
0.0180489495,
0.0098849172,
-0.0323651098,
0.0097073792,
-0.0016742342,
0.0280814115,
0.0282128286,
-0.0769824237,
-0.043779552,
-0.0053478875,
-0.0079163136,
0.02794246,
0.0816255361,
-0.0002107617,
0.003522692,
-0.008946226,
-0.0415216126,
-0.0671935976,
-0.0213033706,
-0.0463013463,
0.0202502999,
0.0160286669,
0.0076255081,
0.041176416,
-0.0905826911,
0.016920507,
0.0055997171,
-0.0286817681,
-0.0317261741,
0.0109344628,
-0.0056410306,
0.0507080555,
-0.0171109717,
-0.0533119477,
-0.0517642684,
-0.0760217458,
-0.0510808937,
-0.0215301383,
0.0359928347,
0.0469983928,
-0.0630716309,
0.0110838348,
-0.0174413025,
-0.0299717896,
-0.0155185023,
0.0113534955,
0.0308547374,
0.0581679158,
-0.0370106511,
-0.0077545587,
-0.0259544384,
-0.0009214587,
-0.0108593414,
0.0093185445,
0.0501412004,
-0.0056810994,
-0.0176771842,
-0.0433941744,
-0.0788857117,
-0.006570837,
-0.0177095216,
-0.0452908985,
0.0575040057,
0.0345680453,
0.0349756591,
0.0586980917,
0.0734181255,
0.0441935882,
0.0281989705,
0.0332622789,
-0.0287804976,
-0.0209491961,
-0.0278928466,
-0.007164205,
-0.0156022022,
-0.0558559299,
-0.0008115434,
0.0487840362,
-0.0074518253,
0.0070862733,
-0.0574953929,
0.095015794,
-0.0207133498,
-0.0188602693,
-0.0267535765,
-0.0520373248,
-0.0039916812,
0.0408809669,
-0.0437150262,
-0.0303486176,
0.0597826838,
0.03933331,
0.0251859035,
-0.0255733281,
-0.0114287343,
-0.044982031,
0.022271147,
-0.0513938963,
0.0024207293,
0.0105549572,
0.0046869232,
-0.0213797335,
-0.0469946153,
0.1053506956,
-0.0602415726,
0.0307895876,
-0.0623819344,
-0.0158694331,
0.0048157843,
-0.0199347958,
-0.0149375964,
-0.0258256681,
-0.0328944996,
-0.0381006449,
0.0259718187,
0.0170866754,
-0.0182786472,
-0.0162633974,
0.0384902097,
-0.0381905437,
0.0169569179,
-0.0195983686,
0.0478641838,
-0.0213305131,
0.0089636417,
0.0658925474,
0.0439798981,
-0.0247033518,
-0.0051813228,
0.0962422565,
-0.0341670699,
-0.0483315401,
-0.0053707166,
-0.0231816806,
0.009240225,
0.0099814916,
0.0208654255,
-0.0195319857,
0.0529685095,
-0.038871251,
0.0228459928,
0.0344645977,
0.0274699926,
-0.0237976536,
-0.0437284298,
-0.0587406717,
0.0164126214,
0.0193269961,
0.001807704,
-0.0596444868,
-0.0087439539,
-0.0361108743,
0.0131820133,
-0.0392367952,
-0.056609109,
-0.0006322329,
-0.0625099093,
0.0542559922,
0.008109252,
0.0137221366,
-0.0084311254,
-0.0419994146,
-0.0002385051,
0.0264113359,
-0.0338620469,
0.0098302346,
0.0292197019,
-0.0112073096,
0.0222604126,
-0.031807743,
0.0326768868,
-0.0165047441,
-0.0193823799,
0.0592139885,
0.0016706995,
0.061058633,
0.0229565166,
-0.0436281413,
0.0123233655,
0.0125950649,
-0.0295055602,
-0.0007771538,
0.0183816142,
0.0059731929,
0.0146806603,
-0.0597131513,
0.0153028285,
-0.0056504905,
-0.0609051138,
0.0091732722,
0.0528985932,
0.0140349688,
-0.022371823,
-0.0582963265,
0.0433797389,
0.020978624,
0.0384481288,
0.0419234671,
-0.0093872827,
-0.0089525888,
-0.0258769318,
0.0090530533,
-0.0038705738,
-0.0535512827,
-0.0309378095,
0.0135374172,
0.0221212525,
-0.0655831471,
0.0070048976,
-0.0349686071,
-0.0200444236,
0.0054228357,
-0.0074465433,
0.0020207921,
-0.0083841709,
-0.0245345775,
0.0230092984,
-0.0079863416,
-0.029021401,
-0.0426533073,
-0.0459151454,
-0.021794375,
-0.0113102412,
-0.0298171062,
0.0083005736,
-0.0481066294,
-0.0409713164,
0.0007730668,
0.0330163427,
0.0317754671,
-0.0038349431,
0.0310166385,
-0.0496425256,
0.0169241857,
0.0509676374,
0.0211467538,
0.0134945344,
0.0233351924,
0.0659544542,
0.0086621689,
0.0001371568,
0.0112783527,
0.0410857424,
0.0322238132,
-0.0138650639,
-0.0308404099,
-0.0467328131,
-0.0155502055,
-0.003298582,
0.0121794622,
0.0279206671,
-0.046929311,
-0.0155431014,
-0.028034687,
0.0088405814,
-0.0101116979,
0.0097192153,
-0.0204549562,
-0.0101133743,
-0.0535433665,
-0.0329304375,
0.0363035649,
-0.0573311932,
0.06120703,
-0.0050432361,
0.067186065,
0.0252258759,
0.0081692962,
-0.0229916554,
0.0049637146,
-0.0186528303,
0.0053559039,
0.0265408363,
0.0196536146,
0.0264927018,
0.0666790977,
-0.0011391038,
-0.0135843055,
0.0064443788,
-0.0044430178,
-0.068794772,
-0.0559808351,
0.0118317734,
0.0322618522,
0.0090267016,
0.0269173943,
0.0277956482,
-0.0042345761,
-0.032927379,
0.0127947908,
0.0297708008,
-0.0057707438,
-0.0310240667,
0.0604065582,
0.0132297557,
-0.0254890528,
0.0063631036,
0.0190205052,
-0.0800490081,
-0.0124416975,
0.008659265,
-0.0219530333,
-0.0292203259,
-0.0407150686,
0.0170752238,
0.0202249996,
0.0117116095,
0.0190319959,
-0.0634530932,
0.0097856652,
0.0436851904,
-0.0066463337,
0.0174351297,
-0.0325943306,
-0.0085183596,
0.0304497145,
0.0168985836,
-0.016801687,
-0.0274060648,
0.0367787853,
-0.0517284833,
-0.0177785363,
-0.0458341204,
-6.227908661e-33,
0.0248108041,
-0.0588872768,
-0.0141491387,
0.0542146564,
-0.0139496727,
-0.0272833519,
-0.0241110343,
0.0555083007,
-0.0193724185,
0.0439400487,
-0.010505368,
0.0268640518,
-0.0071907379,
-0.0147561012,
-0.0034328452,
-0.0505895354,
0.0320024453,
-0.0327476487,
-0.0181765761,
0.0033608645,
0.0116903866,
-0.0171911977,
0.0231286995,
-0.0866319612,
0.0404550023,
0.0056724865,
0.015488713,
0.0425639935,
0.0271862857,
0.0478662103,
0.0034278841,
0.0427992046,
0.0415096171,
-0.0215432309,
0.0076739695,
0.0553022288,
-0.029082315,
-0.0286785234,
0.0150817232,
-0.0509413555,
-0.0249654744,
-0.0168059804,
0.0614343993,
0.0268783495,
-0.0190972686,
-0.0117905019,
0.0272347331,
-0.0184152387,
0.0308655854,
-0.0212602485,
-0.0479577221,
0.0044010831,
-0.0276014637,
-0.0089980494,
-0.0097046951,
0.0061794464,
0.0133739207,
0.0473691635,
-0.0892333388,
0.0515423194,
-0.0251182858,
0.0053309058,
0.0282098725,
-0.0211441852,
0.018982891,
0.0626540408,
0.0128547987,
0.0985665768,
-0.0527799688,
-0.040089421,
-0.0212138444,
0.014039441,
0.0692670643,
0.116877988,
0.0708614886,
-0.0692568347,
-0.0632097274,
-0.0145433974,
0.0513947308,
0.0341150798,
0.0376691297,
0.0061275242,
-0.0050191381,
0.0148136811,
-0.028126549,
-0.0701557249,
-0.02843724,
-0.0133648673,
-0.0206699353,
0.0139115117,
-0.048363544,
-0.0534524508,
-0.003487075,
-0.0288895965,
-0.0104960352,
0.0023182069,
0.0125595629,
0.0090570664,
0.0532589816,
-0.0432336703,
-0.0266967025,
0.0226822309,
-0.006411647,
-0.0023387598,
0.0451081544,
-0.0302829444,
0.0075944345,
-0.0022329157,
-0.0644712672,
-0.0083750859,
-0.0260126553,
0.0710891336,
-0.0060916701,
0.0101658991,
-0.0074474574,
0.0365486629,
0.0203943513,
0.001671097,
0.0357176103,
0.0418534726,
-0.0285278577,
-0.0052444609,
-0.0202059224,
0.0060353745,
-0.0131185064,
-0.0095520318,
0.0157047566,
0.0150934691,
0.0061933897,
-0.0599770471,
-0.018632872,
-0.0294686724,
2.529e-7,
0.0406407118,
0.0447375961,
0.0111909434,
-0.0246758312,
0.0662893355,
0.0338916481,
-0.0036694179,
0.0201818682,
0.0439298823,
-0.0432110094,
0.0080264704,
0.0127214408,
0.0440588109,
0.0421157144,
-0.088859506,
-0.0715522096,
-0.0697073862,
0.0064810822,
0.0104092928,
0.0078998366,
0.0553665347,
0.0029291038,
-0.0085044187,
-0.001885278,
-0.0499217436,
0.0314536914,
0.0699184462,
-0.0601592362,
0.0279008523,
0.0037587741,
0.049724184,
-0.0764191374,
0.00996105,
0.0345836766,
-0.0153047349,
-0.0221142117,
-0.0293364096,
0.0475991257,
0.0403924175,
-0.0570247732,
-0.0009908251,
0.0327540934,
0.0476407483,
-0.0163769759,
0.0001082854,
-0.0692165643,
-0.0249224864,
-0.0598927327,
-0.0480096862,
-0.0306820422,
0.0629969463,
-0.0114427097,
-0.0483735166,
0.0004145673,
0.0091463951,
0.0288334619,
0.0161528271,
-0.0663359985,
-0.0401459709,
-0.0391854495,
-0.0535057932,
-0.0577172674,
-0.012720095,
0.0216505751,
0.0151428171,
0.0281139389,
0.056296289,
1.796018189e-34,
0.029333381,
-0.03297792,
0.013542844,
-0.001955549,
0.0514831543,
0.0276212841,
0.0049672853,
-0.0123833511,
0.0473655388,
-0.0126502551,
0.0038430756
] |
http://arxiv.org/abs/2405.15064v1
|
http://arxiv.org/pdf/2405.15064v1
|
Reframing Spatial Reasoning Evaluation in Language Models: A Real-World Simulation Benchmark for Qualitative Reasoning
| 2024-05-23T21:22:00 | 2024-05-23T21:22:00 |
Fangjun Li, David C. Hogg, Anthony G. Cohn
|
cs.CL
|
cs.CL, cs.AI, cs.DB
|
Spatial reasoning plays a vital role in both human cognition and machine
intelligence, prompting new research into language models' (LMs) capabilities
in this regard. However, existing benchmarks reveal shortcomings in evaluating
qualitative spatial reasoning (QSR). These benchmarks typically present
oversimplified scenarios or unclear natural language descriptions, hindering
effective evaluation. We present a novel benchmark for assessing QSR in LMs,
which is grounded in realistic 3D simulation data, offering a series of diverse
room layouts with various objects and their spatial relationships. This
approach provides a more detailed and context-rich narrative for spatial
reasoning evaluation, diverging from traditional, toy-task-oriented scenarios.
Our benchmark encompasses a broad spectrum of qualitative spatial
relationships, including topological, directional, and distance relations.
These are presented with different viewing points, varied granularities, and
density of relation constraints to mimic real-world complexities. A key
contribution is our logic-based consistency-checking tool, which enables the
assessment of multiple plausible solutions, aligning with real-world scenarios
where spatial relationships are often open to interpretation. Our benchmark
evaluation of advanced LMs reveals their strengths and limitations in spatial
reasoning. They face difficulties with multi-hop spatial reasoning and
interpreting a mix of different view descriptions, pointing to areas for future
improvement.
|
Reframing Spatial Reasoning Evaluation in Language Models: A Real-World Simulation Benchmark for Qualitative Reasoning: Spatial reasoning plays a vital role in both human cognition and machine
intelligence, prompting new research into language models' (LMs) capabilities
in this regard. However, existing benchmarks reveal shortcomings in evaluating
qualitative spatial reasoning (QSR). These benchmarks typically present
oversimplified scenarios or unclear natural language descriptions, hindering
effective evaluation. We present a novel benchmark for assessing QSR in LMs,
which is grounded in realistic 3D simulation data, offering a series of diverse
room layouts with various objects and their spatial relationships. This
approach provides a more detailed and context-rich narrative for spatial
reasoning evaluation, diverging from traditional, toy-task-oriented scenarios.
Our benchmark encompasses a broad spectrum of qualitative spatial
relationships, including topological, directional, and distance relations.
These are presented with different viewing points, varied granularities, and
density of relation constraints to mimic real-world complexities. A key
contribution is our logic-based consistency-checking tool, which enables the
assessment of multiple plausible solutions, aligning with real-world scenarios
where spatial relationships are often open to interpretation. Our benchmark
evaluation of advanced LMs reveals their strengths and limitations in spatial
reasoning. They face difficulties with multi-hop spatial reasoning and
interpreting a mix of different view descriptions, pointing to areas for future
improvement.
|
[
0.0364216156,
0.0283665992,
-0.010062282,
0.0320922807,
-0.0759450048,
0.0161398966,
0.024040306,
-0.0099209286,
0.0052707777,
-0.0180164687,
-0.016241435,
-0.0193485674,
0.0489258468,
0.0192024726,
0.0039434168,
-0.023977343,
0.0262396969,
-0.0405572504,
-0.0048192968,
-0.0030214703,
-0.0547426119,
0.0229954422,
-0.0574928634,
0.020433465,
-0.0315358564,
-0.0112435566,
-0.0231995657,
-0.0117821228,
0.0104042636,
-0.0629295558,
-0.0187078062,
0.0215247907,
0.0507363193,
0.0129921241,
0.0000021728,
-0.0142075252,
-0.0243022423,
-0.0038054385,
0.0143308435,
-0.03323102,
0.0044223079,
-0.045193281,
0.010736011,
-0.011436169,
-0.0267869141,
-0.0073564649,
0.0288766176,
0.029466683,
0.0053593642,
0.0104164267,
-0.026228793,
-0.0363069437,
-0.0197045617,
0.0188303608,
0.0312199239,
-0.0112594813,
-0.01741709,
0.0566362627,
0.0468363352,
0.0250034574,
-0.0130964471,
0.0106742084,
0.0522078201,
0.026321413,
-0.0163992867,
0.0292704534,
-0.0249487832,
0.0395304821,
-0.0042728418,
0.0440672934,
0.0072887498,
-0.0195567701,
0.0283867996,
0.0009700141,
-0.0451590791,
0.0092195235,
-0.0595434159,
-0.0027316655,
-0.0132785076,
0.0144783463,
-0.0489357598,
0.0448269881,
0.0120404055,
0.0020380376,
-0.0223427527,
0.1070296839,
0.0075697987,
-0.0181303732,
-0.0462751687,
0.001018569,
-0.0045006755,
-0.0660269484,
-0.0031835069,
0.0323631577,
-0.0535405613,
0.0382316597,
0.0283955149,
0.0957258493,
0.0181084257,
-0.0625900179,
0.0257478673,
0.0153336683,
0.0112308227,
0.0572023913,
-0.0298533291,
0.1026116163,
-0.0113233244,
0.058148928,
-0.0367301181,
0.0159175359,
-0.0298827291,
0.0055484255,
-0.0137052918,
-0.0000667396,
-0.0534291118,
-0.0458468013,
0.0098735029,
0.0143341981,
-0.0416460894,
0.0507711209,
-0.0080856979,
0.0435554832,
0.0201199204,
-0.0350847915,
-0.0272690859,
-0.0277315304,
-0.0246080644,
0.02877222,
-0.0671057701,
-0.0369174108,
0.00090102,
0.0110216308,
-0.0181260295,
-0.0318893604,
-0.0357822776,
0.0541397296,
0.026650399,
-0.0060997275,
0.0052443468,
-0.0514386408,
-0.0365785211,
-0.0590241514,
0.0370581299,
-0.0826634467,
-0.0063042641,
0.000138224,
-0.0379434265,
-0.0316895582,
-0.0036263089,
-0.0119130882,
-0.0094799409,
0.0603796877,
0.0121107344,
0.0400488488,
0.038464427,
0.0293353777,
0.0282145515,
0.055140987,
0.0573333502,
0.0471607447,
0.0134612899,
0.0472779833,
0.0191694535,
0.0224209689,
-0.00822437,
0.0095341383,
-0.0668952018,
0.0298503321,
-0.0269499309,
0.0515191704,
-0.0511176027,
0.0409617722,
-0.0036164292,
-0.0047978978,
0.0388094783,
0.080450505,
0.0038111564,
-0.0450214222,
0.0287495349,
0.0109055769,
-0.0375793204,
-0.0215822626,
-0.0472298265,
0.0236815121,
0.0117772128,
0.0029415942,
-0.0000945911,
0.0133884717,
0.0222093388,
-0.0450068489,
-0.0549014062,
0.0255409796,
0.0292429067,
-0.0354359895,
0.0291088764,
-0.0156940464,
-0.0343007222,
0.0653269514,
0.0160117727,
-0.0289970022,
-0.0230462383,
-0.0070232167,
-0.0037482425,
0.0614835508,
0.0360920168,
-0.0134229297,
-0.0566782057,
0.0164022092,
0.0137499636,
0.0063626962,
0.0646118224,
0.0095640831,
0.0090933312,
-0.009397557,
-0.0500204228,
-0.0027618583,
0.0453744903,
0.0328373984,
0.0698616579,
0.0070567187,
0.011807234,
-0.0854694471,
0.0278407615,
0.0162637495,
0.0013866032,
0.0249066483,
-0.0937608778,
0.0218008552,
-0.0112338457,
-0.0194782633,
0.0146041419,
0.0112011526,
0.0163398311,
-0.004943883,
0.002466725,
0.0489615984,
-0.0722195432,
-0.0110765304,
0.0143191675,
-0.0017722675,
-0.0219250973,
0.0085600344,
-0.0401529446,
0.0044840733,
0.0261591114,
-0.0236005969,
0.0033656445,
-0.0056797769,
0.0074243513,
-0.0173588078,
-0.0346733965,
-0.0382762216,
0.074584052,
-0.0261875167,
0.0384220853,
0.0012887662,
0.0262605548,
0.0552755967,
0.0032612744,
0.0473472439,
-0.0344064645,
0.0078415442,
0.0126746604,
-0.0358983278,
0.0114533352,
-0.0485357419,
-0.0079799704,
-0.0533902571,
-0.0075462339,
-0.0045393715,
0.0079882126,
0.0108474307,
-0.0673112795,
-0.0260244105,
-0.0072139092,
0.0178486928,
-0.0299677011,
-0.0094516687,
0.0230099093,
0.0332409851,
-0.0376943313,
-0.0162708852,
-0.0024425318,
0.0043529863,
-0.045511663,
0.0450824127,
0.0531036183,
0.0152542386,
0.0166955963,
-0.0277954172,
-0.0422995351,
0.0197158437,
0.0043096095,
-0.0059187757,
0.0561224371,
-0.0310314652,
0.0118898051,
0.0147617524,
-0.0455268845,
0.0104749426,
0.0518343933,
-0.0731616616,
-0.0626201183,
0.0130832735,
-0.0167591721,
0.0191169865,
-0.0068797651,
-0.066552937,
-0.0223677289,
0.0315967277,
-0.0057095247,
-0.0181364082,
-0.0263942908,
0.0603761636,
-0.0303245913,
-0.0079840068,
-0.0415850207,
0.0025286686,
0.009676951,
0.1059768274,
-0.0440360159,
-0.0074640159,
0.0121845044,
-0.0861188173,
0.0444657095,
-0.0328787304,
0.0476295836,
-0.0306855123,
0.1166579723,
0.0071373093,
0.0041143796,
-0.0307548754,
0.0025751947,
-0.009230393,
-0.0408746041,
-0.003189926,
0.0046981536,
0.0124271363,
-0.1232520118,
0.018488232,
0.0275910608,
-0.0311790686,
-0.026086323,
0.027758304,
-0.0494812019,
0.0031872378,
-0.0466011651,
-0.0234659445,
-0.0046406798,
0.0782112852,
0.0081146983,
0.0415263101,
-0.026072558,
-0.0240149871,
-0.0466159768,
0.0120770792,
0.0461567752,
0.0819747001,
-0.0202500932,
0.0106894672,
-0.0387792699,
0.036175929,
-0.0526582599,
0.0247360766,
-0.0354269221,
0.0638853684,
0.0265753344,
0.0297216624,
-0.014731816,
0.014911118,
0.0651833341,
-0.0612988658,
-0.0854671821,
0.0751762688,
0.0737932473,
0.0317997597,
0.0061441134,
-0.0416743644,
-0.007065862,
0.0094065694,
-0.0066599743,
-0.0177891795,
-0.0047108405,
0.0060456702,
-0.051387962,
0.0243457854,
-0.048491165,
0.0049970704,
-0.0794982836,
0.081921339,
-0.0094689261,
-0.1007530391,
-0.0679194853,
-0.0287895761,
0.0083156396,
0.0166290328,
-0.0179450363,
-0.0115966015,
0.0263356697,
0.0011671804,
-0.016811097,
-0.0386296734,
0.0112286797,
0.0272367988,
-0.0502649695,
0.0347134881,
0.00232382,
0.0553276911,
0.0399929583,
-0.0940801576,
-0.0167297833,
0.0698065907,
0.0452602617,
-0.0137064662,
-0.0119836917,
0.0416797735,
0.0157150105,
-0.0013809598,
-0.0348989964,
-0.0279058591,
-0.0497958697,
-0.0101495301,
0.0689950064,
-0.0272796005,
0.0370021611,
-0.0453254282,
0.0775773376,
-0.0041270419,
0.046435263,
0.0184149928,
0.0505086072,
-0.091127634,
-0.0374966674,
-0.0245815404,
-0.0336927995,
-0.0190314483,
0.0147937378,
0.0140693588,
-0.0038905896,
-0.0440056995,
-0.0133225555,
0.011908236,
0.0314942002,
-0.0212076567,
-0.0387516171,
-0.0220183898,
-0.0221563876,
-0.0545697622,
0.0178658459,
-0.0247846022,
0.0415006503,
0.0624432378,
0.0792262182,
0.0118368464,
-0.0540952124,
-0.0098046185,
0.0015161063,
-0.0384653732,
-0.0134596517,
0.0026877194,
0.0449652299,
0.0195968524,
0.0318243504,
-0.0177567322,
0.031052012,
0.0032681287,
-0.0230302606,
-0.0212816168,
0.0271281246,
0.0306528769,
0.0332769044,
0.0211633593,
-0.0045041349,
-0.0147832427,
0.0220529176,
0.0008163766,
0.0213763397,
-0.0565915182,
-0.0150433723,
-0.0125190346,
-0.0298137926,
0.049536027,
0.0383181386,
-0.0398081839,
-0.0060491082,
0.0080251098,
-0.0750719532,
-0.0050811945,
-0.0041255713,
0.0214806665,
0.0256876536,
0.0228882395,
-0.0093147783,
0.0041041705,
-0.009847871,
0.0230110064,
0.0013757213,
-0.0439122505,
0.026518818,
-0.0288197324,
0.0289254729,
-0.0016684749,
-0.0099522779,
0.0147726061,
-0.0517988428,
0.0580925308,
-0.0151793193,
0.0178589001,
0.0344553366,
-0.0350717679,
-0.0655309036,
0.0246545132,
-0.0327288918,
-0.0597357824,
0.0146552371,
0.0497657768,
0.0678678825,
-0.0047654221,
0.0069989557,
-0.0288652964,
0.0179357845,
0.0447462462,
-0.0077185798,
0.0540564507,
-0.0468877256,
0.015442254,
0.0304863006,
-0.0075040869,
0.003169392,
0.0088846749,
-0.0070855571,
0.0288020056,
0.0053496026,
-0.0418958217,
-0.0100345267,
-0.029801961,
-0.0042324569,
0.0412166193,
-0.0084965592,
-0.012469477,
-0.0588190109,
-0.0515737385,
0.0263481978,
0.0869124681,
-0.0113902939,
0.0147154536,
0.0328658335,
0.0567394942,
0.0125022298,
-0.0463538431,
-0.0121626602,
-0.0272498559,
-0.0029807228,
0.0043677134,
-0.0055539338,
-6.22489866e-33,
-0.0026998564,
-0.0814786404,
0.0736203045,
-0.0740954801,
-0.0070190397,
-0.0223126318,
-0.0515722372,
-0.0310677178,
-0.0212124065,
-0.0016747761,
0.0056761075,
0.0135023966,
0.015639279,
0.0127568571,
0.0150452014,
0.0078881076,
0.0042934311,
0.046117425,
-0.0373662412,
-0.0419744477,
0.017266361,
-0.0416183062,
-0.003917207,
-0.0657076985,
-0.0015861824,
0.0086413519,
-0.0031031943,
0.0375796966,
-0.0631393343,
0.0256848503,
-0.0216968898,
-0.0299828034,
0.028574042,
-0.0380769037,
-0.0136406608,
0.0010495782,
-0.0589634664,
-0.0413070694,
0.0058437209,
0.0111664245,
-0.0508998521,
-0.0112174097,
0.0138483103,
0.0380468369,
-0.0191050135,
-0.0332733355,
0.0161351338,
-0.0176098421,
-0.0317971148,
0.0560141653,
-0.044841364,
0.0132550402,
-0.0025547445,
-0.0222494099,
0.0221208893,
0.0033556737,
0.0265176613,
0.0203312263,
-0.0557545274,
0.0137879578,
0.0146392118,
-0.0218830649,
0.0710699484,
-0.0128146783,
-0.0061181663,
0.0781408697,
0.0314898603,
-0.0022214891,
-0.0451450422,
-0.0092266332,
0.0479948744,
0.0795695186,
0.0373824276,
0.0352352224,
0.009750451,
0.0569952093,
-0.0417735092,
0.0195527729,
-0.0945016146,
0.0504762568,
0.036626637,
0.0157184694,
-0.0551934764,
-0.018022934,
-0.0175929982,
0.0325998105,
-0.016100511,
0.0043382933,
0.0010857469,
-0.012525619,
0.0232158471,
0.0355550498,
-0.0004194499,
-0.0215578638,
-0.017670868,
0.014085765,
0.0082718944,
-0.0012939903,
-0.0281801764,
0.0767727792,
-0.0086936913,
-0.0007716358,
0.0056649768,
0.0453880057,
0.0120660765,
0.0195369497,
0.0112531064,
0.013985021,
0.0011097111,
-0.0126982201,
-0.0436242372,
-0.0449356847,
0.0110830599,
-0.0196234509,
0.0428489111,
-0.0052085537,
-0.0056089303,
0.0416837037,
-0.0020495553,
-0.0256602187,
0.0096402913,
0.0273440219,
-0.0617688484,
0.008945954,
-0.0004685003,
-0.0071369512,
-0.0661756173,
-0.0263205413,
0.0389843509,
-0.0486623459,
-0.0103472816,
0.0284063127,
2.872e-7,
0.0211872142,
0.0685893372,
0.040866226,
0.0110799549,
0.0060184542,
0.0127878701,
0.0008576696,
-0.0280058645,
-0.0241601001,
0.0383790955,
0.0427891463,
0.0723957568,
0.0324999578,
0.0400795303,
-0.0584638715,
0.0741350353,
-0.0589122921,
-0.0386769772,
-0.0684662908,
0.0308938939,
0.0878148824,
-0.0155713297,
0.0036075958,
-0.0029933394,
-0.0148218758,
-0.0381088182,
-0.0378446504,
-0.0459636226,
0.0413088091,
0.0396395028,
0.0491345786,
-0.0282073785,
0.0046515702,
-0.0087823663,
0.001239448,
-0.0419950932,
0.0024707033,
0.0502104685,
-0.0137451403,
-0.043460384,
-0.0290913321,
-0.0793722868,
-0.0187252015,
-0.0855381787,
0.0708375499,
-0.0686441436,
0.0119760418,
0.0211228058,
0.0244312622,
-0.0047677825,
-0.0060614329,
0.043311663,
0.0051485384,
-0.0148115382,
0.0092474734,
0.0070230691,
-0.0156777706,
-0.0686730593,
-0.0336936899,
0.0297096353,
-0.0199651662,
0.001109281,
0.0005719797,
-0.0195984393,
0.0032212643,
-0.0190868545,
-0.0563337319,
2.275783236e-34,
0.0120491283,
0.0311896391,
-0.000618033,
-0.0228183512,
0.0508129597,
0.0045057512,
-0.0020138426,
0.0273534674,
-0.0079398425,
-0.0242645461,
0.0132389432
] |
http://arxiv.org/abs/2406.00022v2
|
http://arxiv.org/pdf/2406.00022v2
|
Multilingual Prosody Transfer: Comparing Supervised & Transfer Learning
| 2024-05-23T20:43:24 | 2024-06-18T05:14:21 |
Arnav Goel, Medha Hira, Anubha Gupta
|
cs.CL
|
cs.CL, cs.SD, eess.AS
|
The field of prosody transfer in speech synthesis systems is rapidly
advancing. This research is focused on evaluating learning methods for adapting
pre-trained monolingual text-to-speech (TTS) models to multilingual conditions,
i.e., Supervised Fine-Tuning (SFT) and Transfer Learning (TL). This comparison
utilizes three distinct metrics: Mean Opinion Score (MOS), Recognition Accuracy
(RA), and Mel Cepstral Distortion (MCD). Results demonstrate that, in
comparison to SFT, TL leads to significantly enhanced performance, with an
average MOS higher by 1.53 points, a 37.5% increase in RA, and approximately a
7.8-point improvement in MCD. These findings are instrumental in helping build
TTS models for low-resource languages.
|
Multilingual Prosody Transfer: Comparing Supervised & Transfer Learning: The field of prosody transfer in speech synthesis systems is rapidly
advancing. This research is focused on evaluating learning methods for adapting
pre-trained monolingual text-to-speech (TTS) models to multilingual conditions,
i.e., Supervised Fine-Tuning (SFT) and Transfer Learning (TL). This comparison
utilizes three distinct metrics: Mean Opinion Score (MOS), Recognition Accuracy
(RA), and Mel Cepstral Distortion (MCD). Results demonstrate that, in
comparison to SFT, TL leads to significantly enhanced performance, with an
average MOS higher by 1.53 points, a 37.5% increase in RA, and approximately a
7.8-point improvement in MCD. These findings are instrumental in helping build
TTS models for low-resource languages.
|
[
-0.0201729704,
0.0084731849,
0.0182249229,
-0.0092530744,
-0.0159086604,
-0.018961221,
0.0510660969,
-0.0469362177,
-0.0597804002,
0.0040276647,
-0.0993043333,
-0.0565170906,
0.0413695127,
0.0419943854,
0.052532874,
-0.0695841685,
0.031331867,
-0.0515717752,
0.0038941563,
-0.0153493248,
0.0185185671,
0.0141275534,
-0.0445917659,
0.0234409478,
-0.0031574618,
0.0265462026,
0.0109618604,
0.0036561845,
0.0376025066,
0.0019602284,
-0.0569394417,
0.0147029562,
-0.0163721368,
0.0137744071,
0.0000021522,
-0.074203521,
-0.0405093916,
0.0052573378,
0.0164388828,
0.0089011341,
0.0693504065,
0.0160440616,
-0.0286164992,
0.0332143605,
-0.0025755763,
0.0013342875,
0.0698930547,
0.004106272,
0.0170509443,
0.0799649283,
-0.0197010767,
-0.0118618198,
-0.0126256458,
0.00103835,
-0.0349986777,
0.0004393402,
-0.0065067434,
-0.0252310224,
-0.0218007695,
-0.0949748233,
-0.0410326533,
0.0465848818,
-0.0413697585,
-0.0056891888,
0.0032523216,
-0.047838971,
-0.0239498243,
-0.0362684429,
-0.0116130495,
0.0632986203,
-0.0154027343,
-0.0050892248,
-0.00543524,
-0.0026066327,
-0.0183180273,
-0.0166079476,
-0.0086750491,
-0.0222371165,
0.0274670795,
0.0063119945,
-0.0145637337,
-0.0059446841,
0.0039339256,
-0.057776764,
-0.0452003852,
0.0058004679,
0.0376571938,
-0.0436700843,
0.0575964674,
-0.0125362268,
0.0125563042,
0.0437152907,
0.0840419158,
0.0039186943,
-0.0018707947,
-0.0160515588,
-0.0083931852,
-0.0732458606,
-0.0497437604,
0.0053137373,
0.0301594604,
0.0137373814,
-0.0110147903,
-0.0230497848,
-0.0438300818,
0.035891328,
0.0202628467,
0.0437023006,
-0.0574349575,
-0.0161081143,
-0.0349560902,
0.0258336756,
-0.0101799415,
0.0179776736,
-0.0091315201,
0.0049158721,
-0.0413807929,
-0.0259990338,
-0.0403708182,
-0.0276882332,
0.0276785363,
-0.0231512114,
-0.0557997115,
-0.0109364977,
0.0038689049,
0.1046862602,
0.0039459164,
0.0409481041,
-0.0190072972,
-0.0620057508,
0.0172637235,
0.0231151972,
0.0347195044,
0.0164869092,
0.0610781014,
0.0290366765,
-0.0152984587,
-0.0268302765,
-0.022388529,
0.0336551405,
-0.0564813465,
-0.0435398072,
0.0229760036,
-0.0577085279,
0.0290189069,
-0.016410403,
0.0158843603,
-0.0176529214,
-0.0622526482,
-0.0116805146,
0.0099869948,
0.0046276213,
-0.0284557547,
0.0535955168,
0.003127493,
-0.0364825204,
0.0068289246,
0.0524271764,
0.0538799763,
-0.0001370085,
-0.0022497727,
-0.0558051653,
-0.0229746066,
0.0659961998,
0.0058537577,
0.0340585038,
-0.0019282536,
-0.0364029668,
-0.0182880536,
0.0032019643,
0.0567587391,
0.0190268159,
0.0210897587,
0.0378452651,
0.0879338384,
0.0006277878,
0.0643920675,
0.0210288372,
0.0554893464,
0.0033517666,
0.0128603121,
-0.0286829602,
0.0145615088,
0.0109466622,
-0.0109929284,
0.025725225,
0.0363162756,
0.0851096585,
-0.015013488,
0.003594887,
0.0260080956,
-0.0240365751,
-0.0521879829,
-0.0029755882,
0.0553060137,
0.0359164216,
-0.0513215102,
-0.0031512596,
0.038083382,
-0.0614599921,
0.055738654,
-0.0321744755,
-0.0173246674,
0.0557311773,
-0.0172734875,
0.0044043176,
-0.0073589003,
-0.0235990826,
-0.0293463916,
0.0231523495,
0.037256863,
0.0125164138,
0.0716610327,
-0.0642484501,
0.0087394184,
0.0275542736,
-0.0091861859,
0.0754347891,
-0.0817399994,
0.0867280886,
0.0112670939,
-0.0447305478,
0.0170421246,
-0.0147630526,
0.0052543334,
-0.0340814702,
0.0230728257,
0.0304740649,
0.0305876676,
0.04262558,
-0.0188194588,
0.0274704807,
0.002132718,
-0.0171695184,
0.0288008172,
0.0116738053,
-0.0753817111,
-0.0301581863,
0.0318184532,
-0.0457323305,
0.0196361002,
0.0094585186,
0.0436603874,
-0.09676148,
0.0580176599,
0.0124089438,
0.0297044944,
-0.0121694384,
0.0227939095,
0.0040114727,
0.0069208313,
0.0118244337,
-0.0686717629,
0.042482093,
-0.0044253357,
0.0320292041,
-0.0090750931,
0.0235868692,
-0.010220198,
-0.0172978342,
0.0581474677,
-0.030867938,
0.0326326787,
-0.0013572898,
0.0407405198,
-0.0069514532,
-0.031480968,
-0.0410784483,
0.0099202739,
-0.0104697393,
0.0524036512,
0.0078060371,
-0.0034992998,
-0.0246381722,
0.0037412131,
-0.0343324877,
-0.0392317921,
-0.0064242827,
-0.0112403212,
-0.02315614,
-0.004200778,
0.0161115024,
-0.0219341405,
-0.0014875858,
0.030974593,
-0.0032763358,
-0.047676418,
-0.031997826,
-0.0195240658,
0.0035144305,
-0.0131907137,
-0.020010842,
0.0427591652,
-0.0409429297,
0.0085878847,
-0.0009976077,
0.096142754,
0.0073622582,
-0.021085646,
-0.0215874966,
0.0331994258,
-0.0476566739,
-0.0149046537,
-0.0164314322,
0.0024940548,
-0.0281565879,
-0.0192705635,
-0.0169286355,
0.0036406899,
0.0169977192,
0.0207012352,
-0.0010220654,
-0.0855504349,
-0.0304613709,
-0.0137227057,
-0.0233543813,
0.0477452688,
0.0301176384,
-0.02459394,
-0.0108896168,
0.0108168721,
-0.0358895026,
0.0073069022,
-0.0788319409,
-0.0479520671,
-0.0533855557,
-0.012679724,
0.0116391554,
-0.0456527248,
-0.0373147875,
0.0322981253,
-0.0345623679,
-0.0148860607,
-0.0439589024,
0.0152958585,
0.0264667682,
-0.0579945594,
-0.0138619291,
-0.01197269,
-0.03572575,
-0.0085701281,
0.0023678101,
0.0326111093,
0.0172039568,
-0.0094956635,
0.0153259328,
0.0396316759,
-0.0099639418,
0.0109772328,
0.0202999022,
0.0331248567,
0.0075994218,
0.0135927908,
0.0165806841,
-0.012565,
-0.003585249,
0.0262127519,
0.0311452616,
0.0097028464,
-0.0401588902,
-0.0087374756,
-0.0468781255,
0.0332523324,
0.014927269,
-0.056272734,
0.0144718969,
0.0539258495,
0.0496141575,
0.0146166841,
0.0194578804,
0.0740255192,
0.0056117252,
0.0161455907,
0.0775449872,
-0.0345433243,
-0.0224556345,
-0.0193738583,
0.0151319308,
0.0445560627,
0.0019657968,
-0.014561845,
0.0059142071,
-0.0338655412,
0.0327881202,
0.0529534705,
-0.0012333295,
0.0238196217,
-0.0387224369,
-0.1202109158,
0.0032119851,
0.030311998,
-0.0741378814,
-0.08607243,
-0.0523030534,
-0.0306609273,
0.0206976216,
0.0126818921,
0.001240454,
-0.0453621373,
0.0043881745,
0.046254646,
-0.0812826231,
-0.0636741593,
0.0007942506,
-0.0497462787,
0.0898575634,
0.0692726672,
0.0214544013,
-0.027417643,
-0.0314141549,
0.0105634583,
-0.0157800708,
-0.0122782402,
0.0023173906,
0.0022093265,
0.012616762,
-0.0160162058,
-0.0060865171,
0.0147788683,
-0.0171054881,
0.0312589221,
0.0216076635,
0.0627180263,
0.0313527733,
-0.0325687379,
0.035222996,
0.0505273491,
0.0190903209,
0.0386179686,
0.0336240493,
-0.0963559002,
-0.0038996255,
0.0227769464,
0.0223351065,
-0.0450143814,
0.0062938337,
0.0391908102,
-0.0466459543,
-0.0291253738,
-0.0555739962,
-0.0018699419,
0.0023909574,
0.0727732629,
0.051460471,
0.014684259,
-0.0056333523,
-0.0023709782,
0.0237577055,
0.0160253104,
-0.0100179492,
0.0408160277,
-0.0639844388,
-0.0181330927,
-0.0089759389,
-0.0582905076,
0.0043236427,
-0.0366964191,
-0.0251055956,
-0.0372779816,
0.0057028024,
0.0420925245,
0.0113070942,
0.0344936065,
0.0097048357,
-0.0086185252,
0.0330434814,
0.0425120518,
0.0032744471,
0.0361011289,
0.0978827402,
-0.0144139798,
-0.0056849527,
0.0174485762,
-0.0148154004,
0.047302261,
0.0023820072,
0.0161635168,
0.0604115129,
-0.0840630308,
-0.0236409437,
0.0315721072,
0.0464646965,
-0.014203338,
-0.0057267151,
-0.0617015734,
0.012121168,
-0.0045253406,
0.0673242435,
-0.0551186763,
-0.0321210809,
0.049505651,
0.0437686145,
0.0039303354,
0.0014833797,
0.0028097024,
-0.0258190017,
0.0121419057,
0.0473866425,
0.0279996283,
0.0477435887,
0.0267652404,
0.0209596939,
-0.0349488892,
0.0260202512,
0.0001101758,
-0.0069949306,
-0.0275400504,
-0.0052339002,
0.012825314,
-0.0002870034,
0.0241978317,
0.029621277,
0.0271336623,
-0.0100820083,
-0.0447237603,
-0.0248959847,
0.0418742001,
-0.0075896746,
0.1085309014,
-0.0061190031,
0.0337877236,
-0.0207854733,
0.0265732016,
-0.0635567233,
0.069383502,
-0.017320903,
-0.0381964445,
-0.005192169,
0.0232962165,
0.0195649937,
-0.020595517,
0.0046061785,
-0.0311054587,
0.0390327312,
0.0042638821,
-0.0181087833,
-0.0560168996,
-0.0173924118,
-0.0044509484,
-0.0087212017,
-0.058647804,
-0.046563942,
0.0075962632,
0.001762916,
0.0013976385,
-0.0169424992,
0.0171067212,
0.0136367166,
-0.0703430921,
0.0212120768,
-0.0027268187,
-0.0466353595,
0.0262404568,
0.0155121172,
0.0484188274,
-6.268269259e-33,
-0.0159615744,
0.0412674472,
0.0141232498,
0.0102960337,
-0.0570388958,
-0.0408693403,
-0.0406132489,
0.0012998328,
-0.0293323603,
0.0083093792,
-0.0017342607,
-0.0663927943,
0.0277741719,
0.0359722637,
-0.0435858071,
0.0253310408,
0.0609672964,
0.0112335552,
-0.0138533693,
0.028818788,
0.0342259929,
0.0352219529,
0.0769240335,
-0.0559444577,
-0.0400368497,
0.0097758379,
-0.0098045897,
-0.0520010814,
0.0002329482,
0.0055540227,
0.0311336555,
0.061229948,
-0.0153253358,
-0.0125448825,
-0.0493924692,
0.0295411646,
-0.0288460255,
-0.020557493,
0.0161677338,
0.0420120843,
0.011052751,
0.012785905,
0.0065009575,
0.0019837422,
-0.0438961647,
-0.0070963185,
0.0073025748,
-0.0630662069,
-0.0124888653,
-0.0713896155,
-0.0647153482,
0.0332384594,
-0.0082673877,
-0.0292575154,
0.0351161398,
0.0536297671,
-0.0611863583,
-0.0089354822,
-0.0085139787,
0.0251734797,
0.0155198518,
0.0344316065,
0.054359246,
0.0296915323,
-0.027937118,
0.000820087,
0.0296645332,
-0.0319161266,
0.0803916603,
-0.0096472288,
0.0942536965,
0.0434937999,
0.0387903787,
0.0570797138,
0.0346284956,
-0.0597070195,
-0.0147241261,
-0.0066756387,
-0.0312005207,
-0.0345503688,
0.0126764923,
0.01669663,
-0.0061710426,
-0.006272845,
-0.0804573447,
-0.0237716734,
0.0045557851,
-0.0085081756,
0.0648052469,
-0.0369195826,
0.0014324581,
0.0144827105,
0.094332546,
0.0168542806,
0.0110904882,
-0.0472346209,
-0.0830319971,
0.0078709675,
-0.0246991515,
0.0081674764,
-0.0841240883,
-0.0571493097,
0.0436667688,
0.0104070427,
0.0265006032,
-0.022472918,
0.0087891817,
0.0531923883,
-0.0617585368,
0.0045300554,
-0.0207974035,
0.0319219381,
0.0337043032,
0.0192384124,
0.014392429,
-0.0281196199,
-0.0039025575,
0.0588176474,
-0.0136066936,
-0.0142803611,
0.0243661255,
-0.0081093116,
0.0189016648,
0.0084602162,
-0.0470612869,
0.0329773128,
-0.0053804484,
-0.0211730972,
0.032780461,
-0.0209327731,
-0.0266303979,
-0.0095435446,
3.006e-7,
0.0219772104,
0.0153141273,
0.0092048645,
-0.0625602528,
0.0027965645,
0.0344564468,
-0.0575367436,
0.0136982976,
0.0383277871,
0.0486043356,
-0.009101796,
-0.0073019261,
0.0933157653,
0.0172625054,
-0.0557008199,
-0.021621814,
-0.0293369927,
0.0288851336,
-0.0368604735,
0.0150662269,
0.0080214469,
0.0277498402,
-0.0363848023,
-0.0470458381,
0.0199650824,
-0.0346761756,
0.0221087169,
-0.0392764509,
0.0278401002,
-0.0044377218,
0.0306032263,
-0.0367474817,
-0.0214624386,
0.0133825485,
-0.0684913695,
0.0457688086,
0.0366669819,
0.0387450121,
-0.0001563292,
-0.0482585989,
0.0546177477,
0.027728891,
-0.0526926704,
0.001770968,
0.05381383,
-0.0180951953,
-0.0072040684,
-0.021135103,
-0.0353195108,
0.0134262256,
0.0123505853,
-0.0448386483,
0.0060637477,
0.0196222421,
0.0191847775,
0.0246832203,
-0.0650453344,
0.0078448569,
0.0041655847,
0.0544399731,
-0.0116175013,
-0.0553955398,
-0.025044566,
-0.0175452642,
0.0231545735,
-0.0287684649,
0.0022399903,
2.351913026e-34,
0.0133917267,
-0.0638485402,
0.0640382767,
-0.0468539074,
0.0045459769,
-0.0375328511,
0.0158659406,
0.0315213352,
-0.0073940111,
-0.0166953411,
-0.0357076526
] |
http://arxiv.org/abs/2405.15039v1
|
http://arxiv.org/pdf/2405.15039v1
|
CEEBERT: Cross-Domain Inference in Early Exit BERT
| 2024-05-23T20:36:10 | 2024-05-23T20:36:10 |
Divya Jyoti Bajpai, Manjesh Kumar Hanawal
|
cs.CL
|
cs.CL, cs.LG
|
Pre-trained Language Models (PLMs), like BERT, with self-supervision
objectives exhibit remarkable performance and generalization across various
tasks. However, they suffer in inference latency due to their large size. To
address this issue, side branches are attached at intermediate layers, enabling
early inference of samples without requiring them to pass through all layers.
However, the challenge is to decide which layer to infer and exit each sample
so that the accuracy and latency are balanced. Moreover, the distribution of
the samples to be inferred may differ from that used for training necessitating
cross-domain adaptation. We propose an online learning algorithm named
Cross-Domain Inference in Early Exit BERT (CeeBERT) that dynamically determines
early exits of samples based on the level of confidence at each exit point.
CeeBERT learns optimal thresholds from domain-specific confidence observed at
intermediate layers on the fly, eliminating the need for labeled data.
Experimental results on five distinct datasets with BERT and ALBERT models
demonstrate CeeBERT's ability to improve latency by reducing unnecessary
computations with minimal drop in performance. By adapting to the threshold
values, CeeBERT can speed up the BERT/ALBERT models by $2\times$ - $3.5\times$
with minimal drop in accuracy.
|
CEEBERT: Cross-Domain Inference in Early Exit BERT: Pre-trained Language Models (PLMs), like BERT, with self-supervision
objectives exhibit remarkable performance and generalization across various
tasks. However, they suffer in inference latency due to their large size. To
address this issue, side branches are attached at intermediate layers, enabling
early inference of samples without requiring them to pass through all layers.
However, the challenge is to decide which layer to infer and exit each sample
so that the accuracy and latency are balanced. Moreover, the distribution of
the samples to be inferred may differ from that used for training necessitating
cross-domain adaptation. We propose an online learning algorithm named
Cross-Domain Inference in Early Exit BERT (CeeBERT) that dynamically determines
early exits of samples based on the level of confidence at each exit point.
CeeBERT learns optimal thresholds from domain-specific confidence observed at
intermediate layers on the fly, eliminating the need for labeled data.
Experimental results on five distinct datasets with BERT and ALBERT models
demonstrate CeeBERT's ability to improve latency by reducing unnecessary
computations with minimal drop in performance. By adapting to the threshold
values, CeeBERT can speed up the BERT/ALBERT models by $2\times$ - $3.5\times$
with minimal drop in accuracy.
|
[
0.0099521587,
0.0343864746,
0.0047809547,
-0.0128506329,
-0.0201090388,
0.0107899848,
0.026534196,
-0.0327919833,
0.0115797557,
-0.0262214076,
-0.0635849014,
-0.0275165625,
-0.0210612956,
0.0163037423,
0.0042900192,
-0.0620856881,
0.0274872817,
-0.0459797643,
0.0301374197,
-0.0119798975,
-0.0095008826,
0.0017768447,
-0.0305926111,
0.0440960415,
0.0153500363,
-0.0011758952,
-0.0128954072,
0.0247461386,
0.002126198,
-0.0332229882,
0.0432699546,
0.0572794601,
0.0403860956,
-0.0033689891,
0.0000022471,
-0.0650436059,
-0.0457591675,
0.0023334723,
-0.0273459088,
0.032484252,
-0.0424778238,
0.0002329012,
0.0094132321,
0.0643690005,
-0.0087235877,
-0.0387709439,
0.0479758978,
0.0202703159,
0.0325224586,
0.0677512363,
-0.0064512077,
0.0045295102,
-0.016246438,
-0.0365618877,
-0.0214246344,
-0.068997018,
0.0222279243,
-0.0469898283,
-0.0000021677,
-0.0530745946,
-0.0258894172,
0.0344025008,
0.0249641519,
-0.0062527787,
-0.0456406809,
0.0358200073,
-0.026496768,
-0.0338038765,
-0.0257378239,
0.0539635606,
-0.0113652209,
-0.0301895887,
0.0211637393,
-0.0306096319,
-0.0074066678,
0.0290147141,
-0.008694184,
0.023934707,
-0.0117708407,
0.0083657177,
-0.0143199759,
0.0118518444,
-0.0014894786,
0.0223404057,
-0.0472979769,
0.0381761789,
0.0281066112,
-0.0091329664,
0.0074914866,
0.0423362516,
-0.0530032776,
-0.0379197448,
0.0349184126,
0.0311670545,
0.0031256506,
0.0117208557,
-0.0927028134,
-0.0244624484,
-0.0375863016,
-0.0280353744,
0.0714122877,
0.0380198658,
-0.0581803843,
0.0487330668,
-0.0102838296,
0.001002124,
-0.0265364703,
0.0205604564,
-0.0187507831,
-0.0163857006,
0.0027011223,
0.0208290517,
-0.0743029937,
0.0250111409,
-0.0055936165,
-0.0255400743,
-0.0728240386,
0.0315522589,
0.0265818872,
-0.0454118848,
0.0384761542,
-0.0143485758,
-0.0155435549,
-0.0302929431,
0.0044351327,
0.0307148229,
-0.0273249205,
0.0254509635,
-0.028926583,
-0.0370433666,
-0.0221431814,
0.0305880941,
0.0423778333,
-0.0446027778,
0.0384199955,
0.0875116736,
0.0252760891,
-0.0510649718,
-0.0599959828,
0.0027109303,
0.0036048489,
-0.0549931191,
0.0398364775,
-0.0351756401,
0.0240546782,
0.0181751363,
-0.0104382364,
-0.0096087828,
-0.0164467301,
-0.0041053607,
-0.0171209425,
0.0379952639,
-0.0371286385,
0.0324779078,
0.0397371054,
-0.0138504198,
-0.0107091926,
0.0709966868,
0.0206905138,
0.0032160473,
0.0531176217,
0.0017745007,
0.0115586789,
0.0633448064,
0.0712597296,
-0.0154907415,
-0.0374140963,
-0.013417827,
-0.0255162343,
0.0144763552,
0.0302322693,
0.0157655925,
-0.0465586931,
0.0148404194,
0.0755913407,
0.0375127569,
0.0088932812,
0.0054225093,
0.0653955415,
0.0301263314,
0.0358516537,
-0.0042645722,
-0.0540690981,
0.0185900498,
-0.0184341911,
0.0067352974,
0.0540643148,
0.0246410519,
-0.0435051322,
-0.0281769019,
-0.0041497806,
-0.021113541,
-0.0439785048,
0.0204271302,
0.0747161955,
0.022666974,
-0.0485007763,
0.0597106591,
0.0320234634,
-0.0181901604,
0.0038801311,
-0.0581540577,
0.0138789034,
0.0426658541,
0.022754319,
-0.0325807482,
-0.0517125241,
0.0313998275,
0.0546362586,
0.0414452516,
0.0244333576,
0.0067395046,
0.0525920242,
-0.0362729579,
-0.0517055653,
0.0262205787,
0.0121919271,
0.0366472229,
0.0499588549,
0.0139043164,
-0.0104605956,
-0.088615641,
0.0448000692,
0.0303529613,
-0.0481199585,
-0.0155860782,
0.0169954561,
0.0334120244,
0.0010148014,
-0.01028373,
-0.0270819124,
0.0032773793,
-0.0034339563,
-0.0025182765,
-0.0081020892,
0.0059530265,
-0.0354210027,
-0.0473357365,
0.013033337,
-0.0198957045,
-0.0089970669,
0.0335772671,
-0.0139637254,
-0.0126642343,
0.0214930642,
-0.0251485165,
0.0888039023,
-0.0479300246,
0.0193751492,
0.0037456399,
-0.0088651115,
0.0314339139,
0.0401312113,
-0.0034078276,
0.0061090989,
0.0249269512,
0.0261321515,
-0.0113630714,
0.0068415785,
-0.0428429954,
-0.0541680604,
-0.0268906262,
-0.0082000699,
0.0275065545,
0.0267113466,
-0.0228435285,
0.0149764577,
-0.073540099,
-0.0045833364,
-0.0226437096,
0.0386181884,
0.0205040351,
0.0181628447,
-0.0411591046,
-0.0393994823,
0.0359469242,
-0.023743296,
-0.0509786196,
0.0313068964,
0.0240495354,
0.0084216604,
-0.0119706579,
-0.0334960036,
-0.0079750102,
-0.010364661,
0.0478296541,
-0.025115855,
-0.0360765569,
-0.0332575291,
-0.0012465209,
-0.0205597393,
0.0823934451,
0.0134301921,
-0.0387829952,
-0.015101782,
-0.0462479033,
0.0392476134,
0.0425308719,
0.0020440819,
0.0178320594,
0.0279728658,
-0.0248114876,
-0.045927614,
0.0186485648,
-0.0390012786,
-0.0033807599,
-0.0076371706,
0.0280162953,
0.0007741326,
0.1005688906,
-0.0291815046,
-0.0120186852,
-0.0458808057,
-0.0357836857,
-0.000520565,
-0.0084246043,
0.0085552055,
-0.0209558923,
-0.0128990207,
0.0267975498,
-0.0002039892,
-0.0510347635,
0.0856478661,
-0.0317684785,
-0.0018363276,
-0.0412532017,
0.0239511486,
-0.011128447,
0.0576026253,
0.002385674,
-0.0066232421,
0.0575388782,
-0.0569308065,
-0.0455084369,
0.0193606187,
0.0411660261,
-0.02330634,
-0.0218915921,
-0.0757477432,
0.009218215,
0.0144343516,
-0.0410594605,
-0.0258925334,
-0.0377340987,
-0.0221454315,
-0.0124293733,
0.0022504276,
-0.0370060317,
-0.0632641092,
-0.0148302075,
-0.0225008391,
0.1101517081,
-0.0166252851,
0.0194559339,
0.0003278752,
0.0586608127,
0.0227639675,
-0.0016607258,
0.0043618269,
-0.0606425256,
0.0140634738,
0.0210341942,
0.0306211226,
-0.0182119161,
-0.0298963245,
-0.0030112369,
0.0624584705,
-0.0349698775,
-0.0294957962,
0.042965021,
0.0476784296,
-0.0593376011,
-0.0294166449,
0.1543516517,
-0.000554644,
-0.0205815248,
-0.0329534672,
0.0179971308,
0.0031483651,
-0.0072298972,
0.0344357677,
-0.0385369733,
-0.0460170545,
0.0157329813,
0.0454179049,
0.0185462683,
-0.0238080248,
-0.0101985773,
-0.133900404,
0.0262453612,
-0.0494279377,
-0.0152045935,
-0.0028492233,
-0.0532833636,
0.006178089,
0.0029621648,
-0.0082901251,
-0.0188221633,
0.0019470199,
-0.006134999,
0.0445711836,
-0.0228175987,
-0.0499807261,
0.0179124679,
-0.0756157711,
0.032199461,
0.0562998019,
0.0163653344,
0.0108045209,
-0.0335041024,
0.102590546,
-0.0630540699,
-0.0263984129,
-0.0290039852,
0.0674361289,
-0.0442968085,
-0.0157128274,
-0.0128077017,
0.0337557457,
0.0052525839,
0.0091737267,
-0.0150969448,
0.0554020666,
-0.0022375095,
-0.0310191065,
-0.0216162,
0.0333697684,
-0.0050549125,
0.0555950962,
0.0221551731,
-0.0935610756,
-0.02151368,
0.019299088,
-0.0153986141,
-0.0590592623,
0.0095854569,
-0.0110436464,
-0.0474682041,
-0.0640755817,
-0.042824626,
-0.0067281062,
0.0456714965,
0.0841374546,
0.0066992636,
-0.0177162495,
0.0280148201,
0.0366652086,
-0.0509885326,
0.0227138195,
-0.0040138457,
0.0014826742,
-0.0785501003,
-0.0088801906,
0.0686018765,
-0.0359646752,
-0.0295382123,
0.0030258833,
-0.0398337319,
-0.0650395676,
-0.0025394228,
0.0442348905,
0.017286187,
0.0554852672,
-0.0348581411,
-0.0577813722,
0.0849096701,
0.0240382645,
-0.0248035789,
-0.0127580408,
0.0562713929,
-0.0069520851,
0.0027227348,
0.0342015997,
-0.059264075,
0.0395690128,
0.0337983593,
0.0348241478,
-0.0139566809,
-0.0373809375,
0.0067671086,
-0.0123550082,
0.098875694,
-0.0159674212,
-0.0049295458,
-0.056949947,
0.0355043747,
0.0159398504,
-0.0101534287,
-0.0579816885,
0.0391157307,
0.014413327,
0.0124911424,
-0.034583997,
0.0578206703,
0.0176775828,
-0.0131215267,
-0.0460257567,
0.0373056531,
0.0120548829,
0.0680605546,
-0.0001193135,
-0.0345397443,
0.0044347304,
-0.001288843,
0.0314220823,
0.022909712,
-0.0285999067,
0.0502386093,
0.0283418205,
0.0192623977,
0.0112668443,
0.0094028991,
-0.0502446778,
-0.0121336402,
-0.086168997,
0.048248753,
0.0217364095,
0.0054037026,
0.0695661306,
-0.0085105458,
-0.005440962,
-0.0248602536,
0.0329328217,
-0.0161147211,
-0.0205180459,
0.0228646155,
0.0027173131,
0.000647345,
-0.0000999452,
0.069383949,
0.0209147818,
0.0634267703,
0.0085897138,
-0.0199883226,
-0.0155280121,
-0.0357972756,
0.0504010692,
-0.0152904857,
-0.03849544,
0.021225052,
0.0168076884,
-0.0715636387,
0.0015900225,
0.0235838797,
-0.027113786,
0.0275747683,
0.0417105369,
0.0705392435,
-0.0019615204,
-0.0157846753,
-0.0152201978,
-0.0024036411,
0.0297812186,
-0.0079367245,
0.0211469997,
-7.167592233e-33,
-0.036053367,
0.0110028973,
0.0152357621,
-0.0272387154,
-0.0152559336,
0.0140355816,
-0.035752885,
0.0156517755,
-0.0105807772,
-0.0111151692,
-0.0061947871,
0.0374428742,
0.0331967659,
0.0236714762,
-0.0198077112,
-0.0066187829,
0.0498085432,
-0.0287365932,
-0.0244255848,
-0.0102045583,
-0.0244426522,
0.0776276663,
0.0745830387,
-0.0298273154,
-0.0237232503,
0.0124571966,
-0.0136151994,
0.0239746571,
0.0499212258,
0.0386271924,
-0.0108243199,
-0.0209600907,
-0.0130138872,
-0.0116762202,
-0.019016711,
0.0144207822,
-0.1087896228,
-0.0359256305,
-0.0141154351,
0.0389229096,
-0.0389732048,
0.0338211879,
0.0141355423,
0.010004675,
0.0181306843,
0.0058090356,
0.0448358431,
0.0005220667,
-0.0021502348,
-0.079966493,
-0.0393566117,
0.0474937521,
0.0048677744,
-0.0714235306,
0.020319622,
0.068891488,
-0.0037762325,
-0.0150388908,
-0.0277785361,
0.0260306913,
0.0053402656,
0.0134268245,
0.034914542,
0.0071045551,
0.001807653,
0.0333741233,
0.0210804753,
-0.0107181827,
0.0284333099,
-0.0262909774,
0.0725612491,
0.0215834714,
-0.0203659702,
-0.0198764205,
0.0123693319,
0.0170700103,
-0.0691475794,
0.0347415097,
0.0047810148,
-0.0126986559,
0.0336058922,
-0.0212439653,
-0.0228695478,
-0.0101531986,
-0.0213826001,
-0.0451379642,
-0.0233720765,
-0.0150803551,
0.031354133,
0.0313697867,
-0.043508552,
-0.0192229524,
0.0353629664,
-0.0213662945,
-0.0313201137,
-0.0276368205,
-0.003772699,
-0.0048320186,
-0.0091565512,
0.0448554195,
-0.061956808,
-0.0569702461,
-0.0022926521,
-0.0186930094,
0.0019848163,
-0.0100152045,
0.0301148705,
0.0401971862,
-0.0842762887,
-0.0114442352,
-0.0274027884,
0.0000094507,
0.0121158846,
0.0534456931,
0.0606567189,
0.0042496347,
0.0015996274,
0.0539643429,
0.0089441603,
0.0223802123,
-0.0250323191,
-0.0053498913,
0.0227012485,
0.027305292,
-0.025980657,
-0.0260485373,
-0.0251641255,
0.0098839439,
-0.0338133834,
-0.0828300118,
-0.0339309871,
0.023179099,
3.039e-7,
0.0569635443,
0.0321660005,
0.0372837968,
0.0066290488,
-0.0152822323,
-0.0003914353,
-0.0064646397,
0.0210549813,
0.0047304602,
0.0276006814,
0.035533037,
-0.0041770772,
0.0476321355,
0.001843976,
-0.0853135884,
-0.0423379205,
-0.0304343533,
0.0295655392,
-0.0135152796,
-0.0171075687,
0.0143432049,
0.0865279883,
0.0119467843,
-0.0092604049,
0.0085209161,
-0.0260538645,
0.0662318021,
-0.0116201518,
-0.0018647264,
0.0204189569,
-0.0148218255,
0.0404355898,
-0.0339126587,
-0.001863247,
-0.0282439366,
0.0413997546,
0.0052815531,
0.0400786959,
0.0280021913,
-0.0305607524,
0.0273358561,
0.0177594461,
0.0039927955,
0.0099991653,
0.0748377517,
-0.0928811729,
-0.0272348877,
0.0174935106,
0.0006538684,
-0.0429135002,
0.0133734718,
-0.0390683338,
-0.0191572756,
0.0197698772,
0.041876357,
-0.0376553312,
-0.0158258155,
-0.0870974287,
-0.011449133,
0.0135275619,
-0.040850509,
-0.0356178284,
-0.0410012491,
-0.0301055163,
0.0571751632,
-0.0593851507,
0.0335634984,
2.751634599e-34,
0.0445362143,
-0.0459044538,
-0.0130475219,
0.0294714198,
-0.0092645641,
-0.0014605954,
0.0319532938,
-0.0236705914,
0.025660282,
-0.0450642183,
-0.0281414017
] |
http://arxiv.org/abs/2406.00021v2
|
http://arxiv.org/pdf/2406.00021v2
|
CrossVoice: Crosslingual Prosody Preserving Cascade-S2ST using Transfer Learning
| 2024-05-23T20:30:54 | 2024-06-18T05:26:48 |
Medha Hira, Arnav Goel, Anubha Gupta
|
cs.CL
|
cs.CL, cs.SD, eess.AS
|
This paper presents CrossVoice, a novel cascade-based Speech-to-Speech
Translation (S2ST) system employing advanced ASR, MT, and TTS technologies with
cross-lingual prosody preservation through transfer learning. We conducted
comprehensive experiments comparing CrossVoice with direct-S2ST systems,
showing improved BLEU scores on tasks such as Fisher Es-En, VoxPopuli Fr-En and
prosody preservation on benchmark datasets CVSS-T and IndicTTS. With an average
mean opinion score of 3.75 out of 4, speech synthesized by CrossVoice closely
rivals human speech on the benchmark, highlighting the efficacy of
cascade-based systems and transfer learning in multilingual S2ST with prosody
transfer.
|
CrossVoice: Crosslingual Prosody Preserving Cascade-S2ST using Transfer Learning: This paper presents CrossVoice, a novel cascade-based Speech-to-Speech
Translation (S2ST) system employing advanced ASR, MT, and TTS technologies with
cross-lingual prosody preservation through transfer learning. We conducted
comprehensive experiments comparing CrossVoice with direct-S2ST systems,
showing improved BLEU scores on tasks such as Fisher Es-En, VoxPopuli Fr-En and
prosody preservation on benchmark datasets CVSS-T and IndicTTS. With an average
mean opinion score of 3.75 out of 4, speech synthesized by CrossVoice closely
rivals human speech on the benchmark, highlighting the efficacy of
cascade-based systems and transfer learning in multilingual S2ST with prosody
transfer.
|
[
-0.0348978974,
0.0168749057,
0.0055781612,
-0.0062207631,
-0.00002526,
-0.0282571837,
0.0570227392,
-0.0258295555,
-0.0547108017,
0.0195622426,
-0.0928127021,
-0.0263844673,
0.0454998091,
0.0591525994,
0.0482419655,
-0.0759497285,
0.0435443334,
-0.0347443968,
-0.0236505605,
-0.0213353205,
0.0024618094,
0.009027916,
-0.0143315373,
-0.0159586389,
-0.0101751396,
0.0270829499,
0.0211659167,
0.0082112774,
0.0544928052,
-0.0176176857,
-0.0482875258,
-0.0131669594,
0.0247837175,
0.011122629,
0.0000022537,
-0.0391816534,
-0.04457004,
0.0038998877,
-0.0107769389,
0.0005875136,
0.0183503013,
-0.0027114386,
-0.0328458324,
0.0463668443,
-0.0205418989,
-0.0219731685,
0.0497683845,
0.007711689,
0.0025678286,
0.0453406163,
-0.0315787606,
-0.0216654185,
-0.0021407451,
-0.0249083005,
-0.0018853822,
-0.0309075806,
-0.0277804229,
-0.044817023,
-0.0468485691,
-0.0274158921,
-0.0552001707,
0.0523331501,
-0.0207286756,
-0.0505511723,
0.017024463,
-0.0394908153,
0.0075881509,
-0.0544032715,
-0.0166918654,
0.0682235882,
0.0326587558,
0.0162019338,
-0.0049648522,
0.0072022551,
-0.0346349962,
-0.0570684485,
-0.0144853322,
-0.0026010415,
0.01250719,
-0.0002719185,
0.005184473,
0.018702317,
-0.0003819374,
-0.0705620721,
-0.0521067865,
-0.0180978626,
0.0347117111,
-0.0143347615,
0.0432599373,
0.0169044752,
0.0356679857,
0.0240143649,
0.075113982,
-0.0078351442,
-0.035744261,
0.0019548393,
-0.0093276529,
-0.055458542,
-0.0556793101,
0.0357778184,
0.0416617729,
0.0414902791,
-0.0451102555,
-0.0290911086,
-0.050655894,
0.0298941582,
-0.0329436436,
-0.0036452627,
-0.0345133729,
-0.0249093324,
-0.0260974634,
0.0666643754,
-0.0415543653,
0.0175703652,
-0.0346493907,
0.0061622672,
-0.0173482355,
-0.017552251,
-0.0270457286,
0.0052502612,
0.0376325585,
-0.0170434378,
-0.0754666254,
-0.0193456523,
0.0306668244,
0.1147582084,
0.0271151233,
0.0290793832,
-0.0027230855,
-0.055510208,
0.0025711842,
0.0469200537,
0.0333379507,
0.0047792997,
0.0202674363,
0.014931215,
-0.0146968216,
0.0135897463,
-0.0386212692,
0.0154656684,
-0.0289653987,
-0.0335434191,
0.0171148181,
-0.0188010763,
0.0575740598,
-0.0210089963,
0.0019958422,
-0.0252047349,
-0.024864262,
-0.0006921066,
0.0155481761,
0.0264988076,
-0.0347951502,
0.0139788287,
0.0089292768,
-0.020824017,
-0.0547234714,
0.0212534256,
0.0645076334,
-0.007353663,
-0.0358440205,
-0.048634477,
-0.0551788807,
0.0381389037,
0.0311480779,
0.0169857144,
-0.0485295169,
-0.0016680501,
0.0262311548,
0.0255612731,
0.0708207265,
0.0682888255,
0.0260743424,
0.0059601907,
0.0663309097,
0.004457253,
0.0392073877,
0.0377036333,
0.0708548203,
-0.0036792709,
-0.0228164662,
-0.006818119,
0.0113503933,
0.0213271156,
-0.0123951137,
-0.0342581905,
-0.0087340372,
0.0208462458,
-0.0189763065,
0.0163684357,
0.0335501656,
0.005109203,
-0.0208015833,
-0.027247129,
0.026725186,
0.0613876693,
-0.0611942634,
0.0415812656,
0.0035867137,
-0.0840187296,
0.089192234,
-0.0078823753,
0.0002311253,
0.056492772,
-0.0092897573,
0.0177982803,
-0.0056545818,
-0.0583083294,
-0.0027234573,
0.029515503,
0.013630786,
0.0129730906,
0.1052061245,
-0.0665446222,
0.025832925,
0.0264282785,
-0.0199295413,
0.049877923,
-0.065959692,
0.0733650327,
-0.013961494,
0.0056347954,
0.0696518868,
-0.0081831766,
-0.0566101708,
0.0043459423,
0.0789763629,
0.0160981826,
0.0512311459,
0.0194822941,
-0.0147897629,
0.0138568599,
0.0095763803,
-0.0429271683,
0.0026867006,
0.0256371163,
-0.0680002347,
-0.0210341569,
0.0166455731,
-0.0234147105,
0.0145148141,
-0.0006286205,
0.0291099083,
-0.0545537472,
0.0320340805,
-0.0101920329,
0.0276977923,
-0.0140815089,
0.0090900259,
0.030971339,
0.0341536663,
0.038108062,
-0.0601114258,
0.0494665019,
-0.0499160998,
-0.004821206,
-0.0106165074,
0.014611843,
-0.0400812775,
0.0250420775,
0.0719449669,
-0.0185809545,
0.0253582168,
-0.0107859802,
-0.0062315124,
-0.0151812173,
-0.0071554207,
-0.0035780191,
0.0468440875,
0.0179411862,
0.0553985201,
-0.0195566751,
0.0212560892,
0.0031554091,
0.0159545168,
-0.033183042,
-0.008587732,
0.0243301224,
-0.0239960868,
-0.0077518788,
-0.0340845287,
0.0202092528,
-0.0142673245,
0.0101516154,
0.0205308385,
-0.0176023878,
-0.0770480186,
-0.0219172109,
-0.0267367158,
0.0231168885,
0.0044855461,
-0.015795175,
0.0255963802,
-0.0053697964,
-0.0086526498,
0.0172963794,
0.0678072795,
0.0398149379,
-0.0081054065,
-0.0529855378,
0.0084000891,
-0.042549409,
0.0000734246,
-0.0025112364,
0.0201514363,
-0.0161344185,
-0.0155289732,
-0.0032754412,
-0.0193740204,
-0.0108947009,
0.0372045413,
0.0090716518,
-0.0740257353,
-0.0410833396,
0.0161147956,
-0.023952093,
0.0258461088,
-0.0008656241,
-0.0104223015,
-0.0175050069,
-0.0242947638,
0.004678491,
-0.0174213033,
-0.0508496165,
-0.0527955256,
-0.0676486939,
-0.0003511578,
-0.0071500782,
-0.006997461,
-0.0567805097,
0.020999141,
-0.0023885572,
-0.0318042152,
-0.0340894721,
0.011470438,
0.0180440377,
-0.0617320724,
-0.0264784843,
-0.027990073,
-0.0403564237,
-0.0320572071,
-0.0134831006,
-0.0018829122,
0.0482593141,
0.010051542,
0.0347397998,
0.0364882872,
-0.0097681247,
0.0022983388,
-0.0001537008,
0.0625647753,
0.0278670136,
0.0257683694,
0.0260088574,
-0.0166390613,
0.0428520516,
0.0030648324,
0.0335226431,
0.019409284,
-0.0261327345,
0.0373026691,
-0.0048970287,
0.0264759585,
0.0203323886,
-0.0471465588,
0.0249841344,
0.041515924,
0.0353094116,
-0.0072890501,
-0.0058165938,
0.0796125829,
-0.0261581372,
0.0108393179,
0.0337374844,
-0.0499105752,
-0.0361726694,
-0.0048981048,
-0.0021908225,
0.0333601907,
0.0026123151,
-0.0348665304,
-0.0428183787,
-0.0123929493,
0.0528130457,
0.0223950464,
0.0380667001,
0.0236270763,
-0.0435385928,
-0.0967375264,
0.0533771068,
0.0354239494,
-0.0611599907,
-0.0252029561,
-0.0510773212,
-0.0413186513,
0.0035880092,
0.0536067933,
0.0138683729,
-0.0296198241,
-0.0329422578,
0.0262735933,
-0.0363157801,
-0.032860253,
-0.0121316239,
-0.0521993712,
0.077800177,
0.0539447777,
0.0443201065,
0.000408686,
-0.0429296866,
-0.0364808924,
-0.0577932745,
-0.0172136221,
-0.0243332423,
-0.002287358,
0.0252657756,
-0.0005432445,
0.0054173875,
0.0409978554,
-0.011730209,
0.0200993344,
0.0298407469,
0.0616354197,
0.0042516058,
-0.0471701995,
0.018490484,
0.0330629908,
0.0155738881,
0.0459711477,
0.0086705144,
-0.0574515946,
0.0300923046,
0.0229462143,
0.0124680288,
-0.0355488099,
0.0034250584,
0.0296590794,
-0.0376118235,
-0.06235658,
-0.051760789,
0.0152130397,
-0.0070536244,
0.0436393395,
0.0209129192,
0.01936551,
-0.0202604849,
-0.0193005204,
0.0002905795,
-0.0142057296,
-0.0310755856,
0.0466431528,
-0.0272131097,
0.0191382766,
-0.0020764354,
-0.0415181555,
0.0245115682,
-0.053731434,
-0.0154649382,
-0.0436150767,
-0.0104021728,
0.0143291568,
0.008581656,
0.0294772014,
0.0169069823,
0.0129159326,
0.0674357936,
0.0207516626,
-0.016374208,
0.0086720539,
0.0722454488,
-0.0120546455,
0.0187231991,
0.0412338302,
-0.0069840252,
0.01792253,
-0.0218887571,
0.02398213,
0.0632141009,
-0.026046481,
0.0025942409,
0.0151572088,
0.026048759,
-0.0329268053,
-0.031367626,
-0.0216045231,
0.0247077215,
-0.0445876457,
0.0786697939,
-0.0606807992,
-0.0139730703,
0.0782282427,
0.0242378321,
0.0169176664,
0.0120353457,
0.0056198272,
-0.0274781417,
0.0312990099,
0.0695532858,
0.0084558278,
0.0684769973,
0.0148365004,
0.0394410156,
-0.0143845864,
0.0069977217,
0.0357978083,
0.0165816545,
-0.0442923196,
-0.0072268737,
0.0122375684,
0.0108109824,
0.0310982447,
0.0131475478,
0.0295207929,
0.0070462092,
-0.014574823,
-0.035537459,
0.0375709198,
0.0235810969,
0.1462232322,
0.0257689487,
0.0320478342,
-0.0314116925,
0.0362045914,
-0.0756327882,
0.0564914718,
-0.0222778935,
-0.0528366193,
0.012458588,
0.0495909862,
-0.008821018,
-0.0226890314,
-0.0025106575,
-0.0260231774,
0.0335530192,
0.0044410541,
-0.0106309019,
-0.031426847,
-0.049582094,
-0.0130042462,
-0.0244365819,
-0.0343558602,
-0.0423967801,
0.018174639,
-0.0182111524,
0.0002558286,
-0.0409500413,
-0.0086530028,
-0.0004417285,
-0.0301025379,
-0.00967809,
-0.0080311317,
-0.0340530649,
0.0151371863,
-0.0119302832,
0.0287707988,
-6.914949844e-33,
-0.0269339941,
0.0598987602,
0.0302222148,
0.0177321341,
-0.083510153,
-0.0037359912,
-0.0807133317,
-0.0154226413,
-0.0278554764,
0.0126254587,
0.0028201721,
-0.0431039482,
0.0214995462,
0.0179200098,
0.0019750677,
0.0304776449,
0.054407347,
-0.0070398883,
-0.0252424758,
0.0685559586,
0.0334214084,
0.0392793082,
0.0587819703,
-0.0651451498,
-0.0477752052,
0.0343541279,
0.0081703607,
-0.0126844998,
-0.0195105914,
-0.0179096088,
0.0144544616,
0.0247204416,
-0.0213188026,
-0.0201189127,
-0.0365039483,
0.044804655,
-0.0897581726,
0.0180331748,
-0.002073952,
0.0178738311,
-0.0303184465,
0.0055253617,
0.0338675343,
0.0051903189,
-0.0182840861,
-0.0499609374,
-0.0126180528,
-0.0409899428,
-0.0084870225,
-0.0621384755,
-0.0597139224,
0.0257385392,
0.0056562317,
-0.031081073,
0.0544133,
0.0600696728,
-0.0365640633,
-0.0259858835,
-0.0477694757,
0.0045356904,
-0.0689516962,
0.0313937999,
0.0639402717,
-0.0262358133,
-0.0320094973,
-0.0184838511,
0.093588613,
-0.0367462002,
0.0465656221,
-0.0076577528,
0.0536300614,
0.053552635,
0.0528908223,
0.0362366885,
0.0170462672,
-0.0669009387,
-0.0412401259,
-0.0072706374,
0.0077839387,
-0.011751133,
-0.0177187957,
0.0466210321,
0.0458397567,
0.0053038024,
-0.0821464285,
-0.0608511902,
0.0073447851,
-0.0298773982,
0.0361649804,
-0.0280814618,
0.0068603097,
-0.0328615084,
0.1062032133,
0.0184284914,
-0.0003190839,
-0.0316194445,
-0.0472795814,
-0.0170021243,
-0.0222197752,
0.0242323484,
-0.0701751038,
-0.0486061648,
0.0370290913,
0.0093641402,
0.0132059352,
-0.0248090569,
-0.0151030915,
0.0539730228,
-0.0492123701,
-0.0013096891,
-0.0435727052,
0.0305391792,
0.0128454436,
-0.008201017,
0.0239963196,
-0.03097561,
0.0349288844,
0.0414692461,
-0.0110064149,
0.0255391859,
0.0192121118,
0.013877403,
0.0515278019,
0.0404950865,
-0.0473467857,
0.033937186,
-0.0004657937,
-0.009261569,
0.0545923337,
-0.0336760059,
-0.0310747158,
-0.0061713988,
3.041e-7,
-0.0144594945,
0.0430014618,
0.0228679162,
-0.063127622,
0.0177502278,
0.0404204205,
-0.0403252542,
0.0262076166,
0.0240646452,
0.0073886644,
-0.0229457598,
0.0062534413,
0.1356005818,
0.0514357239,
-0.0364273414,
-0.0495673232,
-0.0206148159,
0.0264623966,
-0.0227723885,
0.0146941785,
0.0375350155,
0.0261688009,
-0.0327895023,
-0.0643204972,
0.0304040108,
0.0111523438,
0.0184212662,
-0.0015749992,
0.0106879035,
-0.0242543686,
-0.0098676691,
-0.0441292748,
-0.0347377285,
0.0154050104,
-0.0624943897,
0.0489970744,
0.0003989397,
0.0410014316,
-0.0168480817,
-0.0745673478,
0.0436009839,
0.0520228073,
-0.0384718888,
0.0066200169,
0.0341905616,
0.0180175547,
-0.0125048533,
-0.0250478666,
-0.0570756793,
0.0369179994,
0.0181576908,
-0.0104817431,
-0.0075898562,
0.0009484414,
0.0140550118,
0.0577604659,
-0.0806971863,
0.0027459434,
-0.0050756456,
0.0130513255,
-0.0420363992,
-0.0446482375,
-0.0169296954,
0.0203789286,
-0.0352892131,
0.0280329511,
-0.0094208103,
2.502680737e-34,
0.0059308973,
-0.0756351948,
0.0509575903,
-0.0353202075,
0.0283205472,
-0.0601475313,
0.0110974293,
0.0224964879,
-0.019246215,
-0.0097816093,
-0.0436677001
] |
http://arxiv.org/abs/2405.15032v2
|
http://arxiv.org/pdf/2405.15032v2
|
Aya 23: Open Weight Releases to Further Multilingual Progress
| 2024-05-23T20:10:38 | 2024-05-31T14:47:55 |
Viraat Aryabumi, John Dang, Dwarak Talupuru, Saurabh Dash, David Cairuz, Hangyu Lin, Bharat Venkitesh, Madeline Smith, Jon Ander Campos, Yi Chern Tan, Kelly Marchisio, Max Bartolo, Sebastian Ruder, Acyr Locatelli, Julia Kreutzer, Nick Frosst, Aidan Gomez, Phil Blunsom, Marzieh Fadaee, Ahmet Üstün, Sara Hooker
|
cs.CL
|
cs.CL
|
This technical report introduces Aya 23, a family of multilingual language
models. Aya 23 builds on the recent release of the Aya model (\"Ust\"un et al.,
2024), focusing on pairing a highly performant pre-trained model with the
recently released Aya collection (Singh et al., 2024). The result is a powerful
multilingual large language model serving 23 languages, expanding state-of-art
language modeling capabilities to approximately half of the world's population.
The Aya model covered 101 languages whereas Aya 23 is an experiment in depth vs
breadth, exploring the impact of allocating more capacity to fewer languages
that are included during pre-training. Aya 23 outperforms both previous
massively multilingual models like Aya 101 for the languages it covers, as well
as widely used models like Gemma, Mistral and Mixtral on an extensive range of
discriminative and generative tasks. We release the open weights for both the
8B and 35B models as part of our continued commitment for expanding access to
multilingual progress.
|
Aya 23: Open Weight Releases to Further Multilingual Progress: This technical report introduces Aya 23, a family of multilingual language
models. Aya 23 builds on the recent release of the Aya model (\"Ust\"un et al.,
2024), focusing on pairing a highly performant pre-trained model with the
recently released Aya collection (Singh et al., 2024). The result is a powerful
multilingual large language model serving 23 languages, expanding state-of-art
language modeling capabilities to approximately half of the world's population.
The Aya model covered 101 languages whereas Aya 23 is an experiment in depth vs
breadth, exploring the impact of allocating more capacity to fewer languages
that are included during pre-training. Aya 23 outperforms both previous
massively multilingual models like Aya 101 for the languages it covers, as well
as widely used models like Gemma, Mistral and Mixtral on an extensive range of
discriminative and generative tasks. We release the open weights for both the
8B and 35B models as part of our continued commitment for expanding access to
multilingual progress.
|
[
0.011209094,
0.0802105591,
0.0224535745,
0.0610699691,
0.001476203,
0.0167509727,
0.073190704,
-0.0066126115,
-0.0118431868,
0.0116345854,
-0.0821373314,
-0.0967191905,
-0.0068791523,
0.021927353,
0.037094947,
-0.0307945255,
0.0382903889,
-0.0522369556,
-0.018941069,
-0.0271370951,
-0.0122968489,
0.0269296002,
-0.007641213,
0.0308996364,
-0.0022075132,
-0.0094764102,
0.0366224386,
0.010159621,
-0.0152313961,
-0.0060698823,
0.0143847009,
0.0672486871,
0.0127660939,
0.0874652937,
0.000002075,
-0.0525097698,
0.0106991669,
-0.0010915992,
-0.0247009695,
0.0274031982,
0.0429926068,
-0.0350332409,
-0.0011148313,
-0.0097923819,
-0.0058780769,
-0.0396660864,
0.0745938718,
-0.0016959039,
0.0072876471,
0.0540605374,
-0.0208691992,
-0.0719599277,
0.0450501442,
-0.0686602071,
-0.025872834,
-0.0402173959,
-0.0005973675,
-0.0187588762,
-0.0240318961,
-0.0119626261,
-0.0237612817,
0.0231221542,
0.0347554311,
0.0168066788,
-0.0241711549,
0.0093218284,
-0.0023406493,
-0.0771781057,
0.0402598567,
0.0476036221,
-0.0349744633,
-0.0476071201,
-0.0079818191,
0.0002243328,
-0.0457714349,
-0.0372830518,
-0.0354189314,
0.0471080206,
0.0213482138,
0.0029647709,
0.0430431403,
0.0049786209,
0.0436019003,
0.018384276,
0.0007548252,
0.0285255481,
0.0488749221,
-0.0024368681,
0.032074701,
0.0425720401,
0.0015219188,
-0.0198448431,
0.0147854118,
-0.0060047656,
-0.0242740102,
-0.0098847887,
-0.0157472398,
-0.030830821,
0.0455921888,
-0.0558948554,
0.0429304056,
0.0448561534,
-0.0704967827,
0.0508588701,
-0.0209046062,
0.0556529835,
-0.0327973627,
0.0249819774,
-0.0400989838,
-0.0380185694,
0.0455097966,
0.0056461054,
-0.0243498869,
-0.0373993181,
-0.0370728672,
-0.01251732,
-0.0587299652,
-0.0085779242,
0.007744554,
-0.0384575948,
0.0370071679,
-0.0299337264,
-0.0582070462,
-0.0236585829,
0.0371980146,
0.0543265045,
-0.0277418029,
0.0271835197,
-0.0081519531,
-0.062947318,
-0.0136327157,
0.0418604687,
0.0093112299,
-0.035862226,
0.0418155454,
0.0603578277,
-0.0239803325,
-0.037789084,
-0.0310099255,
0.0258693285,
-0.0371167213,
-0.0891541243,
-0.0018223133,
-0.0676013082,
0.0355428681,
0.000043492,
0.0422998033,
-0.0489782616,
-0.0229576938,
-0.0182594564,
0.0059494991,
0.0358681902,
-0.0047884658,
-0.0061585507,
0.0228701178,
-0.0035900918,
0.037621215,
0.0777086988,
-0.0092939092,
0.0054221358,
0.0480525494,
-0.0153360218,
0.0033753405,
0.0527786911,
0.0262484662,
0.022728866,
-0.0106340973,
-0.0300447289,
0.0134386243,
0.006808924,
0.0555357523,
0.0187538993,
-0.0259437263,
-0.055684492,
0.0546537489,
0.0188758038,
0.0010404185,
0.0527188815,
0.0788513049,
-0.0279495437,
0.0157905631,
-0.0277361013,
0.0068342062,
0.0246324912,
-0.0179471187,
-0.0010899828,
0.0096120453,
0.0533533543,
-0.0500147417,
-0.0162030812,
-0.0076256115,
-0.0270731933,
-0.0038400076,
0.0254952069,
0.0662767589,
-0.0250859503,
-0.0021820795,
0.0355975479,
0.0859080106,
-0.0634136349,
0.0315325297,
-0.0416239724,
0.0462773181,
0.0450584106,
-0.0254612975,
-0.0121084247,
-0.0381118096,
0.0204434805,
0.0060384301,
0.074651435,
0.0750169605,
-0.0121532576,
0.0784391686,
-0.0791577995,
0.0283912942,
-0.0165246017,
0.0098303258,
0.0747890994,
0.0004272553,
0.0226910952,
0.0012347801,
-0.0184253044,
-0.0041275644,
0.0336083435,
-0.058723364,
0.0104814926,
-0.000573032,
0.0056133824,
0.0339777283,
0.0278458092,
-0.0304902885,
-0.0329252668,
0.0436400361,
-0.0433486588,
0.0078498563,
0.0087524801,
-0.0493554585,
-0.0239229128,
0.0360310785,
0.0344586484,
0.03642454,
0.001130534,
-0.0265327524,
-0.054871358,
0.0429061987,
0.0077569489,
0.0029489635,
-0.0705056787,
0.0229773112,
-0.0101904478,
0.0574678406,
0.0103440732,
0.0470003188,
0.0007332279,
0.0275294315,
0.0055247745,
-0.037206687,
0.0351372845,
-0.0553726591,
-0.0510864258,
0.0619212799,
0.0125806276,
0.0379530862,
0.030697519,
0.0305608045,
0.0033566279,
-0.0032264476,
-0.1001825258,
-0.0107444227,
-0.0589133427,
0.0044080862,
-0.0105392728,
0.0375741199,
-0.0273298305,
-0.0351325199,
-0.0183132309,
-0.028477652,
0.010812372,
0.038675569,
-0.0196586363,
0.0112012234,
0.0138723329,
-0.0190445594,
-0.0701838583,
0.0385475233,
0.0627133623,
-0.0176501591,
-0.0411126129,
-0.0285551958,
-0.0136455558,
-0.0309934281,
0.0087119723,
-0.015524691,
-0.0356096439,
-0.0327335335,
0.0025766918,
0.0246912669,
0.0447263233,
-0.0364929363,
0.0014499426,
0.0310842898,
0.0048120352,
-0.0218255688,
0.010835967,
-0.0000375867,
-0.0236653555,
-0.0480011702,
-0.0113984095,
-0.0325675867,
0.0903748572,
-0.0255122241,
-0.0393333323,
-0.0569512583,
-0.0400642119,
-0.0286751632,
-0.0353657007,
-0.0248944592,
0.0168863125,
0.0064906213,
0.0390885137,
-0.0339130498,
-0.0437323116,
0.0524720289,
-0.049636215,
-0.0214922149,
-0.0384581462,
-0.0425230451,
-0.0376101099,
0.0310110617,
0.0130005144,
0.0443986766,
-0.0367612988,
-0.0309799332,
-0.0518785305,
0.0130210994,
-0.0010655072,
-0.0051406114,
0.0524120145,
-0.046529565,
0.0382299349,
-0.0161999594,
-0.0500861108,
0.021318743,
-0.0108638024,
-0.0310378913,
0.0139149707,
-0.0044824444,
-0.0196365789,
-0.016182296,
-0.0006400304,
0.023854034,
0.0712964162,
0.0023220154,
0.0477062427,
0.0089477217,
0.0206640586,
0.0199931134,
-0.0179251134,
0.0027340886,
-0.05651832,
-0.008940421,
0.0259307548,
0.0184348188,
0.009071188,
-0.0109514631,
0.0289674383,
0.0610956363,
0.0023134148,
-0.0539250821,
-0.0191739053,
0.0435968116,
-0.0090424614,
-0.0085698804,
0.0712433755,
0.0061254096,
-0.0727695078,
-0.0008546241,
-0.0349995904,
0.0455208272,
0.0256105568,
0.0096696569,
-0.0857224911,
-0.0238748603,
-0.014335651,
-0.0274163838,
0.0317365453,
0.0146812033,
0.0129090445,
-0.1003780589,
-0.0053036446,
-0.0062390198,
-0.0532734096,
0.009352223,
-0.0076345229,
-0.024078967,
-0.0735900998,
0.0237601753,
0.004571877,
0.0073313829,
-0.0367214493,
0.0411900096,
0.0186665934,
-0.0562419668,
0.0599043332,
-0.0358433612,
0.0197650548,
0.0249757748,
0.0323621966,
-0.0474386029,
-0.0422613919,
0.0455248952,
-0.0114181712,
-0.0293586645,
-0.0219253115,
0.0184686445,
0.028781984,
-0.034228798,
0.0340209939,
0.0017006889,
-0.0353581645,
-0.0255323276,
-0.0344062932,
0.0206128694,
0.0418058671,
-0.0100875804,
0.0070950137,
-0.0220874567,
-0.0109436102,
0.0330336131,
-0.0085635027,
-0.0635290593,
-0.0147533491,
-0.0050311736,
0.0073486855,
-0.0308007114,
-0.0228756852,
0.0142648779,
-0.0239783302,
-0.0664998069,
-0.0556680076,
-0.0298761502,
-0.0037005495,
0.026436979,
-0.0097058378,
0.0029656475,
-0.0304652583,
0.0012012433,
-0.0066321013,
0.0309171677,
-0.0140622305,
0.0381551497,
-0.0697486326,
-0.0054338546,
0.0354109406,
-0.009736428,
-0.036515262,
0.0458148532,
-0.0853205621,
-0.0548797473,
0.0152792726,
0.0475739241,
0.0279106013,
0.0144451465,
-0.0199882034,
-0.0539423227,
0.0317047611,
-0.0315281115,
0.0107617099,
0.0375601016,
0.01432487,
-0.0305565316,
-0.0244032796,
0.0473715365,
-0.0368974097,
0.0332244001,
0.0193310473,
0.0585997365,
-0.0151592437,
-0.0232303236,
-0.0090769734,
-0.0153404931,
0.0144865736,
-0.0474130772,
-0.0260732826,
-0.0832729265,
0.0253387801,
-0.0101714376,
0.0007284289,
-0.03097618,
-0.0099220136,
0.0675391033,
0.0090688635,
-0.0103375567,
0.0213497356,
0.0293260254,
0.0227247607,
-0.0333427154,
0.0176926311,
0.0511390418,
0.0361646824,
0.0081445444,
0.0225135554,
-0.0187347699,
0.0250432417,
-0.0249651466,
0.0459181927,
-0.0485564284,
0.0204211231,
-0.0385244265,
-0.0102381185,
-0.0167096686,
0.0155460937,
-0.0131787909,
-0.0092675732,
-0.0479360633,
0.0568783581,
0.083761014,
-0.0038335177,
0.0819014162,
0.0106495935,
-0.0246018115,
-0.0316703022,
0.0186425541,
-0.0060929582,
-0.0387464836,
0.0147398319,
-0.014625296,
-0.0246232003,
0.0129176918,
0.0847804546,
0.0129523566,
0.0624063499,
0.0106475186,
0.0352947004,
-0.0145655312,
-0.0449504256,
0.0045037465,
-0.0165895894,
-0.0350424275,
0.0070242872,
-0.0188763924,
-0.0711788386,
0.026505148,
-0.0007873517,
0.0015729229,
0.0040593403,
0.002398873,
0.0818598196,
-0.0083647147,
-0.0298901275,
-0.0493613817,
-0.0424460247,
-0.0011752994,
-0.000821989,
0.0136750164,
-6.632521898e-33,
-0.0087058386,
0.0516579971,
0.0143187493,
-0.0658235103,
0.0055658445,
0.0011287479,
-0.0108984206,
0.0001616405,
-0.0441495888,
0.0025718273,
0.0021866322,
0.0137444688,
0.0264989696,
0.0478192717,
0.007997416,
0.0145125901,
0.0400439054,
0.0188799351,
0.0210574754,
0.0362040363,
0.0370860025,
0.023048168,
0.0975379646,
-0.003542434,
0.0025881419,
-0.0017632607,
-0.0269056428,
0.0139829172,
-0.0232818443,
0.0805101544,
-0.038272392,
0.0561455525,
0.0100016529,
-0.0118797971,
-0.0086434288,
-0.0262271278,
-0.0502961315,
-0.0048021059,
-0.0012031612,
0.0439007096,
-0.0114269536,
-0.0184555389,
-0.0196026955,
-0.0367828086,
0.0073837209,
0.002182245,
-0.0006811383,
-0.0244152118,
-0.0315456018,
-0.0543072894,
-0.0790180191,
0.0228351653,
0.0115211327,
0.0265442077,
0.0178778041,
0.0206350107,
-0.0015355777,
-0.0065723285,
-0.0597868226,
0.0055514914,
0.0011543728,
0.0373308659,
0.0479008593,
0.0325536914,
-0.003171338,
0.0159549285,
0.0451891534,
-0.0325778909,
0.0220449511,
0.0195550118,
0.0540766157,
0.0222238414,
0.0324144922,
0.0683263987,
0.0223539919,
-0.0308169983,
-0.1010328457,
-0.0125774471,
0.0316102766,
0.035660658,
0.026051674,
0.0148133663,
-0.0083559286,
-0.0229603332,
-0.0243060235,
-0.0724222511,
-0.0425098464,
-0.0179432798,
0.0056520551,
0.0125503615,
-0.0093688723,
0.0382761806,
0.0547600798,
-0.0204809718,
-0.0302992295,
0.0460011773,
-0.0077862972,
0.0215180367,
-0.0084109912,
0.0803220868,
-0.0610148162,
-0.0493114889,
-0.0121207358,
-0.0117923422,
-0.0123892706,
-0.0068107429,
0.0844620392,
0.0297104865,
-0.0648401976,
0.0018975566,
-0.0324859507,
0.0109453117,
0.038452208,
0.0718490481,
0.0251936223,
-0.0025227084,
-0.009416122,
0.0590484329,
0.0388186499,
0.017415246,
-0.0111651476,
-0.018872546,
0.0126371002,
0.043602284,
-0.0143913198,
0.0300831012,
-0.0178974587,
-0.0398159698,
0.0394061543,
-0.0316880979,
0.0088832034,
0.0350684486,
2.632e-7,
0.0288955811,
0.0230400693,
-0.0053961943,
-0.0280030016,
0.0030758576,
-0.0005346974,
-0.0292656589,
0.0024268494,
0.0041754292,
0.0180214327,
-0.0078542065,
0.0213802494,
0.0803688988,
0.0259676818,
-0.051307153,
0.0310029369,
-0.0330947638,
0.0113048395,
-0.0453202687,
-0.0155798895,
-0.029358888,
0.1470240355,
-0.0150942653,
-0.0162040014,
-0.0024493032,
-0.0432828441,
0.1044362187,
-0.0238872245,
-0.0210018698,
0.0176997483,
0.0118605243,
0.0178599376,
0.0152845187,
-0.0159694105,
-0.0294265468,
0.0469612889,
0.0544343032,
0.066310443,
-0.0131634343,
0.0033732809,
0.0182790384,
0.0038655507,
-0.0465039685,
-0.0501751266,
0.0419392399,
0.0302144364,
-0.0274234246,
-0.0409125052,
-0.0267310143,
0.0095498869,
0.0483442955,
-0.063611418,
-0.003312031,
0.030630948,
0.0404547751,
-0.0432296433,
-0.0482532382,
-0.0713095292,
-0.0394286551,
0.0124503253,
-0.0426239818,
-0.0145693887,
0.0219808146,
-0.0140149947,
0.0085789533,
-0.0392083079,
0.035183575,
2.083233588e-34,
0.0403065681,
0.0090499399,
0.0073480061,
0.0030737263,
0.0130112823,
-0.037370611,
0.0006321181,
-0.0046234597,
-0.0489854999,
-0.0615276359,
-0.0163930301
] |
http://arxiv.org/abs/2405.15028v1
|
http://arxiv.org/pdf/2405.15028v1
|
AGRaME: Any-Granularity Ranking with Multi-Vector Embeddings
| 2024-05-23T20:04:54 | 2024-05-23T20:04:54 |
Revanth Gangi Reddy, Omar Attia, Yunyao Li, Heng Ji, Saloni Potdar
|
cs.CL
|
cs.CL, cs.IR
|
Ranking is a fundamental and popular problem in search. However, existing
ranking algorithms usually restrict the granularity of ranking to full passages
or require a specific dense index for each desired level of granularity. Such
lack of flexibility in granularity negatively affects many applications that
can benefit from more granular ranking, such as sentence-level ranking for
open-domain question-answering, or proposition-level ranking for attribution.
In this work, we introduce the idea of any-granularity ranking, which leverages
multi-vector embeddings to rank at varying levels of granularity while
maintaining encoding at a single (coarser) level of granularity. We propose a
multi-granular contrastive loss for training multi-vector approaches, and
validate its utility with both sentences and propositions as ranking units.
Finally, we demonstrate the application of proposition-level ranking to
post-hoc citation addition in retrieval-augmented generation, surpassing the
performance of prompt-driven citation generation.
|
AGRaME: Any-Granularity Ranking with Multi-Vector Embeddings: Ranking is a fundamental and popular problem in search. However, existing
ranking algorithms usually restrict the granularity of ranking to full passages
or require a specific dense index for each desired level of granularity. Such
lack of flexibility in granularity negatively affects many applications that
can benefit from more granular ranking, such as sentence-level ranking for
open-domain question-answering, or proposition-level ranking for attribution.
In this work, we introduce the idea of any-granularity ranking, which leverages
multi-vector embeddings to rank at varying levels of granularity while
maintaining encoding at a single (coarser) level of granularity. We propose a
multi-granular contrastive loss for training multi-vector approaches, and
validate its utility with both sentences and propositions as ranking units.
Finally, we demonstrate the application of proposition-level ranking to
post-hoc citation addition in retrieval-augmented generation, surpassing the
performance of prompt-driven citation generation.
|
[
0.0280950461,
0.0240109842,
0.0136181712,
0.0147292418,
0.0381853618,
0.0011517367,
0.0001442434,
-0.0102363182,
0.0104714492,
-0.0588104576,
0.024921963,
-0.010185631,
-0.0154494736,
-0.0225238521,
0.072088331,
-0.0361025892,
0.0352230482,
0.0319187604,
0.0281764828,
-0.0069093085,
0.0008572922,
0.0156556983,
0.0179887861,
0.0296760574,
0.0515434854,
-0.0080089467,
-0.0236650761,
-0.0000407412,
0.0884452611,
-0.0362100787,
0.0150853526,
0.0425606668,
0.0357413553,
-0.0145679805,
0.0000017048,
-0.0099826492,
-0.0138139091,
0.0053171753,
0.0140763633,
0.0215949211,
-0.0125574963,
0.0226041321,
0.0521112867,
0.010095288,
-0.0385764241,
0.0092570428,
0.0303710327,
0.0260530841,
0.07795313,
0.0284577999,
0.0056473822,
-0.0248734299,
0.0153756551,
-0.0470874049,
0.0053092465,
-0.0180691723,
-0.0051739174,
-0.0149404304,
0.0143457735,
0.1054845676,
0.0067348685,
0.0205431394,
0.0054137944,
-0.0146085741,
0.0666618124,
0.0443317443,
-0.0041288906,
0.0553740785,
0.022708429,
0.1079976186,
-0.02230216,
0.0129563995,
0.015767036,
0.0085264752,
0.0257932488,
-0.0322088748,
-0.0348187201,
0.0188597608,
0.0157232434,
0.012483093,
-0.0001388515,
0.0264005717,
0.0190064292,
0.0082902396,
0.0119688185,
-0.0212414078,
0.0171947572,
-0.0493817218,
-0.0001780476,
0.0145463794,
-0.0312865153,
0.0079454435,
-0.0078183273,
0.0139795421,
-0.0362223424,
-0.016889656,
0.0130270841,
-0.0214654934,
0.0493317991,
-0.0222420543,
0.048577141,
0.0141945481,
-0.0032873226,
0.035779763,
-0.0413143672,
0.0213927198,
-0.027344631,
0.0006444118,
-0.0277570821,
-0.0114934212,
-0.0338123739,
-0.0264232103,
-0.0282382984,
0.0649871901,
-0.0551310107,
0.005944191,
-0.057987731,
0.025574591,
-0.036876522,
-0.01514252,
-0.0035973361,
0.0513319708,
-0.0354312994,
0.0305772461,
-0.0180681944,
0.0165421888,
-0.0689467415,
0.0186145026,
0.021073522,
0.0066628447,
-0.0361986607,
0.0066755824,
0.0126264682,
-0.025665855,
0.0056131617,
0.0244953502,
-0.0345274247,
-0.0295584276,
-0.0333166197,
-0.0379346833,
-0.0062289652,
-0.0540410727,
-0.0092337253,
-0.0716373548,
0.0393107161,
0.0232875049,
-0.0203274824,
-0.0118948892,
0.0085945316,
0.0439392217,
-0.0310324263,
0.028292466,
0.0366363525,
-0.0570411794,
-0.00883066,
0.0354606956,
0.0451964401,
0.0370512493,
-0.0123030115,
0.0342996083,
0.0115776286,
0.0379548445,
0.002660708,
-0.0364386477,
0.0461854376,
0.0237709265,
0.0237070955,
0.0019479529,
-0.0017080928,
0.0576573983,
0.0677738264,
0.0554531291,
-0.0476934612,
-0.0019553627,
0.0407411046,
-0.0170671996,
0.0156884026,
0.0170870665,
0.0694471449,
0.0090107797,
-0.0309953261,
0.0251070168,
-0.0039118091,
0.0521643721,
-0.0184736829,
0.000814762,
0.1020803303,
0.0255804602,
-0.0282619037,
-0.0133609818,
0.0343835428,
0.013779209,
-0.0072456924,
-0.031382177,
-0.0128467269,
0.0261066239,
-0.0551330447,
0.0417927392,
0.0129979486,
-0.0633829311,
0.0195000824,
0.0049107689,
0.0154123055,
0.0177665614,
0.055349607,
-0.032401897,
0.0128986798,
-0.0272531584,
0.0145100849,
0.0223109573,
0.0056549925,
-0.0140091479,
0.032231424,
-0.0637796596,
-0.0252889562,
-0.0608064122,
-0.0000126055,
0.0268743206,
-0.0000894304,
0.0209067091,
0.0352209322,
-0.0149691096,
-0.0097001931,
0.0033602591,
-0.0069142818,
-0.0292383712,
0.0375151932,
-0.0060005202,
-0.0029186765,
0.0251905844,
-0.0138207143,
0.0256565902,
0.0519242287,
-0.057907436,
-0.0535835363,
-0.0366267115,
-0.0115158223,
-0.047598768,
-0.0003507801,
-0.0470034592,
0.0075762174,
0.0061905859,
-0.0782430917,
-0.0149455834,
0.0042704381,
-0.0473416559,
-0.0072500464,
-0.0172971673,
-0.0323993713,
0.0372354686,
0.0154991597,
-0.0634267032,
0.101737909,
-0.0353249684,
0.0711045861,
0.075866513,
0.0031526547,
0.0158668812,
-0.0441246778,
-0.0646554232,
0.0014674403,
-0.0102019198,
0.0288903154,
0.0354190804,
-0.0058343164,
-0.0390673801,
-0.0079150703,
-0.0720985681,
0.0339241028,
-0.0006099343,
0.0021982198,
-0.0121707404,
0.0214619059,
0.0391599014,
-0.0267231278,
-0.0574875437,
-0.0401061028,
-0.0195540115,
0.0433678627,
0.0581723973,
0.0299856067,
0.0465760529,
-0.0254911631,
-0.0347873308,
0.0059244931,
0.066560477,
-0.0338196456,
0.0628706738,
-0.0574567057,
-0.002292797,
-0.0634359643,
-0.0060037682,
-0.0157829188,
-0.0076657073,
-0.0010422806,
-0.0250878707,
0.0435102656,
0.0581175052,
0.0336747169,
0.0241758954,
0.0387799516,
0.0040442292,
-0.0542881414,
0.0142021133,
-0.0341801979,
-0.0004894164,
0.0035511651,
-0.0087084221,
-0.0419557579,
0.0274803471,
0.0106810033,
-0.0219251737,
-0.0320542008,
0.0013983283,
0.0189115815,
-0.0003107594,
-0.0765095875,
0.027100198,
0.005753038,
0.0146060428,
-0.0193862971,
-0.0145838428,
0.0653757229,
0.0088687614,
-0.0033150834,
-0.0462098569,
-0.0034745315,
-0.0530274361,
0.0305102281,
-0.0023206105,
0.015482231,
-0.0124915028,
0.0049733664,
-0.0762014985,
0.0173060242,
0.0232959129,
-0.0440897308,
0.043641232,
-0.0030688827,
-0.0167102646,
-0.0414120033,
-0.0729114935,
-0.0284170806,
-0.0476981215,
0.0283841658,
0.0407583416,
0.0433434881,
-0.0211697631,
-0.0360731967,
-0.0429285467,
0.0441973768,
0.0218622219,
-0.0289236605,
-0.0376427509,
-0.0320388861,
0.0986644328,
0.0573834628,
0.04012255,
0.0185314603,
-0.0002370423,
-0.0203079041,
0.06525065,
0.0208456386,
-0.0454373397,
-0.0195962284,
-0.0141133508,
0.021809537,
-0.0294959154,
-0.0347259119,
-0.0088510569,
0.0193630438,
-0.0231100786,
-0.0046808845,
0.0645360872,
0.0321632847,
-0.032338772,
-0.0097572599,
-0.0210003182,
0.0619430579,
0.0077352729,
0.040918339,
-0.1175471246,
-0.0130084567,
-0.0374446511,
0.0344940871,
-0.0079255998,
-0.0153419813,
-0.0043519763,
-0.1225044951,
0.0107789254,
-0.0052498425,
0.026448926,
-0.0247432552,
-0.0515775755,
-0.0166414902,
-0.0320490226,
-0.0392804369,
0.0115549909,
0.002800154,
0.0132131297,
0.0194572043,
0.0525792018,
-0.0353794247,
0.0100142686,
-0.0521392524,
0.0105039366,
-0.0150253456,
0.0747669041,
0.0066816667,
-0.0477613024,
-0.0530574918,
-0.0672403648,
-0.0112201953,
0.0005605407,
0.0385694094,
-0.0108902799,
0.0179325175,
-0.0199951474,
0.0115277488,
0.0048711938,
-0.0022936105,
-0.009610422,
0.0611633174,
0.007116898,
0.0158006344,
-0.0602936223,
-0.0028833807,
-0.0068677692,
0.0530315042,
0.0104240244,
-0.0187973138,
-0.0210926663,
-0.0563749559,
0.0498291813,
-0.0438498966,
-0.0454507135,
-0.0301150214,
0.008677721,
0.0306237936,
-0.0767769516,
-0.0511350259,
-0.0280830748,
0.0299980268,
-0.0180270467,
0.0064993477,
-0.0558941662,
-0.0165017527,
-0.0680244938,
0.0053394702,
0.0049649579,
0.0149688954,
-0.0106115285,
0.009274492,
0.0504754335,
-0.0271044616,
-0.0514154322,
0.008851273,
0.0167463105,
-0.0475660525,
-0.0240216553,
0.0619317964,
0.002795323,
0.0388459414,
-0.0165257938,
-0.0462808385,
0.1226740927,
0.0611990988,
0.0061382586,
0.0212340839,
0.0498577915,
-0.0202997737,
-0.0214663278,
0.0205888543,
-0.0627706572,
0.0027807083,
-0.0528988466,
0.0330449305,
-0.0069439858,
-0.0291718934,
-0.0012573506,
-0.0389015786,
0.0443732925,
-0.0062004467,
0.014836207,
-0.0274221748,
0.0276433881,
-0.0083067389,
-0.0334132016,
-0.0268796198,
0.018954698,
-0.0010737642,
0.0158126429,
-0.0219861679,
0.0381862186,
0.0211342089,
0.0571451001,
-0.0511711314,
-0.0100120772,
-0.0064630951,
0.0087106097,
-0.0268154908,
0.009173586,
-0.0429603644,
-0.0325811766,
-0.0068066036,
0.0359156206,
-0.0350316241,
0.0212488454,
0.0365447626,
-0.0115897367,
0.0047078156,
0.0632678121,
-0.0982399434,
0.047330156,
-0.037539769,
-0.0638171285,
0.0509579256,
-0.0001747436,
0.0607720353,
0.0152075365,
-0.0341666751,
-0.0008809803,
0.0229884516,
-0.0102101052,
-0.0319921002,
0.011175734,
0.030056797,
-0.0090410197,
0.0465377197,
0.0756097436,
-0.0082463399,
0.0274321549,
-0.0409046412,
-0.0247060154,
-0.0228189789,
-0.0299013928,
-0.0441892371,
0.0796202943,
0.0179513376,
-0.0095559228,
-0.0474160276,
-0.0262095928,
0.0080096591,
0.0050585023,
-0.0113992337,
-0.0256766696,
0.050377205,
0.0066111628,
0.0067330813,
-0.0195370559,
-0.061138209,
-0.0014245309,
-0.0442208797,
-0.0360005312,
-0.0273709223,
-6.20390727e-33,
-0.0293537453,
-0.0370843895,
-0.0128418412,
-0.006266633,
-0.027574962,
0.0421010219,
-0.0220310837,
-0.0073942496,
-0.0324094892,
0.0220418032,
-0.0055491813,
-0.0019334264,
-0.016660165,
0.0340052173,
0.0199245289,
0.0117887584,
0.0525826365,
0.0164051112,
-0.0225086119,
0.0565677695,
0.0659292936,
0.0111854281,
0.0292193238,
-0.0491616949,
0.0458567329,
0.0024292239,
-0.0292091276,
-0.0005083383,
0.0137751512,
0.0686592311,
-0.0114251608,
0.0317492746,
0.0313596502,
-0.0597818643,
0.0102278637,
0.0240743142,
-0.0880569667,
-0.0259390194,
-0.0154507039,
-0.0784891322,
-0.0867233053,
-0.0536789522,
-0.0097521255,
-0.0213447139,
0.0470876358,
-0.0544957593,
0.0507966243,
-0.0167796277,
-0.0325066596,
-0.0652530342,
-0.0069872593,
0.0153141953,
-0.0188441221,
0.0046616383,
-0.0314159766,
0.0607511401,
0.0102700787,
0.0770862252,
-0.0381286144,
0.0128554897,
-0.045587983,
0.0389054939,
0.0049656783,
-0.0277273227,
-0.0485265553,
0.0461749583,
-0.011586044,
-0.0396534689,
-0.0219986532,
-0.0458593667,
-0.0450150296,
0.0214771871,
0.0628351644,
0.0759712309,
0.031893909,
0.0401209593,
-0.0676946342,
0.043274235,
-0.0079800077,
0.030939946,
0.02604785,
0.0199943483,
0.0074794288,
-0.0094412426,
-0.0415922813,
-0.0255666077,
-0.0289130807,
-0.024738783,
0.0035638511,
0.0389096402,
-0.0415592343,
0.049084492,
0.0346534997,
-0.0243637059,
-0.0138873253,
0.0845758542,
-0.0347642601,
0.0242549684,
0.0329927914,
-0.0155157214,
-0.0481962115,
0.0051574656,
0.0364979766,
-0.0184404533,
-0.0023224368,
0.0147278281,
0.0520173647,
0.0028529356,
-0.0501276702,
-0.0002198064,
-0.002854018,
0.0035287053,
-0.0664246529,
-0.0029457128,
-0.0106636519,
0.0149617326,
-0.0031123753,
0.0611839071,
0.0326557085,
0.0600268245,
-0.0124056051,
-0.0112712113,
0.01989186,
0.0204141159,
-0.0050319643,
0.0114153409,
-0.0156665053,
0.0479917414,
-0.0044194506,
-0.0596590452,
-0.0264873449,
0.0090604499,
2.516e-7,
0.0272001307,
0.0044938331,
-0.0189831313,
-0.0156421363,
-0.0006118761,
0.0331884027,
0.0516154319,
0.0859244391,
-0.0058516064,
0.0344421864,
-0.0037398173,
0.0010624082,
0.0102558425,
0.0008545892,
-0.0902606845,
0.0131098581,
-0.0716274381,
-0.0308246333,
-0.0031179152,
-0.0452034958,
0.0507835038,
0.0697929561,
-0.0051345648,
0.0261473879,
0.0033141295,
-0.0051357029,
0.0906932577,
0.0489331856,
0.000572919,
-0.0161752906,
0.0658938587,
0.0293397252,
-0.013092503,
-0.0010402176,
-0.0489118323,
0.0406002849,
0.041373685,
0.0343505107,
0.0268939305,
-0.0656743497,
0.0229632072,
-0.0332325846,
0.0081548849,
-0.0282560792,
0.0488269888,
-0.0470180847,
-0.0147737386,
0.0076398416,
-0.0561928079,
0.0315589644,
0.0474935696,
-0.0188368168,
-0.0573806688,
-0.0026327579,
-0.0174733493,
0.0056014955,
-0.009538101,
-0.0744262189,
-0.0252474919,
0.0264836419,
-0.009347178,
-0.0401935428,
-0.0028529814,
-0.0244327821,
0.03761274,
-0.0672484487,
0.0115206512,
2.386103611e-34,
0.0474340543,
-0.0234377943,
0.0083068395,
-0.0077021355,
0.024102753,
0.0129425125,
-0.0166311599,
-0.0101014888,
0.0502941497,
-0.0583278611,
-0.0150726102
] |
http://arxiv.org/abs/2405.15025v1
|
http://arxiv.org/pdf/2405.15025v1
|
OAC: Output-adaptive Calibration for Accurate Post-training Quantization
| 2024-05-23T20:01:17 | 2024-05-23T20:01:17 |
Ali Edalati, Alireza Ghaffari, Masoud Asgharian, Lu Hou, Boxing Chen, Vahid Partovi Nia
|
cs.LG
|
cs.LG, cs.CL
|
Deployment of Large Language Models (LLMs) has major computational costs, due
to their rapidly expanding size. Compression of LLMs reduces the memory
footprint, latency, and energy required for their inference. Post-training
Quantization (PTQ) techniques have been developed to compress LLMs while
avoiding expensive re-training. Most PTQ approaches formulate the quantization
error based on a layer-wise $\ell_2$ loss, ignoring the model output. Then,
each layer is calibrated using its layer-wise Hessian to update the weights
towards minimizing the $\ell_2$ quantization error. The Hessian is also used
for detecting the most salient weights to quantization. Such PTQ approaches are
prone to accuracy drop in low-precision quantization. We propose
Output-adaptive Calibration (OAC) to incorporate the model output in the
calibration process. We formulate the quantization error based on the
distortion of the output cross-entropy loss. OAC approximates the
output-adaptive Hessian for each layer under reasonable assumptions to reduce
the computational complexity. The output-adaptive Hessians are used to update
the weight matrices and detect the salient weights towards maintaining the
model output. Our proposed method outperforms the state-of-the-art baselines
such as SpQR and BiLLM, especially, at extreme low-precision (2-bit and binary)
quantization.
|
OAC: Output-adaptive Calibration for Accurate Post-training Quantization: Deployment of Large Language Models (LLMs) has major computational costs, due
to their rapidly expanding size. Compression of LLMs reduces the memory
footprint, latency, and energy required for their inference. Post-training
Quantization (PTQ) techniques have been developed to compress LLMs while
avoiding expensive re-training. Most PTQ approaches formulate the quantization
error based on a layer-wise $\ell_2$ loss, ignoring the model output. Then,
each layer is calibrated using its layer-wise Hessian to update the weights
towards minimizing the $\ell_2$ quantization error. The Hessian is also used
for detecting the most salient weights to quantization. Such PTQ approaches are
prone to accuracy drop in low-precision quantization. We propose
Output-adaptive Calibration (OAC) to incorporate the model output in the
calibration process. We formulate the quantization error based on the
distortion of the output cross-entropy loss. OAC approximates the
output-adaptive Hessian for each layer under reasonable assumptions to reduce
the computational complexity. The output-adaptive Hessians are used to update
the weight matrices and detect the salient weights towards maintaining the
model output. Our proposed method outperforms the state-of-the-art baselines
such as SpQR and BiLLM, especially, at extreme low-precision (2-bit and binary)
quantization.
|
[
-0.0322748907,
0.0584433489,
0.0203393716,
0.04330872,
0.0101203676,
-0.0010129134,
-0.0083429934,
0.0275879651,
-0.0358606316,
-0.0304094739,
-0.0214497596,
-0.0319176316,
0.0274780765,
-0.0323589332,
0.0250534657,
-0.0177412722,
0.023487309,
-0.015285302,
-0.0092460532,
-0.0004893564,
0.0091010658,
0.0206068717,
0.0256207269,
0.0235135872,
-0.0029798329,
-0.0099267829,
-0.0123218158,
0.0454207808,
0.0215313807,
-0.0115490854,
0.0103413714,
0.0498812608,
0.0409876816,
0.0345189683,
0.0000022295,
-0.0222869217,
-0.0288792755,
0.0078131473,
-0.0506932288,
0.0610082112,
0.03665879,
-0.0131536787,
-0.0480254441,
0.0425812788,
0.0052580987,
-0.0063558156,
0.0790994391,
-0.0191391241,
0.0452099629,
0.0499130338,
0.0051894486,
-0.0368575975,
0.0400210358,
-0.0256535374,
-0.0331175812,
-0.0210036822,
0.0028748994,
0.0257223304,
-0.0679941028,
-0.0324186906,
-0.0140717691,
0.0268028919,
0.01572931,
-0.008470742,
-0.0238466989,
-0.0031293156,
-0.0421421155,
0.0153753171,
-0.0279560536,
0.0305722356,
-0.0726083219,
-0.0468361713,
-0.0174586549,
-0.0134968972,
-0.0198418126,
-0.1083665267,
-0.0144979516,
0.024438072,
0.0322524235,
0.0434070639,
-0.0164168868,
-0.022412261,
0.0503086783,
-0.008602323,
-0.0428066961,
0.036098361,
0.0066425442,
-0.0283707157,
-0.0372234732,
-0.0083346255,
-0.0586313158,
-0.0350191742,
0.006165118,
-0.0008611543,
0.0085844938,
-0.0003945185,
-0.0815343931,
-0.0407376811,
-0.0414481275,
-0.0453798473,
0.0625328496,
0.0545456111,
-0.0193443783,
0.0080113141,
-0.0391351245,
0.0736515597,
-0.0152203627,
0.0354456082,
-0.0476904437,
0.025496481,
-0.0136469156,
-0.0000207291,
-0.0501353219,
0.0234651659,
0.0126796169,
-0.0085527664,
-0.0519424491,
-0.0049765212,
-0.0172065552,
-0.0019946625,
0.0305666104,
0.0159980971,
-0.0401937105,
-0.0252492409,
0.0365761444,
0.0457693674,
0.0303355083,
0.0144615974,
-0.0075119296,
-0.039397005,
0.0088483794,
0.0327763334,
-0.003357827,
-0.003215244,
0.0283776335,
-0.0096740415,
0.025876984,
-0.0331870243,
-0.0182775334,
0.0074032564,
-0.0041936208,
-0.0597097762,
-0.0295322742,
-0.0413882546,
-0.0132701686,
-0.0137739722,
0.0144229708,
-0.0410823859,
-0.0038672646,
-0.0272551961,
-0.0134988362,
0.045941595,
-0.0434929878,
-0.005912784,
0.0185357723,
-0.003731048,
0.0509501249,
0.1161265895,
0.0723009109,
0.0280448124,
0.004314689,
-0.0008275379,
-0.034853518,
0.0665170327,
0.0335682295,
0.0086052511,
-0.0926574767,
-0.0351845324,
-0.0134288315,
0.0450126082,
0.1038903072,
0.0725325197,
-0.0317461714,
-0.0050331177,
0.0391664878,
0.0211935956,
0.0638634935,
-0.0057263998,
-0.0119987912,
0.0033610961,
0.0124757057,
0.0114431074,
-0.0242768005,
0.0112394569,
0.0069428175,
0.0109097557,
0.0723865256,
0.0354545861,
-0.033964999,
-0.0313655324,
-0.0166881233,
-0.0129569601,
-0.0129310172,
0.0118409544,
0.0321138576,
-0.0026045551,
-0.0138681699,
0.1368291527,
0.0204219203,
-0.0618245006,
0.0133604836,
-0.0738422871,
-0.0300211441,
0.0418943204,
0.0364215374,
0.010256107,
0.022761561,
0.0098319668,
-0.0057849274,
0.0103345281,
-0.0022145722,
0.0084626675,
0.0216197241,
-0.066009894,
-0.0331421308,
0.0178316627,
-0.0006016225,
0.0173899662,
0.0072671217,
0.0099417102,
0.0206911527,
-0.0328201614,
0.0011300234,
0.0235842802,
0.0234801769,
-0.0270590968,
0.0333263986,
-0.0062617119,
0.0230829213,
0.0439662561,
-0.0351939686,
-0.0004282072,
0.0350154527,
-0.0334469527,
0.0439507067,
-0.0494815148,
0.0086793629,
-0.0147471717,
0.0698655546,
0.0252267588,
0.0400487892,
0.0554955527,
-0.060400784,
-0.0490875952,
0.0365758501,
-0.0226996448,
0.0392710008,
-0.071621269,
0.014000589,
-0.038111899,
0.0287420582,
-0.025684474,
-0.0120583456,
-0.0302756187,
0.0005570322,
-0.0195409507,
-0.0041734371,
-0.0100518502,
-0.0603655279,
-0.057262294,
-0.0036273259,
0.0070759393,
0.0635569245,
0.0169591568,
-0.0010523942,
0.0284296256,
0.011232744,
-0.0963641033,
0.001659872,
-0.0579863526,
0.03933727,
0.0355012976,
0.0272701439,
0.0088573731,
-0.0292996336,
-0.0401898287,
-0.0002868165,
-0.0219726991,
-0.0277927946,
-0.0113790492,
-0.0198648926,
-0.0349003524,
-0.0308155268,
-0.0395885743,
0.0043278551,
0.0409488343,
-0.0573134795,
-0.0215125345,
-0.1145828515,
0.0101518203,
-0.0109288683,
0.0441517085,
0.0191558469,
-0.0461828113,
-0.0001009668,
-0.0099342261,
-0.0213274658,
0.0790443271,
0.0055900817,
-0.0214080196,
0.0811949447,
0.0501346029,
-0.0262475647,
0.0162197482,
-0.0058033913,
-0.0678982362,
-0.0263943914,
-0.0397767834,
-0.0322776437,
0.0786921531,
-0.0215703771,
0.0177137088,
-0.0344486982,
-0.0334790237,
-0.0097125368,
-0.0205189697,
-0.0460434817,
-0.0206451844,
-0.0246771108,
0.0402170345,
0.0105006201,
-0.0502401441,
0.0097444281,
-0.0608369969,
0.0015508219,
-0.00418757,
-0.0069351629,
-0.006763333,
0.0274320599,
0.033557035,
0.0507231243,
-0.0003484908,
-0.0070455596,
-0.0724254325,
-0.0451526754,
0.0033792623,
-0.0392825082,
-0.0354460254,
-0.0129617928,
0.0178579036,
-0.0275287181,
-0.0158697702,
0.0391485319,
0.0320337713,
-0.0261359345,
0.0246847365,
0.0164941344,
-0.0400788486,
0.0361606739,
0.0716174543,
-0.0156766474,
0.0358384177,
-0.0052576913,
0.0120758638,
-0.013749674,
0.043957334,
0.0392217822,
0.0188973341,
0.0152657405,
-0.0208196323,
-0.0370248482,
0.0206306595,
0.033910729,
0.0923718512,
-0.0655366406,
-0.0123289358,
0.0595417321,
0.0308696367,
-0.0092050126,
0.0342418551,
0.0526021458,
-0.0373376235,
0.007452081,
0.0827243552,
0.016173711,
-0.0629441515,
-0.011257139,
0.0139371799,
0.0254153907,
-0.0024583649,
0.0047942186,
-0.1025172099,
-0.0242391173,
0.0157416463,
0.0363413654,
0.0125818541,
0.0227136519,
-0.0013111257,
-0.0748274997,
0.010697077,
0.0349782109,
-0.0073148329,
-0.0108454181,
-0.0234350003,
-0.0161068738,
-0.0525613949,
0.0649376512,
0.0331103802,
0.0199278165,
-0.0103732543,
0.0662326962,
-0.0279255696,
-0.0458952151,
0.0010977531,
-0.0223511662,
0.0491405427,
0.0918381065,
0.0114056058,
0.0005860121,
0.0168011859,
0.0273157544,
-0.0154940188,
-0.0361533165,
0.0279055573,
0.0007944653,
0.0137688518,
0.0000824037,
0.0328476839,
-0.0040283278,
0.009007751,
0.0198837258,
-0.0252974816,
0.0168881416,
0.015001243,
0.0064280154,
0.0083164629,
-0.0196975041,
-0.0361331701,
0.0289560352,
0.0310291424,
-0.0222429112,
-0.0435579829,
-0.0278576557,
0.0075970446,
-0.0512140244,
0.0165790468,
0.0139004793,
-0.0327535309,
-0.0729651004,
-0.0164857134,
0.0243787356,
-0.0107667837,
0.007452115,
-0.0159561336,
-0.0225712545,
0.0222853199,
0.0216572843,
-0.0908081606,
-0.0190239511,
-0.0201253593,
0.0301265456,
-0.042920623,
0.0207632799,
-0.0096033169,
-0.0019517058,
0.0016898694,
-0.012655735,
-0.1437946409,
-0.0661230609,
0.0217525866,
0.0314591974,
0.0151503067,
0.0463223793,
-0.0139104314,
-0.0850116536,
0.0685878024,
0.061735481,
-0.0070639942,
0.0612840168,
0.0242598038,
-0.0152648417,
-0.02095473,
0.0443400703,
-0.0290758815,
0.020459285,
0.0011081076,
0.033558663,
-0.0448488034,
-0.0833788961,
-0.0057066809,
-0.002682281,
0.0624051541,
-0.0316616036,
-0.0380712524,
0.0230887216,
0.0421683751,
0.0002598555,
0.0063556125,
-0.0372384377,
0.0045884689,
0.0292763673,
0.0034178693,
0.0093702078,
0.0059735575,
0.0309324786,
-0.0115307849,
0.0034180437,
0.0013655225,
-0.0436194949,
0.0507373214,
0.0601817071,
-0.0089349467,
-0.0182676129,
-0.0012726225,
0.0129887406,
0.029694749,
-0.0224025007,
0.0086125508,
-0.0101702269,
0.0269103367,
-0.0355073661,
0.0529483929,
-0.0279304963,
-0.0135771045,
-0.0911889449,
0.032524962,
0.0663845912,
0.0577686466,
0.0597111769,
0.0114739267,
-0.0262495261,
-0.0363614373,
0.0283774361,
-0.0108386362,
-0.0023614087,
0.0428676233,
-0.0060787397,
-0.0328641012,
0.0237768721,
0.013022,
0.0380968973,
0.009063472,
-0.0288666692,
-0.0031093748,
0.0098907677,
-0.0105096987,
-0.0510753989,
0.0096085379,
-0.0220702291,
-0.0090340907,
0.0121491216,
-0.0466727987,
0.0001371888,
0.047008235,
0.0233432185,
-0.0276279524,
0.0197996777,
0.0347051993,
-0.0136495167,
-0.0339666344,
-0.0312064737,
-0.0468925461,
-0.0029735533,
-0.0146907326,
0.0042815809,
-6.482179844e-33,
-0.0379999615,
0.0272973832,
-0.0103537822,
-0.0169269927,
-0.0216285717,
0.0353732631,
-0.0018258624,
-0.0017630699,
-0.0220095348,
-0.0332506113,
-0.0002552195,
0.0161164198,
0.0256619453,
-0.0018480928,
-0.0301238615,
0.0259008389,
0.0855878368,
-0.035475526,
-0.0248105135,
0.0248788577,
0.0141283879,
0.0231479201,
0.0729373693,
0.0183510091,
-0.0265424829,
0.0258922838,
-0.0459172167,
-0.0029525768,
0.0374228917,
0.0296126716,
-0.0211535227,
0.0229389574,
-0.0224255677,
-0.002158799,
-0.011707522,
-0.0278521981,
-0.0572205707,
0.021990804,
0.0039361855,
0.0629042313,
0.0086980639,
0.0039171246,
0.0295121484,
0.022546595,
-0.0810460225,
-0.0449953154,
-0.0138768703,
-0.0119448658,
-0.0398849882,
-0.0681382045,
0.0109414533,
0.0166160166,
0.0129989358,
-0.0014314498,
0.0634628162,
-0.0090668108,
-0.0298075322,
-0.0166557096,
-0.0924353376,
-0.0121004488,
0.0400278568,
0.0946101919,
0.0722811818,
0.0659208745,
-0.0163656361,
0.0039257775,
0.0611193739,
-0.0210209545,
-0.0099147102,
0.049625732,
0.0555417053,
-0.021528095,
0.0406639911,
0.0341793112,
0.1042674407,
-0.0086578298,
-0.053458415,
-0.0052747801,
-0.0042140475,
0.0287643671,
0.0453961045,
0.0694570616,
-0.0106874723,
-0.0377856828,
-0.0481743664,
-0.0364814438,
-0.0289233364,
-0.0437784754,
-0.0112941945,
0.0081070103,
-0.0277642105,
-0.018711552,
0.0192722995,
-0.0500865802,
0.0056755585,
-0.021871781,
-0.0214950908,
0.0277435295,
-0.0054588201,
0.0532414876,
-0.0259994864,
-0.0338722765,
0.0169104133,
0.0006791531,
0.017057566,
-0.0165170263,
0.028795233,
0.0499472208,
-0.0548099019,
-0.0016815905,
-0.029729303,
-0.0187623631,
0.0077309813,
0.0439002328,
0.0140027171,
-0.0052244784,
0.0110708158,
0.0262245089,
-0.0103355674,
0.0136513952,
-0.0086382888,
-0.0061187618,
0.0245568566,
0.0584342554,
-0.0548386723,
0.0040469691,
-0.0243794937,
-0.0065067825,
0.0308080856,
-0.0580485985,
-0.0541637503,
0.0183709133,
2.873e-7,
0.0418741405,
0.0073880758,
0.0223931707,
-0.0140006766,
-0.0138304289,
0.0021397015,
-0.0066835135,
0.0386912785,
0.018668646,
-0.0247471184,
-0.0023743929,
0.0152869495,
0.045450259,
0.0146756042,
-0.0616634972,
-0.0085027535,
-0.0601980872,
0.0279558431,
-0.0185972899,
0.0093072476,
0.0076681571,
0.1167188659,
-0.0009648203,
-0.0256300699,
0.0478398241,
-0.0253370069,
0.015862681,
0.0083685657,
0.0204139762,
0.0135017727,
0.0500420146,
0.0134801678,
-0.0268009305,
-0.0342508517,
0.001204418,
-0.0032640656,
-0.0387567841,
0.0565576702,
-0.0276727229,
-0.0131627116,
0.0026620498,
0.0226679686,
-0.0198195167,
-0.054892879,
0.0274997242,
-0.0026051446,
-0.0069458112,
-0.1133245155,
0.0105001163,
-0.0110258404,
0.0212042537,
-0.0182911903,
0.0244932305,
0.0160368644,
0.0338949151,
0.005002995,
-0.018175941,
-0.0767075941,
-0.0571132526,
0.0725866407,
0.0138041768,
-0.0561433844,
-0.0276754126,
-0.0086993538,
0.01946529,
-0.045675043,
0.0327033959,
2.520567996e-34,
0.0239907783,
-0.0284752212,
0.0265628677,
0.0225433093,
0.003028308,
-0.0238036234,
-0.0221420471,
0.0072189257,
-0.0199275967,
-0.0533250831,
-0.0223147329
] |
http://arxiv.org/abs/2405.15012v1
|
http://arxiv.org/pdf/2405.15012v1
|
Extracting Prompts by Inverting LLM Outputs
| 2024-05-23T19:35:03 | 2024-05-23T19:35:03 |
Collin Zhang, John X. Morris, Vitaly Shmatikov
|
cs.CL
|
cs.CL, cs.LG
|
We consider the problem of language model inversion: given outputs of a
language model, we seek to extract the prompt that generated these outputs. We
develop a new black-box method, output2prompt, that learns to extract prompts
without access to the model's logits and without adversarial or jailbreaking
queries. In contrast to previous work, output2prompt only needs outputs of
normal user queries. To improve memory efficiency, output2prompt employs a new
sparse encoding techique. We measure the efficacy of output2prompt on a variety
of user and system prompts and demonstrate zero-shot transferability across
different LLMs.
|
Extracting Prompts by Inverting LLM Outputs: We consider the problem of language model inversion: given outputs of a
language model, we seek to extract the prompt that generated these outputs. We
develop a new black-box method, output2prompt, that learns to extract prompts
without access to the model's logits and without adversarial or jailbreaking
queries. In contrast to previous work, output2prompt only needs outputs of
normal user queries. To improve memory efficiency, output2prompt employs a new
sparse encoding techique. We measure the efficacy of output2prompt on a variety
of user and system prompts and demonstrate zero-shot transferability across
different LLMs.
|
[
0.0111287246,
0.0388786085,
-0.0153684495,
0.0908594951,
-0.0241642408,
-0.0027756975,
-0.0205145665,
0.0538557172,
0.0343581922,
-0.0078887409,
-0.0363259837,
-0.0193376038,
-0.0225957483,
0.0239525978,
0.0551512986,
-0.0054107425,
0.0411514975,
-0.0446840301,
-0.0606383793,
0.0027534072,
-0.0258615427,
-0.0078616161,
0.0362713449,
0.0214471724,
0.0244815089,
-0.0120214224,
-0.0398477204,
0.0502944887,
0.0104125654,
-0.0051213657,
-0.0491725542,
0.0091365632,
-0.0157117285,
0.0364440903,
0.000002017,
-0.0367052071,
-0.0244471617,
0.0006433535,
-0.1065403521,
0.0124726128,
0.0011112016,
-0.0009336727,
-0.0364158452,
0.0165001489,
-0.0223814286,
0.033046294,
0.0503519624,
0.0077957138,
0.0633880347,
0.0713053197,
0.013489157,
0.0438178517,
0.0037692252,
-0.0225010682,
0.0386907645,
-0.0269004963,
-0.011445649,
-0.080799602,
0.0103327082,
0.0435159802,
0.0174884982,
0.0168084372,
0.0223078486,
-0.0236260649,
-0.0610616021,
0.0080879703,
0.0186710879,
-0.0062089618,
-0.0396314114,
0.0505824164,
-0.0181985125,
0.0066614463,
0.008665937,
0.0096738841,
0.0388033241,
-0.0216532461,
-0.047859963,
0.0399099514,
-0.0063609653,
0.011288004,
0.0039190068,
0.0024300334,
0.0219583977,
-0.0365496092,
-0.0364591591,
0.0951104537,
0.0343629979,
-0.0087117208,
0.0104263518,
-0.0123011554,
0.0486178845,
-0.0419587903,
-0.0160237774,
0.043603465,
-0.0542021319,
0.0190649014,
-0.0312716402,
0.0026025802,
0.0099763582,
-0.0121203102,
0.0699870586,
0.0417782255,
-0.0486456528,
0.0382350981,
-0.0418045186,
0.0407274589,
-0.0160965547,
0.0767236874,
-0.0282493476,
-0.009200749,
0.0053114733,
-0.0161551535,
-0.0300855115,
0.056693323,
-0.0214419831,
0.0247193743,
-0.0045100441,
0.0477158464,
0.0014977382,
0.0320281796,
-0.0269696936,
0.0040723062,
-0.0367529392,
-0.0167329088,
0.0078258747,
0.0038106625,
-0.0039692419,
-0.0294954032,
0.0228393637,
-0.092169039,
-0.0432449728,
0.0307476614,
-0.0252462793,
-0.0256650317,
-0.0078091901,
0.1167026982,
-0.0156712867,
0.0003804433,
-0.0685175434,
-0.0477362759,
0.0124023603,
-0.037327081,
0.0064916406,
-0.0827653632,
0.000603778,
-0.0053228647,
0.0024959748,
-0.0684607476,
0.0017141578,
0.0119037377,
-0.0028173341,
0.0758622661,
0.0023370688,
-0.0389236547,
0.0139901992,
0.0225553792,
-0.0001311468,
0.0512196086,
-0.020816505,
0.0920974016,
0.0281555951,
-0.0339302607,
0.0343770795,
0.0714018047,
0.0331056379,
0.0178661291,
-0.058828637,
-0.0064470498,
0.0401286967,
-0.0222454555,
0.0426578894,
0.0477322675,
0.0080225375,
-0.0332947187,
0.0327251665,
0.0775354952,
-0.0195366312,
-0.0368996784,
0.0115305856,
-0.0383092724,
-0.0326304771,
0.0483152419,
-0.0384365395,
0.0256161243,
-0.0657511428,
-0.0243756268,
0.0142380623,
-0.0112039158,
-0.0171945542,
-0.0182462651,
0.0259993412,
-0.0218868963,
-0.0320491865,
0.002767571,
0.0327441059,
0.0095628947,
-0.0383178666,
0.0675462857,
0.0350316167,
-0.0368172005,
-0.029639706,
-0.0771435127,
0.0071300524,
0.0340379328,
0.0158428848,
-0.0138173904,
0.0210152753,
-0.0094883936,
-0.0178670324,
0.0452910103,
0.020511752,
-0.0010121801,
0.0323014073,
-0.0062076012,
-0.0145732891,
0.004154732,
0.0129443221,
0.0815677866,
-0.026143156,
-0.0201924294,
0.0250718649,
-0.0033176574,
0.0024185521,
0.0082846424,
-0.0547922887,
-0.0800166056,
0.0335634872,
0.0231665727,
0.0040931427,
0.0039054903,
0.0204273388,
0.0212346725,
0.071740292,
-0.0509786345,
-0.0003401575,
-0.0253120251,
-0.0618014149,
-0.042083092,
0.033565186,
-0.0054944344,
0.0087991646,
0.0405505225,
-0.0429310799,
-0.0378158912,
0.0090090651,
-0.0322314166,
0.0584314615,
-0.023962358,
0.0197076052,
0.0297096912,
-0.0201940406,
-0.0663924739,
0.0271777902,
-0.0262005627,
0.005383499,
0.0365722217,
-0.0018224115,
-0.0009442475,
-0.0653090402,
-0.0468796529,
0.0517507307,
-0.0406890027,
0.0074372813,
0.0084959101,
0.0096325558,
-0.0103437919,
0.001389705,
-0.0532676391,
0.0119812051,
-0.0227921214,
0.0307914242,
0.0379381888,
0.0380683877,
0.0047774352,
-0.0489429422,
-0.0540893823,
0.0117425546,
0.0233508833,
0.0233471077,
0.0338348374,
0.0341885686,
-0.0039613605,
-0.0434661135,
-0.0116438856,
0.0031793257,
0.033505071,
-0.062105529,
-0.0314267613,
-0.0602802522,
0.0789290443,
0.0431515686,
0.0239793509,
-0.024076974,
-0.0225372221,
0.02409303,
-0.0378311649,
-0.0070934147,
0.0518509969,
0.0065590302,
0.0454278067,
0.0291593485,
0.0140496623,
-0.0743260831,
-0.0105273062,
0.007795081,
-0.070293732,
-0.0156384688,
-0.0607984588,
0.0492727049,
0.0581938773,
0.0209905542,
0.0332494639,
-0.0165120307,
0.0050558886,
0.0272796527,
-0.0310883429,
-0.0572002977,
0.0265673865,
-0.0568499416,
0.0619317107,
-0.0115325181,
-0.0670524165,
0.0516592674,
-0.063199088,
-0.0004756,
0.041292347,
-0.0524746738,
0.0153965019,
0.0654249191,
-0.024364097,
0.0232473481,
0.0235225707,
-0.068338871,
-0.0443862602,
-0.0157416016,
-0.0073341513,
-0.0199660826,
0.0485109463,
-0.0154197114,
-0.0089859646,
-0.0370282046,
-0.0418555699,
-0.0501698181,
-0.0554800928,
0.0035281787,
-0.0070693856,
0.003718805,
-0.0195644759,
-0.0099353613,
-0.0115121305,
0.0261627883,
0.0314453132,
-0.0088826483,
-0.0100016724,
-0.026729729,
0.0044600754,
0.0460669324,
0.0122319572,
-0.0275058821,
-0.0172556136,
0.0347932577,
0.0706522837,
0.0241889302,
0.0106593668,
0.0185362305,
-0.023877738,
0.0519189388,
0.0031595058,
-0.0203624275,
0.0394501239,
0.0275393501,
-0.0387177616,
-0.021768149,
0.0802993402,
0.0158568751,
-0.0565356463,
0.0018342998,
-0.0336440019,
0.0242722277,
0.0286047291,
0.0394516103,
-0.0899819061,
-0.0398028865,
0.014355164,
0.059892159,
-0.003129611,
-0.0412375443,
-0.0184592493,
-0.162839964,
-0.0007317511,
0.0207187347,
-0.0105667887,
-0.0179910623,
-0.0179041438,
-0.0006633537,
0.0188804436,
0.0172393676,
0.010519404,
-0.0116834678,
-0.0158606768,
0.0274092779,
0.0304431953,
0.008485659,
0.0093335183,
0.0076290607,
0.0505434982,
0.0944745317,
-0.0171576571,
-0.0127350045,
0.0008194027,
0.0354005769,
-0.0114801433,
0.0146446712,
-0.0066911122,
0.039479997,
0.0124621168,
-0.0231271498,
0.0257273447,
0.0204713587,
-0.0268121976,
-0.0007252739,
0.0409196615,
0.0545772761,
-0.0437315628,
-0.0014966114,
0.0127972038,
-0.0322828591,
0.0279340092,
0.0819400549,
0.0443887413,
-0.01539387,
-0.0218772981,
-0.0887965709,
-0.0168429669,
-0.0143890567,
-0.0455955714,
-0.0031597246,
0.003093513,
-0.0120990416,
-0.0409507006,
-0.0300130267,
-0.0006849999,
0.0701949,
0.0045314003,
-0.0022165501,
-0.0220669415,
-0.0395841151,
-0.0099885473,
-0.0137054734,
-0.0609347485,
0.0086506475,
-0.0953001007,
-0.0047104801,
0.0093950704,
-0.0304806419,
-0.0063423775,
0.0086477464,
-0.0443611555,
-0.0690258294,
-0.001709624,
0.0665865242,
0.0264640804,
0.0380545631,
-0.0118868193,
-0.0164127108,
0.0510795116,
0.0768473074,
-0.0149714714,
0.0199454576,
0.0474576913,
0.0150566231,
-0.011258618,
0.0330318026,
-0.0513181798,
-0.0040594009,
0.0766077563,
-0.0134282792,
-0.0505641662,
-0.0575607978,
0.0247032549,
0.0591473579,
0.0424272716,
-0.0122829955,
-0.0371394269,
0.0200100411,
0.0377331115,
-0.0285920985,
-0.0106530935,
-0.0401009284,
-0.0036498681,
0.0032743781,
-0.0375301428,
0.0489297025,
-0.0459361188,
-0.0177012999,
0.0561488383,
-0.0631770715,
0.0006246328,
0.0302375816,
0.0229579285,
0.050654117,
0.0260086879,
-0.0143191256,
0.0634621456,
0.0420586914,
0.0734615549,
-0.0222569741,
0.0473206416,
0.0245234054,
0.0154972421,
0.0041530617,
0.0601160154,
-0.070709087,
-0.0431315787,
0.0003714511,
0.0478246175,
0.0306258015,
-0.0114185316,
-0.0023650979,
-0.0001331358,
0.015312667,
-0.0403165445,
0.0096163005,
0.0266745593,
-0.0223611854,
0.0371154249,
0.0044103218,
-0.007800906,
-0.005356831,
0.0666545928,
-0.0061566187,
0.0473327637,
0.0017787718,
0.002025451,
-0.058864668,
0.0016328064,
0.0593926609,
-0.0188789759,
0.0024533097,
-0.0210180506,
-0.0322257914,
-0.0490215309,
-0.0202435907,
-0.0000061018,
0.0005015226,
-0.0070026899,
0.0229599103,
0.0500784442,
0.0209928099,
0.0157516114,
-0.0392474271,
0.0301560517,
0.0152871208,
-0.05905056,
-0.0210859682,
-6.224932456e-33,
-0.0844905153,
-0.0550575592,
-0.0219888631,
0.0096228328,
0.0305774026,
-0.015700493,
-0.0459382311,
0.0010572152,
-0.0256372821,
-0.0259755142,
-0.0393797718,
-0.0296647362,
0.0314567499,
0.0085723465,
-0.0124997078,
-0.0249272976,
0.0339384899,
0.0084107993,
-0.0441340432,
0.025465779,
0.0753190592,
0.0420753732,
0.0449403897,
-0.0457605943,
0.0313728824,
-0.013081504,
0.0103813848,
0.0341123901,
0.0055590072,
0.0352475867,
-0.0005187344,
0.0334270447,
0.0083956802,
-0.0526595153,
-0.0195437055,
0.0386283398,
-0.0171749722,
-0.0346379839,
0.0455803536,
-0.0163887423,
0.0206654277,
-0.0675730109,
-0.0054123709,
-0.0294551235,
0.0068766857,
0.0114788497,
-0.0069768182,
0.0081200469,
-0.0241902582,
0.001606277,
-0.0411701947,
-0.0070736888,
0.0127133904,
-0.0266956743,
0.0162439458,
0.0226006564,
0.0167565476,
-0.0223205984,
-0.0298584905,
0.0089062992,
0.02852365,
0.0551897138,
0.0894653723,
0.0592252314,
-0.0394508541,
0.0351232663,
0.0072450219,
-0.0051754504,
-0.0269037243,
-0.0064974469,
0.0049398984,
0.0409642048,
0.0255293157,
0.0145181557,
0.0193216484,
-0.0289876051,
-0.0241446663,
0.0520015433,
0.0278905239,
-0.0056904815,
0.0163718779,
0.0219749082,
-0.0141575765,
-0.0364366025,
-0.0091295317,
-0.0100577893,
-0.0139893303,
-0.0362715609,
0.0384376608,
-0.0289756935,
-0.0157533977,
-0.0698496401,
0.0588673539,
-0.0357837714,
0.0018201445,
0.0094034635,
-0.012546448,
-0.0141303912,
-0.0120275719,
0.0185609646,
-0.0059250323,
-0.0012965635,
0.0138549767,
-0.0015725447,
0.030637892,
0.0272529013,
0.0179296583,
0.0330923013,
-0.0622825362,
-0.0046509225,
-0.0652075112,
0.0362259112,
0.0115528684,
0.0274517741,
0.0149321044,
-0.045159556,
0.0083156088,
0.0297753513,
-0.0311005246,
-0.0490087792,
-0.0211027227,
-0.0003649879,
0.0043246727,
0.0035095939,
-0.0466477908,
0.0079432307,
-0.0135284038,
0.0467178263,
0.0523518398,
0.0070881373,
0.0051347348,
-0.0123680728,
2.765e-7,
0.0133214202,
-0.0318615697,
0.0263496581,
0.0326760747,
-0.0083370935,
0.061614994,
-0.0236677732,
0.0373115391,
0.0095126349,
-0.0622772537,
-0.0219686143,
0.000237403,
0.0412084199,
0.0104979975,
-0.0906000882,
-0.0196300521,
-0.0567870773,
-0.0198779833,
-0.0393396653,
0.0114867762,
0.0727130026,
0.081918478,
-0.0150813432,
-0.0299546998,
0.0329307988,
-0.0340400077,
-0.0018739478,
-0.0586884506,
-0.0012995672,
-0.0294903368,
0.0519228131,
-0.0101003367,
0.0149440803,
-0.0065036747,
-0.0118262265,
-0.0190259349,
-0.0206457358,
0.0655159578,
0.0014179085,
-0.0728050694,
-0.0199528504,
-0.0200717691,
0.0106445439,
-0.0266450569,
0.0387543961,
0.0398589298,
-0.0155656571,
-0.0604868121,
0.0233241655,
-0.0440050438,
-0.0238154531,
-0.0397305898,
0.0087618455,
0.0112085016,
0.0252103936,
-0.004965656,
-0.0198862124,
-0.0352869704,
-0.0454886183,
-0.0260849074,
-0.0314872786,
-0.0599622168,
-0.0494063087,
0.0267537981,
-0.0165230073,
-0.0543432049,
-0.0057831309,
2.820978205e-34,
-0.0161298905,
0.0276788007,
0.0238567926,
0.0557296649,
0.0273423903,
0.0191035755,
0.012611703,
0.0266879909,
0.0088985413,
-0.0707482174,
-0.0326808244
] |
http://arxiv.org/abs/2405.15007v1
|
http://arxiv.org/pdf/2405.15007v1
|
RE-Adapt: Reverse Engineered Adaptation of Large Language Models
| 2024-05-23T19:23:40 | 2024-05-23T19:23:40 |
William Fleshman, Benjamin Van Durme
|
cs.CL
|
cs.CL, cs.AI, cs.LG
|
We introduce RE-Adapt, an approach to fine-tuning large language models on
new domains without degrading any pre-existing instruction-tuning. We reverse
engineer an adapter which isolates what an instruction-tuned model has learned
beyond its corresponding pretrained base model. Importantly, this requires no
additional data or training. We can then fine-tune the base model on a new
domain and readapt it to instruction following with the reverse engineered
adapter. RE-Adapt and our low-rank variant LoRE-Adapt both outperform other
methods of fine-tuning, across multiple popular LLMs and datasets, even when
the models are used in conjunction with retrieval-augmented generation.
|
RE-Adapt: Reverse Engineered Adaptation of Large Language Models: We introduce RE-Adapt, an approach to fine-tuning large language models on
new domains without degrading any pre-existing instruction-tuning. We reverse
engineer an adapter which isolates what an instruction-tuned model has learned
beyond its corresponding pretrained base model. Importantly, this requires no
additional data or training. We can then fine-tune the base model on a new
domain and readapt it to instruction following with the reverse engineered
adapter. RE-Adapt and our low-rank variant LoRE-Adapt both outperform other
methods of fine-tuning, across multiple popular LLMs and datasets, even when
the models are used in conjunction with retrieval-augmented generation.
|
[
0.0170484725,
0.0636470094,
-0.0192513671,
0.054418806,
-0.010165073,
0.0366406068,
-0.0353744365,
-0.0173211601,
-0.0125258844,
-0.0638231039,
-0.0954095721,
0.0065333471,
-0.0398457795,
-0.0015211902,
0.0381598994,
-0.0030333926,
0.0621003769,
-0.0302207712,
-0.0622684173,
0.011382699,
-0.0264840201,
0.039153222,
0.0101034595,
0.021699449,
0.0551986247,
0.0115854992,
-0.0134452768,
0.0369428135,
0.0251486693,
0.0205407292,
0.0436278842,
0.0181748886,
0.041002553,
0.008015627,
0.0000019121,
-0.0267794225,
-0.0476090573,
-0.0269585289,
-0.0278074294,
0.0246892497,
0.0328118503,
-0.0004691394,
-0.022576049,
0.0345488712,
-0.0323079601,
0.0085783657,
0.0193647686,
-0.0021462215,
0.0747282282,
0.131807223,
0.016094014,
-0.0390276276,
0.0124185998,
-0.0408100672,
-0.017614048,
-0.0454925001,
0.0155814961,
-0.0096929837,
-0.003107474,
0.0370959118,
-0.0285874773,
0.0190275516,
-0.0054683355,
0.0153683042,
-0.0223934334,
-0.0172260553,
-0.0030642219,
0.0134872179,
-0.0102605093,
0.0433954969,
-0.0131633794,
-0.0289856326,
0.002985399,
0.0055262451,
0.0145269511,
-0.0114352405,
-0.0290546827,
0.0166678429,
0.0360390097,
0.0173788313,
0.0492165796,
-0.0759323612,
0.0412514359,
0.0028464906,
-0.0448564738,
0.0531859435,
0.0361943804,
-0.0117453365,
-0.0157838594,
0.0108767301,
0.0007230271,
0.0000888025,
-0.0091641974,
0.0308312513,
-0.0350036174,
0.0357154384,
-0.0730497986,
-0.0268795863,
-0.0153349945,
0.0235229693,
0.1094285399,
0.0516597293,
-0.029783912,
-0.0054695457,
-0.0568492338,
0.0363727994,
-0.0100548631,
0.0280035473,
-0.0662952811,
0.0320415869,
-0.0461909473,
0.0146932043,
-0.0449462868,
0.0011360594,
-0.0615030117,
0.0266518276,
-0.0238482486,
0.0538084432,
0.0151089896,
-0.0394076034,
0.0156098511,
-0.0205961615,
-0.0808649138,
-0.0218314324,
0.0179511439,
0.0426947661,
0.0012553242,
0.0137991058,
0.0250291452,
-0.0574079677,
-0.0127816666,
0.0035172142,
0.0023342471,
-0.0209750403,
0.0343974456,
0.068956472,
-0.0205447469,
-0.0387442,
-0.0769095719,
-0.0178560205,
-0.0125898756,
-0.0500836298,
0.0311428718,
-0.0398421772,
0.0131265856,
-0.0179153644,
0.01585689,
-0.062542513,
-0.0004732638,
-0.0066717616,
-0.0529281721,
0.04462808,
-0.0109103201,
0.0076878252,
-0.0080361171,
-0.0053913649,
0.0175396428,
0.0717611685,
0.0138801336,
-0.0015526811,
0.0164018739,
-0.0181973055,
-0.0244035702,
0.0599337704,
0.0252326652,
0.0218052287,
-0.0145292329,
-0.0137076331,
-0.005760842,
0.0012534736,
0.0553283021,
0.0044990974,
-0.0074929572,
0.009777084,
0.0196101014,
0.0043313419,
0.0470687635,
0.0198847353,
0.0686225072,
-0.0121059855,
-0.026834663,
0.0356382541,
-0.0739284232,
0.0136709735,
-0.0302118156,
0.0141572785,
-0.0173483994,
0.0522726774,
-0.0158295035,
-0.0402878672,
0.0044123661,
0.0225997195,
-0.0211928301,
-0.0062994971,
0.0360247642,
0.0448175445,
-0.04735085,
0.0342601538,
0.0406986624,
-0.0448182747,
0.0244330298,
-0.0741570741,
-0.0011075337,
0.0651813596,
0.059344586,
-0.0072974707,
-0.0492376275,
0.0202755909,
0.0474035963,
0.0173274111,
0.0398765691,
-0.0033525061,
-0.0063213198,
-0.0608333014,
-0.0314732827,
-0.019234851,
0.0514944158,
0.0384607278,
0.0126500083,
-0.0185947474,
0.0331371091,
-0.0350808091,
-0.0270785,
0.0433801301,
-0.082283482,
-0.050317958,
-0.0011262923,
0.0092581771,
0.0040596714,
-0.0009630402,
-0.0219056606,
0.0023882361,
0.0415411964,
-0.0161096547,
0.0318324864,
-0.0116162412,
-0.0256979335,
-0.0196257904,
0.0218214598,
-0.0120950621,
0.0052748267,
0.0360709652,
-0.0528060049,
-0.0438028872,
0.0256284941,
-0.0341467187,
0.0414596125,
-0.0348446406,
0.0514911003,
0.0034401831,
0.0078524584,
-0.0564369448,
0.0293179043,
0.009656297,
-0.0185496379,
0.055179175,
-0.0156654343,
-0.0210380722,
-0.0247260015,
0.0010974123,
-0.0210423935,
-0.0084294491,
0.0249403492,
-0.0036660058,
0.0438812226,
-0.0452392735,
-0.0476245657,
-0.042586688,
0.0372630842,
-0.048475299,
0.0573622286,
0.048563756,
0.0221503675,
-0.015938513,
-0.0159332883,
-0.0155176977,
0.0167964213,
0.0099559184,
0.0268447697,
0.0115922065,
-0.0015231573,
-0.0167918541,
-0.0535972305,
-0.0043612588,
0.0126507767,
-0.0033836665,
-0.0474235453,
0.0261298362,
-0.1054312587,
0.0420958363,
-0.0134957535,
0.0326138586,
-0.0021257021,
-0.0602178164,
-0.0054105027,
-0.0358662568,
0.0428156257,
0.0833841413,
-0.0156899076,
0.0346711352,
0.0595496185,
-0.0147948554,
-0.0757257342,
0.020792691,
-0.0199117642,
-0.0869912803,
-0.0542261787,
-0.0310041253,
0.0108705619,
0.0641501471,
-0.0225078575,
-0.0314260758,
-0.0570797585,
-0.0088904016,
-0.0003215994,
-0.0164827947,
-0.0105376709,
0.0151632959,
-0.0219507366,
0.0160210729,
-0.0141685912,
-0.0466410778,
0.0477152318,
-0.0121421888,
0.0094394078,
-0.0564691238,
-0.0162396338,
-0.0273676589,
0.01613646,
0.0143522397,
0.0226165392,
0.0090042502,
-0.0181588475,
-0.069636859,
0.012594481,
-0.0004670516,
-0.0196901076,
0.084032245,
-0.0528056361,
-0.0011629101,
-0.0082220752,
-0.0209896546,
-0.0445413142,
-0.0053300294,
-0.0062371134,
-0.0032639592,
0.0163254067,
-0.0440462567,
-0.0834209621,
-0.0404537655,
0.0200905949,
0.0424016193,
-0.0198118668,
-0.0120588569,
-0.0471416116,
0.0028682349,
0.0570933074,
0.0280580875,
-0.0181621294,
-0.0388370566,
-0.0039545633,
0.0239854548,
0.0307309236,
-0.0382261463,
-0.0282228552,
0.0340877362,
0.0604591481,
-0.0190123823,
-0.0089091696,
0.0400348827,
0.0663034171,
-0.0655999035,
-0.0575223491,
0.0947358608,
0.0112887667,
-0.0331223384,
0.0069287415,
-0.0061795996,
0.0575165786,
0.0252396744,
0.0612532906,
-0.0568544865,
-0.0474320836,
0.0274510849,
0.0611456297,
0.0190000162,
0.0050267573,
-0.0069669327,
-0.1125154272,
0.0383726656,
0.0568834804,
-0.03424941,
-0.0374541245,
-0.006294711,
0.0053495332,
0.0040199365,
0.0489544682,
0.0371023901,
-0.0388918035,
-0.0029055341,
0.015801182,
0.017505642,
-0.0606320687,
0.0015536814,
-0.0079443343,
0.0732375011,
0.0953859389,
0.0062444573,
0.0305663832,
-0.0157226361,
0.0509701073,
-0.0303851627,
0.0467371866,
-0.0176440012,
0.0345887803,
-0.0079048881,
-0.0205811448,
0.0376975834,
0.0363574736,
-0.0467023328,
0.0101223923,
0.013423725,
0.0502178781,
-0.0149605051,
0.0068221805,
-0.0199429374,
-0.0354305543,
-0.0160909723,
0.0410656519,
0.0393159166,
-0.0359511375,
0.0100111086,
-0.0404212698,
-0.0133662075,
-0.0740170851,
-0.0405420214,
0.0262973923,
-0.0678744465,
0.005909116,
-0.0658013001,
-0.013128045,
0.0103412215,
0.0204073135,
0.0325181298,
-0.0148929302,
-0.0048945975,
-0.0315141715,
-0.0585648045,
0.0418248475,
-0.0733775347,
-0.0028927263,
-0.0912984088,
-0.0454546399,
0.0523446649,
-0.0172689687,
0.0017410312,
-0.0034218566,
-0.0623845048,
-0.0730325207,
0.0074084331,
0.0682517663,
0.0148706585,
0.0682589412,
0.0014063373,
-0.0310772285,
-0.004945165,
0.0234234147,
-0.0200009272,
0.0611831844,
0.029946791,
-0.0457610823,
-0.0275755078,
0.0426177755,
-0.0604120642,
0.0265797544,
0.0592757501,
0.0452117808,
0.011729368,
-0.0413987339,
0.0157948956,
0.011460403,
0.0335617438,
-0.0165681206,
-0.0146333631,
-0.0232871044,
0.030257469,
-0.0148033397,
-0.0019552503,
-0.0494497754,
-0.0012703072,
0.03480836,
-0.0274321344,
-0.0044286055,
-0.0186685771,
-0.0118042417,
0.0541058071,
-0.0679305047,
0.0382223949,
-0.0020336704,
0.0209197961,
0.0374576636,
0.018966632,
-0.0060999487,
0.0243998189,
0.0257777013,
0.0635985434,
-0.0277995225,
0.0225385614,
0.0094735175,
0.0230976138,
-0.0320674181,
0.0505937897,
-0.0514647737,
-0.016083315,
-0.0482204519,
0.0170703884,
0.0572468601,
-0.009117607,
0.0781148896,
0.0051648617,
-0.0071955561,
-0.0232918542,
0.0179889984,
-0.003334552,
-0.0080219191,
0.010050253,
0.0154557014,
-0.0009733177,
0.0091427034,
0.04541694,
-0.0001834952,
0.0894865468,
-0.0146430107,
0.0342537463,
-0.0184958577,
-0.0415615477,
0.0195270982,
-0.0119656865,
-0.0022183219,
-0.0000890255,
0.006492666,
-0.0715335086,
-0.0391386077,
0.0161168594,
-0.0031975699,
-0.0264737103,
0.0176646542,
0.039596688,
-0.0279684402,
-0.0218844134,
-0.0375204943,
0.0106153125,
-0.0095594302,
-0.0457656011,
-0.0173220932,
-6.280400361e-33,
-0.0563589483,
0.0166414455,
0.0358615145,
-0.0284019336,
-0.0343372114,
-0.0249364339,
-0.0358941332,
-0.0110261738,
-0.0198584609,
-0.0079142684,
-0.0089966645,
-0.0001957431,
0.011524233,
0.0294106919,
0.0039471202,
0.0121844904,
0.0549749807,
-0.0109193446,
-0.0310792737,
0.0223625582,
0.0271073692,
0.0332198553,
0.0597650781,
-0.0347017236,
-0.0032021739,
0.0047049453,
-0.0149759399,
0.0341335721,
-0.0143854776,
0.0682992637,
0.0305101275,
0.011943493,
0.0175206419,
-0.0144420806,
-0.0410944186,
-0.005295367,
-0.079036653,
-0.0125483833,
0.0242964439,
0.0108154425,
-0.020041788,
-0.0626266673,
0.042197749,
-0.0101473927,
-0.0212664921,
-0.0206914991,
0.0031612273,
-0.0490281694,
0.0012342952,
0.0144459251,
-0.0674536899,
-0.0054810015,
0.0264590476,
-0.0470400825,
0.0223031864,
0.0363196023,
0.0046686837,
-0.0131683303,
-0.0236953292,
0.021127332,
0.0289311148,
0.1022860706,
0.0837286115,
0.0346711725,
-0.0484093763,
0.021175433,
0.0095482832,
-0.0102904029,
0.0428503007,
0.0012871944,
0.055138927,
0.0526724011,
-0.0117898956,
0.023483742,
-0.0020172321,
-0.0143920016,
-0.0369241014,
0.0228972659,
-0.0036039851,
0.036985904,
0.0060025244,
0.0316637754,
-0.0147587452,
0.0090463143,
-0.043574594,
0.0336260349,
-0.0305369664,
-0.020787945,
0.0500907786,
0.0117042791,
-0.0366775878,
-0.042127803,
0.0405869447,
-0.0508731753,
0.0196514521,
0.0049160933,
-0.0291872732,
-0.0155897811,
-0.0266693085,
0.0334598012,
-0.0083640814,
-0.0788104534,
0.0376289524,
-0.02291237,
-0.0025443081,
-0.0054574478,
0.0269937757,
0.0445361473,
-0.0618338659,
-0.0003575593,
-0.0406784229,
0.0246605296,
-0.0039255293,
0.0548871346,
0.0184013788,
0.0043537738,
0.0062211598,
0.054051403,
0.0209001042,
-0.0251205452,
-0.0178377535,
0.0015074166,
0.0346934609,
0.0385184772,
-0.0350532606,
0.0291877631,
-0.0191148501,
0.0225645136,
0.0033755065,
-0.0316984728,
-0.0106907962,
0.0118071362,
2.636e-7,
0.0480095521,
0.0378609933,
0.0080483016,
0.0191356409,
0.0113397883,
0.022610547,
-0.0306639336,
0.0245242156,
0.0634571016,
0.0002067017,
0.0232102443,
0.0338972397,
0.0789825022,
-0.0309712254,
-0.0667875931,
0.0134553919,
-0.0003107154,
0.0212372355,
-0.0316701829,
-0.0330584832,
0.0067153969,
0.095977515,
-0.0019220497,
-0.0235641189,
0.0036907594,
-0.0205384251,
0.0353139564,
-0.0314944834,
-0.0068883221,
0.0253445972,
-0.0064977058,
0.0556180812,
-0.0205477588,
-0.0195444822,
-0.0446900688,
0.0372113995,
-0.0588837452,
0.0756115541,
0.0012193202,
-0.0268074237,
0.0203098319,
-0.0003578079,
-0.0097796265,
-0.0018507693,
0.0301118847,
-0.0419688635,
0.0088759931,
-0.0282462854,
-0.0183467921,
-0.0078002382,
-0.0445796363,
-0.0380371772,
-0.0092026414,
0.0360728912,
0.0465930402,
0.0026581481,
-0.0543441065,
-0.0829810053,
-0.008538546,
0.0252729729,
-0.0376992598,
-0.0423849709,
-0.0371332727,
0.0033023446,
-0.0000341652,
-0.0816371664,
-0.0062515321,
2.490998574e-34,
0.0226715934,
-0.0208685659,
0.0223149266,
0.0487394668,
0.0260434765,
-0.0050052064,
-0.0094445888,
0.0382084809,
0.0054838476,
-0.0962424055,
-0.0319938287
] |
http://arxiv.org/abs/2405.14992v1
|
http://arxiv.org/pdf/2405.14992v1
|
Linking In-context Learning in Transformers to Human Episodic Memory
| 2024-05-23T18:51:47 | 2024-05-23T18:51:47 |
Li Ji-An, Corey Y. Zhou, Marcus K. Benna, Marcelo G. Mattar
|
cs.CL
|
cs.CL, cs.LG
|
Understanding the connections between artificial and biological intelligent
systems can reveal fundamental principles underlying general intelligence.
While many artificial intelligence (AI) models have a neuroscience counterpart,
such connections are largely missing in Transformer models and the
self-attention mechanism. Here, we examine the relationship between attention
heads and human episodic memory. We focus on the induction heads, which
contribute to the in-context learning capabilities of Transformer-based large
language models (LLMs). We demonstrate that induction heads are behaviorally,
functionally, and mechanistically similar to the contextual maintenance and
retrieval (CMR) model of human episodic memory. Our analyses of LLMs
pre-trained on extensive text data show that CMR-like heads often emerge in the
intermediate model layers and that their behavior qualitatively mirrors the
memory biases seen in humans. Our findings uncover a parallel between the
computational mechanisms of LLMs and human memory, offering valuable insights
into both research fields.
|
Linking In-context Learning in Transformers to Human Episodic Memory: Understanding the connections between artificial and biological intelligent
systems can reveal fundamental principles underlying general intelligence.
While many artificial intelligence (AI) models have a neuroscience counterpart,
such connections are largely missing in Transformer models and the
self-attention mechanism. Here, we examine the relationship between attention
heads and human episodic memory. We focus on the induction heads, which
contribute to the in-context learning capabilities of Transformer-based large
language models (LLMs). We demonstrate that induction heads are behaviorally,
functionally, and mechanistically similar to the contextual maintenance and
retrieval (CMR) model of human episodic memory. Our analyses of LLMs
pre-trained on extensive text data show that CMR-like heads often emerge in the
intermediate model layers and that their behavior qualitatively mirrors the
memory biases seen in humans. Our findings uncover a parallel between the
computational mechanisms of LLMs and human memory, offering valuable insights
into both research fields.
|
[
0.0020015985,
0.0002782671,
-0.0326418765,
-0.0035281884,
-0.0087050982,
0.0420135781,
0.0045917928,
-0.0187170208,
0.0161267389,
-0.0197381005,
-0.0400945432,
-0.0380128734,
-0.0058210082,
-0.0422392935,
0.02928048,
-0.0149195325,
0.0560522415,
-0.0296575949,
-0.0315629654,
-0.0016999829,
-0.0323083512,
-0.0490066409,
0.0132043017,
-0.0009387076,
-0.0087581538,
-0.0227810238,
-0.013553476,
-0.0251188315,
0.0110379476,
-0.010133313,
0.012451129,
0.0433885269,
0.011474885,
0.0309280325,
0.0000021219,
-0.0624689944,
-0.0071555418,
-0.0021728382,
-0.0424435809,
-0.0060499529,
0.0026922494,
-0.0052169138,
-0.0396191515,
-0.0081055304,
-0.0368496291,
0.0117655173,
0.039698191,
0.0662502646,
0.035895966,
0.050533317,
-0.0114146918,
-0.0439357199,
0.0505951419,
-0.0047587147,
0.0554164834,
0.0035265316,
0.0452840626,
-0.0350613184,
-0.0804023147,
-0.0197613928,
-0.0316741578,
0.0340121537,
0.0442264676,
0.0601975434,
0.0160306133,
0.0554721355,
-0.0107126674,
-0.0173231121,
-0.0139093306,
-0.0041020196,
0.0140556954,
-0.0541094318,
-0.0066070464,
0.0381038785,
0.0097678211,
0.0164041705,
-0.0631872863,
0.0519916266,
-0.0092972787,
0.0053517469,
0.034994524,
-0.0114636794,
0.0151319997,
-0.0214802101,
-0.0561654828,
0.0505234301,
0.0119637959,
-0.0712314621,
0.0126444893,
0.0020173234,
0.013369441,
-0.032472834,
0.0119212288,
0.0419322327,
0.0179483164,
0.0177090522,
0.0025198495,
-0.0276421402,
0.0305980388,
-0.0507374182,
0.0013445832,
0.0246409103,
-0.0614652336,
0.0325447097,
-0.0473629944,
0.0665849149,
-0.0413715392,
0.0392108001,
0.0047757337,
0.0106954845,
-0.0371611416,
-0.0258947648,
-0.0360611975,
0.0217298344,
0.0637473464,
0.0013679597,
-0.0356401354,
0.0709145144,
0.0125704473,
-0.0152272237,
0.0012022387,
-0.0111246733,
-0.028815411,
-0.0013734861,
-0.0137704508,
0.0274909064,
-0.0186139829,
-0.0248183161,
0.0132697001,
-0.0239694752,
-0.0423416384,
0.0158659332,
-0.0126788914,
-0.0293467566,
-0.005438006,
0.0481039025,
-0.0065998356,
-0.0606341995,
-0.0410447679,
0.0089883851,
-0.0047364077,
-0.0580390431,
0.023437731,
-0.0306133442,
-0.0069831014,
0.0541428924,
-0.0101134526,
-0.0497157387,
-0.0051502483,
0.0324078389,
-0.014649801,
0.0284798667,
0.0370885693,
0.0132902758,
0.0636035204,
0.0004149448,
-0.0037658145,
0.0225403588,
0.0221927743,
0.012576621,
-0.0209395345,
-0.0284764394,
0.0176864993,
0.0235303827,
0.0110855987,
-0.0220489372,
0.0014008374,
-0.0289143063,
0.0165405665,
0.0101942541,
0.0128647834,
0.0472338721,
-0.0193609372,
-0.0214042943,
0.0540192351,
0.0515097417,
0.0268252958,
0.0558500104,
0.002892548,
-0.0342471153,
0.0203110483,
0.0456965454,
-0.0992626175,
0.0645062253,
-0.0342773907,
0.0192553066,
-0.0311532021,
0.0160566922,
-0.0284174234,
-0.00424108,
0.0221556276,
0.018121751,
-0.0144741694,
0.0048773289,
0.0489818007,
0.0020968483,
-0.0530791506,
0.062208157,
0.0027364041,
-0.0457861535,
0.0056448006,
-0.043616917,
0.0152533809,
0.0430377163,
0.002031588,
-0.0333984531,
-0.0216222014,
0.0030621458,
0.0242365543,
0.0268966462,
0.015079462,
0.0195124224,
-0.0316853561,
-0.0288216006,
0.0108923772,
-0.0012402062,
-0.0052527734,
-0.0023483147,
0.0002720294,
-0.0212845076,
0.0100310966,
0.0075463904,
-0.0193884578,
0.0355444029,
-0.0916312411,
-0.0197402593,
-0.0625432357,
0.0331691951,
-0.0191682372,
-0.0280249026,
0.0048641716,
0.0056909737,
0.0732972249,
-0.0284797717,
-0.005379966,
0.0545289367,
-0.0833437443,
-0.0021779998,
0.0043422137,
-0.0163200591,
0.0037391891,
-0.0133642061,
0.0471935347,
-0.0526785813,
-0.0049504577,
-0.0412123464,
0.1170277148,
-0.0492512807,
0.0524282493,
-0.0168310553,
0.0230703466,
-0.0418666974,
0.083171919,
-0.0387059748,
-0.0411745645,
0.0375791825,
-0.0146433515,
0.0010019796,
-0.0449712276,
0.004450636,
-0.0053607812,
-0.012201393,
-0.0083272494,
0.0164459329,
-0.0615884513,
-0.0843565837,
-0.0407946706,
-0.0708879977,
-0.0048900787,
-0.024148548,
0.0160547476,
0.0113060288,
-0.015679758,
-0.0025724864,
0.0006411896,
-0.0116307987,
0.0108049745,
-0.0051513501,
0.0481893085,
-0.0044115628,
-0.0104056224,
-0.0013549974,
-0.0400921628,
-0.0091245854,
-0.0289116465,
0.0436789356,
-0.0073453411,
-0.0254387707,
-0.0839040801,
0.0374345109,
0.028892884,
0.0097946264,
-0.029178502,
-0.0235478431,
0.0296364725,
-0.007058118,
0.1059430912,
0.1324012578,
0.0411294661,
0.0241075438,
0.05252086,
-0.0252916273,
-0.0368062779,
0.0041452968,
-0.0475910343,
-0.0171846319,
-0.0325165913,
-0.00468934,
-0.0119921062,
0.0876262486,
0.0081789577,
0.0182410926,
-0.0463979915,
0.0127821816,
-0.0013311151,
-0.0353092737,
0.011289537,
-0.0020791593,
-0.0688567236,
0.0376930758,
-0.0072577498,
-0.0407642089,
0.0634617954,
0.0226045512,
0.0005079519,
0.0286283586,
0.0178965982,
-0.0210338254,
0.0273746494,
0.0497521013,
0.0309566725,
0.0256951395,
-0.0211663451,
-0.0656633899,
0.0458898135,
-0.0128390584,
-0.0187912639,
0.048524458,
0.0089011919,
-0.022055557,
0.0283548832,
-0.0142392199,
-0.0897713527,
0.0193697866,
-0.0323526859,
-0.0091268094,
0.003397845,
-0.0038216028,
-0.0700996146,
0.007180809,
0.0807606354,
0.0390499309,
0.0094901752,
-0.0443612784,
0.006952093,
0.0048605474,
0.0124961901,
0.029765131,
0.0525361001,
-0.0390616395,
0.0507320054,
0.0455655903,
0.0397771075,
0.0109162293,
0.0002311948,
0.0092155309,
0.0438139439,
-0.0704645887,
-0.0313321427,
0.0361978933,
0.0073953206,
-0.0527375005,
-0.0565428063,
0.111342378,
0.0290923268,
-0.0994782075,
-0.0051449863,
-0.0428076126,
0.003581563,
0.0514646173,
0.0272815302,
-0.0392575152,
-0.01092037,
0.0166594181,
0.0013616098,
-0.0093416693,
-0.0130011635,
0.0258802585,
-0.0679056123,
0.0161069054,
-0.0084339855,
-0.0785052776,
0.0087868813,
-0.0438247733,
-0.0023619202,
-0.0151741989,
0.0018985978,
0.0521529764,
-0.0345360264,
-0.0338235348,
0.0154227642,
-0.0038055154,
-0.017713882,
0.0252599213,
-0.0185569171,
0.0452259369,
0.071962975,
-0.0006761311,
-0.0036998645,
-0.0567799211,
0.0170702003,
-0.0456029028,
0.0394833796,
0.0127857272,
0.0439681634,
-0.0079700369,
0.0076152915,
-0.0437073782,
0.0265920702,
-0.0211698692,
-0.0494718589,
-0.0036081558,
0.0803862289,
-0.0262465142,
-0.0127405347,
0.0234525297,
0.0013975508,
-0.0336636789,
0.0788877234,
0.019920025,
-0.0531598367,
-0.0115523608,
-0.0106783723,
0.0651280954,
0.0116112391,
-0.0693058074,
0.0363333784,
0.0174706038,
-0.0275318138,
-0.0618047081,
-0.0183713045,
-0.0029839159,
-0.0100990878,
0.0310117844,
-0.0251358561,
0.0284534246,
-0.0246352684,
-0.0510935821,
-0.0004948335,
-0.0290714484,
-0.0412326381,
-0.1112804562,
-0.0337412879,
0.0359686688,
-0.0674314648,
0.0031585512,
0.0061934353,
-0.0527879409,
-0.027539691,
0.106387049,
0.0616430342,
0.0203168914,
0.0479491502,
0.0243079811,
-0.0178158451,
0.0455643125,
0.0145662241,
-0.0164000336,
0.0414492153,
0.0159693845,
0.0017687824,
-0.014503425,
0.0134554682,
-0.0364373401,
0.0244879033,
0.0236794986,
0.0188399777,
-0.0283867717,
-0.0179214645,
0.0710036457,
-0.0485247336,
0.073185727,
-0.0059712306,
0.0174222626,
-0.0008410027,
0.014038628,
0.014819447,
0.0157315806,
0.019206766,
-0.0353081636,
-0.0029277105,
0.0232154895,
0.0186535642,
0.0324332044,
0.0020359994,
-0.00843764,
-0.0100536775,
0.0066972151,
-0.0243683271,
-0.011418676,
0.0073103365,
0.0547096618,
0.0077789696,
0.0205074511,
0.0054945783,
0.051631622,
-0.0052761692,
0.1015900224,
0.0202606078,
-0.0083477413,
-0.0452980623,
0.0187249109,
-0.1045511663,
0.008165326,
0.0305178929,
-0.0362156853,
0.0607967526,
-0.0632979274,
0.0321845524,
-0.0269256793,
0.0072807604,
-0.0096710948,
-0.0281361323,
0.0171811245,
0.0165436417,
-0.0253137257,
0.0131067391,
0.0169833489,
0.0329520665,
0.0520517342,
-0.0265996996,
0.0499063991,
-0.0065979031,
-0.0039809714,
-0.0133839352,
-0.02147476,
-0.009496348,
-0.0073002554,
-0.023479281,
-0.0360923931,
-0.0756706223,
-0.0452919193,
0.022760639,
-0.0282387473,
0.0330140926,
-0.0491238981,
0.0230030399,
0.0409769751,
-0.0230733212,
-0.0398054644,
0.011888871,
0.0282603838,
-0.0152227003,
0.0007110424,
-0.0190355908,
-6.764678319e-33,
-0.021461565,
-0.0285061207,
0.0334948152,
0.0359189957,
-0.0269811787,
-0.0264609847,
-0.0380591713,
0.0056077791,
-0.0033794912,
-0.0183249637,
-0.0441232808,
0.0169268511,
0.0074675973,
0.0047912793,
0.0343123786,
-0.0083030127,
0.0305079166,
0.001487293,
-0.030837059,
0.0171583444,
0.0480948947,
0.0514207482,
0.0509545319,
-0.018286543,
0.0050782524,
-0.0167976283,
0.0568027124,
0.0131258797,
-0.0260719061,
0.051414229,
-0.0413774811,
0.0431654789,
0.0332444571,
-0.0958475769,
-0.0160983913,
0.0371758081,
-0.0978519544,
-0.002942201,
-0.0114774024,
-0.0187182352,
0.0005344598,
-0.0560391024,
0.064420782,
-0.0412877165,
-0.0288185608,
0.000223544,
0.0533405952,
-0.0115829315,
-0.0440449566,
-0.0157575682,
-0.0723628476,
-0.0258025639,
-0.0064761383,
-0.0081676506,
0.0160419829,
-0.0069034756,
-0.0118589215,
0.0377054401,
-0.0809609294,
0.0329903923,
-0.0336758904,
0.0137439454,
0.0238042064,
0.0491686203,
0.0170255695,
0.0483714081,
0.0685124919,
0.0663891137,
-0.0330974199,
-0.0432066321,
0.0424332097,
0.0074347737,
0.0125155868,
-0.0108029377,
0.0274052881,
-0.0177224874,
-0.0184976757,
-0.0075226035,
-0.0140806576,
0.0733300894,
0.0539998002,
-0.0103253126,
-0.0063178884,
-0.0179908779,
-0.0269927979,
0.0058344947,
-0.0346025974,
-0.0224852785,
0.0143839354,
0.006223876,
0.0073488392,
-0.0191644579,
0.0366797075,
-0.0271556433,
0.0552820861,
0.0268143136,
-0.0099560926,
0.0123920748,
-0.0172134619,
0.0205740388,
-0.0590303764,
0.0081432322,
-0.0062224665,
0.0452431291,
0.0194916483,
0.0240364596,
0.0358750299,
0.0286070388,
-0.0388590433,
-0.0238579251,
-0.0280007944,
-0.0000630628,
0.0082295286,
-0.0413001329,
0.017181458,
0.0315566398,
0.0044657965,
0.0203294214,
0.0221339948,
0.0013378924,
-0.0583805367,
0.0113751115,
0.0000115402,
0.0101233097,
-0.0104369903,
0.0112585938,
-0.04559296,
0.0806788951,
0.0222526211,
-0.0672914088,
0.0025685451,
-0.0114329075,
2.975e-7,
-0.0024217749,
0.0949503556,
0.0061616115,
-0.0192131568,
0.0474673472,
0.0017818973,
-0.010413669,
-0.008928596,
0.0432218425,
-0.0229366887,
0.0141141117,
0.0606530793,
0.0327307023,
-0.0055846809,
-0.0504005924,
-0.0366785415,
-0.0389978848,
0.0312783532,
-0.0172536317,
0.0204590447,
0.0150142247,
0.0480899625,
0.0143700084,
-0.0250399429,
-0.0068037473,
0.0153258508,
0.070608601,
-0.0478060171,
0.0356310233,
-0.0023445333,
-0.0031942006,
0.0533712655,
0.0035653838,
0.0435821898,
0.0049763829,
-0.0016587904,
-0.0196584146,
0.0243453961,
0.0161125325,
-0.0478458293,
0.030997511,
0.0237907786,
-0.0110817784,
-0.0458229482,
0.0425223559,
-0.1059497967,
-0.0396452285,
0.0159288924,
-0.0277726166,
-0.0156465955,
-0.0226109307,
0.0181508306,
-0.0039954693,
-0.0134677617,
-0.024878731,
0.0278820544,
0.0035754433,
-0.0842422694,
-0.0302478857,
0.0423675552,
-0.0382477045,
-0.047031831,
-0.0510781519,
-0.0552660562,
0.0146731865,
-0.0476375073,
-0.0003902673,
2.42121852e-34,
0.0097493967,
0.0122353407,
0.0250407383,
0.0314508937,
0.0047096917,
-0.0390762463,
0.0351596922,
0.0130073195,
0.017792603,
-0.1088998392,
-0.0111730061
] |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.