ambean/construct-validity-review · Datasets at Hugging Face

mundlerSWTBenchTestingValidating2024

SWT-Bench: Testing and Validating Real-World Bug-Fixes with Code Agents

Include

null

A benchmark for generating code tests (unit tests) from natural language user GitHub issues.

null

Specific Application (A single use case, where the benchmark is likely to be examples of that use case)

Automatic code test generation (i.e. generating unit tests for issues)

Yes

The ability to generate valid tests to reproduce an issue in a codebase.

Comprehensive

null

Given a GitHub issue in natural language, you must write tests to reproduces the described issue.

A GitHub issue (taken from SWE-Bench), code that contains the issue and code with a 'golden patch' that has the issue fixed. The goal is to write unit tests that fail on the faulty code but pass after the patch is added.

Very comprehensive details about task definition.

Real task examples (e.g. GitHub issues), Modified from another benchmark (e.g. translation into another language)

1900

Yes

Length of the GitHub issue in tokens, original GitHub repository

Specific criteria (items were taken from a larger set based on specified rules)

Structured response (e.g. valid JSON, API call alone)

Whether the faulty code fails on the test and the gold-standard code passes it.

null

SWE-bench, which originates from real GitHub issues

Academia

Yes

null

Test

null

Simple Mean

Yes

Description length in tokens, original GitHub repository

null

https://github.com/logic-star-ai/SWT-Bench

SWT-Bench

Widely-agreed

Yes

The benchmark is itself realistic

No

Yes

Limitations in how the phenomenon was operationalised - all problems are in Python.

simple mean

Outputs alone

Complete real task (e.g. providing medical advice to real people interactively)

null

Single cohesive phenomenon

Not applicable

null

Agents

Coding

null

['Real task', 'Another benchmark']

['Criterion']

['Structured']

['Reward']

['Widely-agreed']

['Yes']

['Realistic']

['Yes']

['Complete']

['Mean']

davidsonEvaluatingLanguageModel2024

EVALUATING LANGUAGE MODEL AGENCY THROUGH NEGOTIATIONS

Include

null

The paper introduces a dynamic framework for evaluating LLMs using negotiation games in self-play and cross-play settings. They find that only closed-source models are able to successfully complete the task and that stronger LLMs don't always win over weaker opponents.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Alignment

Yes

Alignment metrics of interest are internal and external faithfulness as defined in Section 2.3, and the ability to follow instructions. [...] We measure instruction-following behavior of staying within the maximum number of words allowed to generate notes/messages (note/msg instruct) and the ability to correctly format internal offer indications using valid JSON (format instruct). [... (from 2.3)...]. . In natural language processing (NLP), faithfulness is a concept used to describe how accurately a model’s reasoning explains its answers/actions. To measure internal faithfulness, agents are asked to summarize acceptable offers for each Issue in their mental notes. [...] If Alice makes an offer to Bob for fewer slices than she stated as acceptable, we register this as an instance of internal unfaithfulness.

Subset

The paper is a bit unfocused in what it measures. The title says "Agency", the authors mainly note "Alignment" as motivation, and there is also a degree of "Negotiation skill" and "Theory of Mind".

The task is a series of negotiation games, where LLMs are given rules, a persona, protocols, and goals. Agents do both internal deliberation and external negotiation, and the game ends when a completion criteria is reached.

A single task is one round of a negotiation game that is either self-play or against another model.

null

Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions)

null

Yes

prompts, game settings, issues

Targeted items (creators defined a task space and chose tasks within it strategically)

Extended interaction (e.g. conversation, calling an API and processing the response)

Exact Match (accuracy, F1, precision, recall), Number of rounds completted

null

The authors generate a list of Games, Issues. It seems these were crafted manually

Academia

Yes

null

This "benchmark" defines too many phenomena to fit neatly in the framework

Test

null

Negotiation

Simple Mean

Yes

Scores are reported for different types of games.

null

https://github.com/epfl-dlab/LAMEN/

null

Contested

Partially

Yes

No

No comparisons made

It is an entirely constructed scenario (no available realistic setting)

No

null

mean with variance

Outputs alone

Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

The tasks simulates agent negotiations (so no humans involved)

Composite phenomenon

Yes

null

Alignment

null

['Author-crafted']

['Targeted']

['Interaction']

['Exact match', 'Reward']

['Contested']

['Partially']

['Not possible']

['No']

['Constructed']

['Mean', 'Std']

helweMAFALDABenchmarkComprehensive2024

MAFALDA: A Benchmark and Comprehensive Study of Fallacy Detection and Classification

Include

null

The paper introduces MAFALD, a benchmark that provides a unified classification of fallacies and provides a taxonomy. It features manually annotated data with explanations, a tailored annotation scheme, and an evaluation method for subjective NLP tasks. Various language models and human performance are evaluated on fallacy detection and classification in a zero-shot learning setting.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

fallacies in reasoning

Yes

A fallacy is an erroneous or invalid way of reasoning. A fallacy is an argument where the premises do not entail the conclusion. Sub-elements: Fallacy of credibility, fallacy of logic, appeal to emotion

Comprehensive

null

Given a text, detect fallacies and classify them

Level 0: binary classification (fallacy or not), Level 1: groups fallacies into Aristotle’s categories: ‘Pathos’ (appeals to emotion), ‘Ethos’ (fallacies of credibility), and ‘Logos’ (fallacies of logic, relevance, or evidence), Level 2 contains fine-grained fallacies within the broad categories of Level 1. For instance, under fallacy of credibility, we have specific fallacies such as appeal to tradition, ad populum, and guilt by association.

null

Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions), Crowd-sourced task examples (e.g. Prolific-created tasks), Modified from another benchmark (e.g. translation into another language)

9735

No

null

Convenience sample (creators found a set of tasks that was readily accessible), Targeted items (creators defined a task space and chose tasks within it strategically)

Multiple choice

Exact Match (accuracy, F1, precision, recall)

null

Academia

Yes

null

Test

null

Simple Mean

Yes

3 levels (different granularity)

null

GitHub

MAFALDA

Widely-agreed

Yes

No

No comparisons made

No

Yes

No

null

simple mean/sum

Outputs alone

Representative task (e.g. answering medical licensing exam questions)

null

Composite phenomenon

Yes

null

Reasoning

Logical

null

['Author-crafted', 'Crowd-sourced', 'Another benchmark']

['Convenience', 'Targeted']

['Multiple choice']

['Exact match']

['Widely-agreed']

['Yes']

['No comparison made']

['No']

['Representative']

['Mean']

niuRAGTruthHallucinationCorpus2024

RAGTruth: A Hallucination Corpus for Developing Trustworthy Retrieval-Augmented Language Models

Include

null

This paper targets word-level hallucinations in various tasks and domains in the RAG setting. It presents approximately 18,000 responses generated using RAG from diverse LLMs which are annotated at the word level for hallucination intensity. Hallucination frequencies are benchmarked across various LLMs, and hallucination detection methods are assessed versus a small LLM fine-tuned using the proposed dataset, RAGTruth.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

hallucination detection, specifically for RAG applications

Yes

"Hallucination in the context of LLMs usually refers to a situation where the model generates content that is not based on factual or accurate information"

Subset

null

For a given reference-response pair, determine if it contains hallucinated content at the response level and span level.

A single item consists of source information (reference), an LLM-generated response (which may contain various degrees of hallucination), annotation of the location and type of hallucination (if any), and a brief annotated explanation of the hallucination observed.

Additional meta-data regarding the model and inference hyperparameters used to generate each sample is provided, along with details regarding the source and task type for the reference texts.

Real task examples (e.g. GitHub issues), Crowd-sourced task examples (e.g. Prolific-created tasks), Modified from another benchmark (e.g. translation into another language), Procedurally-generated task examples (e.g. Creating instances from a template), LLM-generated task examples (e.g. Filtered from responses to a prompt)

2700

Yes

source information index, generating model, temperature, whether quality issues are present in the sample, task type of the data, source of the original content, prompt used to generate the response, base content for RAG

Random sample (creators defined a task space and sampled from it), Targeted items (creators defined a task space and chose tasks within it strategically)

Short free response (e.g. single word or number), Free response (e.g. summary paragarph), Structured response (e.g. valid JSON, API call alone)

Exact Match (accuracy, F1, precision, recall)

null

Mix (multiple authors from industry and academia)

Yes

null

Test, Train

15090 (train)

null

Simple Mean

Yes

by task type (QA, summarization, data-to-text writing)

null

https://github.com/ParticleMedia/RAGTruth

RAGTruth

Widely-agreed

Yes

No

No comparisons made

No

Yes

Benchmark statistics and quality checking are described. Hallucination density is assessed across models used to generate the data, in relation to context length, and versus position in the text.

null

Outputs alone

Complete real task (e.g. providing medical advice to real people interactively)

null

Composite phenomenon

Yes

null

Retrieval

null

Factuality

['Real task', 'Crowd-sourced', 'Another benchmark', 'Procedurally-generated', 'LLM-generated']

['Random', 'Targeted']

['Short free response', 'Free response', 'Structured']

['Exact match']

['Widely-agreed']

['Yes']

['No comparison made']

['Yes']

['Complete']

null

wangIELMOpenInformation2022

IELM: An Open Information Extraction Benchmark for Pre-Trained Language Models

Include

null

They introduce a new open information extraction (OIE) benchmark designed to evaluate the relational knowledge stored in pre-trained language models (LMs) such as BERT and GPT (published in 2022). Their method involves transforming these pre-trained LMs into zero-shot OIE systems to assess their performance on both existing and novel factual OIE datasets. Their results show that pre-trained LMs achieve competitive performance, even surpassing state-of-the-art supervised OIE methods on certain datasets without any additional training data.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

open information extraction i.e. answering “fill-in-the-blank” questions when given a pre-defined relation category

Yes

"In this work, we set up a new open information extraction (OIE) benchmark, called IELM, towards testing the general and open relational information stored in pre-trained LMs."

Comprehensive

For definition_integrity - the paper looks at both standard OIE and factual OIE.

"In this work, we set up a new open information extraction (OIE) benchmark, called IELM, towards testing the general and open relational information stored in pre-trained LMs. We refer to OIE as it is a task that is designed to extract open relations from massive corpora without requiring a pre-defined relation category."

"For open information extraction (OIE), we take an input as a NP-chunked sentence and output a set of triples. Below is an example. Input DylanNP was born in MinnesotaNP, and was awarded Nobel PrizeNP. Output (Dylan; born in; Minnesota), (Dylan; awarded; Nobel Prize). NP denotes the noun phrase."

null

Crowd-sourced task examples (e.g. Prolific-created tasks), Based on knowledge graphs (KG) e.g. Wikidata

27,400,440 triples 6,096,709 arguments 5,418 predicates 9,925,937 documents

No

null

Convenience sample (creators found a set of tasks that was readily accessible)

Structured response (e.g. valid JSON, API call alone)

Exact Match (accuracy, F1, precision, recall)

null

Academia

No, link is broken

null

Test

The dataset size above is summed over 4 datasets in Table 2.

Output is a set of triples

null

Yes

Metrics are reported for each OIE dataset (CaRB(existing), Re-OIE206 (existing), TAC KBP-OIE (novel), Wikidata-OIE (novel)).

null

https://github.com/cgraywang/IELM - This repository is empty.

IELM

Widely-agreed

Yes

No

Yes

They carry out an error analysis: "We argue that we are measuring a lower bound for what LMs know. To further understand the shortcomings of the current method, we conduct an error analysis of the errors in precision on all datasets. We choose BERTLARGE for the study. We sample 100 documents from the Wikidata-OIE dataset, and manually check the reasons for the errors." They find error from: incorrect arguments, missing pairs in predicate mapping, correct triples that are not covered by Wikidata, and incorrect predicate phrases.

The authors carry out some error analysis: "We argue that we are measuring a lower bound for what LMs know. To further understand the shortcomings of the current method, we conduct an error analysis of the errors in precision on all datasets. We choose BERTLARGE for the study. We sample 100 documents from the Wikidata-OIE dataset, and manually check the reasons for the errors"

Outputs alone

Representative task (e.g. answering medical licensing exam questions)

null

Composite phenomenon

No

null

NLP

Extraction

null

['Crowd-sourced', 'Procedurally-generated']

['Convenience']

['Structured']

['Exact match']

['Widely-agreed']

['Yes']

['No comparison made']

['Yes']

['Representative']

['Other']

heTGEAErrorAnnotatedDataset2021

TGEA: An Error-Annotated Dataset and Benchmark Tasks for Text Generation from Pretrained Language Models

Include

null

TGEA (Text Generation Error Annotation) is an error-annotated dataset with multiple benchmark tasks for text generation. Following the authors hierachical error taxonomy, crowdsourced workers manually labeled 12k erroneous sentences with semantic information, including error types, associated text spans, error corrections and rationals behind errors.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Text generation error analysis

Yes

"The key interest of this dataset is detecting and annotating text generation errors from PLMs."

Subset

null

The task requires models to analyze machine-generated Chinese text to detect, locate, classify, correct, and explain generation errors according to a comprehensive taxonomy of error types.

A single item consists of machine-generated Chinese text with annotations marking error spans, associated spans, corrections, error type classifications, and explanatory rationales.

null

LLM-generated task examples (e.g. Filtered from responses to a prompt)

47,058

Yes

error type classification, token counts, error span locations, span distances, error distribution

Targeted items (creators defined a task space and chose tasks within it strategically), Specific criteria (items were taken from a larger set based on specified rules)

Short free response (e.g. single word or number), Free response (e.g. summary paragarph)

Exact Match (accuracy, F1, precision, recall), n-gram (BLEU, ROUGE, chrF), Distribution (perplexity, calibration, correlation)

null

Mix (multiple authors from industry and academia)

Yes

null

Test, Train

train (37,646), Dev (4,706), test (4,706)

null

None, Separate metrics for each sub-task with no single aggregated score

Yes

Erroneous text detection, Erroneous and associated span detection, Error type classification, Error correction, Rationale generation

null

https://download.mindspore.cn/dataset/TGEA/

TGEA

Widely-agreed

Yes

No

No comparisons made

No

Yes

The authors validate their benchmark with inter-annotator agreement statistics for different tasks, Cohen's Kappa coefficients, a rigorous quality control protocol, annotation verification on sampled texts, and human performance baselines.

Simple means for performance metrics; agreement percentages and Cohen's Kappa for annotation reliability.

Outputs alone

Representative task (e.g. answering medical licensing exam questions)

null

Composite phenomenon

Yes

null

Factuality

null

['LLM-generated']

['Targeted', 'Criterion']

['Short free response', 'Free response']

['Exact match', 'Soft match', 'Distribution']

['Widely-agreed']

['Yes']

['No comparison made']

['Yes']

['Representative']

['Mean', 'Other']

huangCEvalMultiLevelMultiDiscipline2023

C-EVAL: A Multi-Level Multi-Discipline Chinese Evaluation Suite for Foundation Models

Include

null

The paper introduces C-EVAL evaluation suite for assessing advanced knowledge and reasoning abilities of foundation models in Chinese, It spans four difficulty levels and 52 disciplines. It also introduces C-EVAL HARD a subset of challenging subjects that require advanced reasoning.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Knowledge and reasoning in Mandarin Chinese and on questions situated in the Chinese context

No

null

Comprehensive

null

Multiple choice questions from real-world human exams in China at different difficultly levels (e.g., high school, college) and different disciplines (e.g., STEM, humanities).

An MCQ question with four possible answers.

null

Human exam questions (e.g. GRE questions)

12342

Yes

topic area (e.g., STEM, humanities) and difficultly level (e.g., middle school)

Convenience sample (creators found a set of tasks that was readily accessible)

Multiple choice

Exact Match (accuracy, F1, precision, recall)

null

Academia

Yes

null

Test, Train, Validation

Dev: 260, Valid: 1346

null

Simple Mean

Yes

Subject/exam (and by extension difficulty)

null

https://github.com/hkust-nlp/ceval/tree/main

C-EVAL

Contested

They follow the lead of popular knowledge and reasoning benchmarks, so it's hard to say here.

Not sure about this. Compared to other similar benchmarks, yes. In general, probably not.

Yes

No

null

simple mean

Outputs alone

Representative task (e.g. answering medical licensing exam questions)

null

Composite phenomenon

Yes

null

Knowledge

Cultural

null

['Human exams']

['Convenience']

['Multiple choice']

['Exact match']

['Contested']

['Partially']

['No comparison made']

['No']

['Representative']

['Mean']

myungBLEnDBenchmarkLLMs2024

BLEnD: A Benchmark for LLMs on Everyday Knowledge in Diverse Cultures and Languages

Include

null

The paper introduces BLEND, a novel benchmark comprising hand-crafted question-answer pairs designed to evaluate LLMs on everyday cultural knowledge across 16 countries/regions and 13 languages, including low-resource ones. It demonstrates significant performance disparities among models, showing cultural and linguistic biases, especially in underrepresented regions.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

cultural knowledge and multilingual cultural commonsense understanding

Yes

knowledge of everyday cultural practices that are specific to different countries and regions. This includes understanding what people commonly do, eat, or experience in their daily lives within a specific cultural and linguistic context. Specifically, dimensions such as food, sports, celebrations, education, family, and work-life are considered.

Subset

null

The task is to evaluate large language models on their ability to correctly answer short-answer and multiple-choice questions about everyday cultural practices from various countries and regions, using either local languages and English. Human evaluation is conducted on short-answer questions with annotators coming from the tested regions.

"Al-en-06": { "question": "대한민국 학교 급식에서 흔히 볼 수 있는 음식은 무엇인가요?", "en_question": "What is a common school cafeteria food in your country?", "annotations": [ { "answers": [ "김치" ], "en_answers": [ "kimchi" ], "count": 4 }, { "answers": [ "밥", "쌀밥", "쌀" ], "en_answers": [ "rice" ], "count": 3 }, ... ], "idks": { "idk": 0, "no-answer": 0, "not-applicable": 0, "others": [] } },

null

Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions), Crowd-sourced task examples (e.g. Prolific-created tasks), Procedurally-generated task examples (e.g. Creating instances from a template)

52.6k

Yes

null

Convenience sample (creators found a set of tasks that was readily accessible), Targeted items (creators defined a task space and chose tasks within it strategically), Specific criteria (items were taken from a larger set based on specified rules)

Multiple choice, Short free response (e.g. single word or number)

Exact Match (accuracy, F1, precision, recall), LLM post-processing (extracting answers, reformatting for automated scoring)

null

Academia

Yes

null

Test

null

Simple Mean

Yes

by language (native and English)/country (region)

null

https://github.com/nlee0212/BLEnD

BLEnD

Contested

Yes

No

No comparisons made

No

Yes

for short-answer questions, there is a human evaluation, which to some extent can represent the validity of the questions

null

simple mean, Anova for p-values, Tukey-HSD

Outputs alone

Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

null

Composite phenomenon

Yes

null

Knowledge

Cultural

null

['Author-crafted', 'Crowd-sourced', 'Procedurally-generated']

['Convenience', 'Targeted', 'Criterion']

['Multiple choice', 'Short free response']

['Exact match', 'LLM post-processing']

['Contested']

['Yes']

['No comparison made']

['Yes']

['Constructed']

['Mean', 'Tests']

yaoWebShopScalableRealWorld2022

WebShop: Towards Scalable Real-World Web Interaction with Grounded Language Agents

Include

null

The paper introduces WebShop, a simulated online shopping environment where agents try to follow natural language instructions to find and buy the right products. WebShop benchmark is designed to test how well agents can search, navigate, and make decisions on the web. The authors train models using imitation and reinforcement learning, and show that the best ones can even handle similar tasks on real sites like Amazon and eBay.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Natural language understanding and sequential decision-making in web environments.

No

To evaluate agents that can understand human-provided natural language instructions and perform grounded actions in a realistic web environment, e.g generating search queries, navigating results, selecting options, and (at the end, if succesful) purchasing a product that matches the instruction.

Subset

null

The task is to follow a natural language instruction to find and purchase a product in a simulated ecommerce environment. Agent must search, navigate pages, select product options, and choose the best match based on the instruction.

Natural language instruction - specifying a desired product (including attributes, options, and price constraints), with the starting state of the simulated shopping environment. The agent must then complete the task by navigating and interacting with the website to find and purchase a suitable product.

null

Real task examples (e.g. GitHub issues), Crowd-sourced task examples (e.g. Prolific-created tasks)

500

Yes

product category, product attributes, product options, product price

Convenience sample (creators found a set of tasks that was readily accessible), Targeted items (creators defined a task space and chose tasks within it strategically), Specific criteria (items were taken from a larger set based on specified rules)

Multiple choice, Free response (e.g. summary paragarph), Extended interaction (e.g. conversation, calling an API and processing the response)

reward is computed based on the final product chosen by the agent, compared against known attributes, options, and price of the target product.

null

Academia

Yes

null

Here the evaluation is fully automated, which allows for easier reproduction - which seems like a significant advantage compared to others.

null

“[...] a total of 12,087 instructions into an i.i.d. distributed train / development / test split of 10,587 / 1,000 / 500 instances"

null

Simple Mean

Yes

Paper reports breakdowns by reward components: attribute match score, option match score, price match, and type match.

null

https://webshop-pnlp.github.io/

WebShop

Contested

Yes

No

No comparisons made

Yes

They discuss the performance gap between models and humans, quite detailed analysis of error types (e.g. failure in option matching or limited exploration), evidence of sim-to-real transfer to Amazon and eBay, aiming to indicate the external validity, as well as component-wise ablations and choice oracle (the model doesn't have to chose) experiments to diagnose bottlenecks

The authors report average task score and success rate across trials. They also include standard deviation/error bars in some result plots (e.g. Figure 4), mainly to show the variation across multiple runs.

Outputs alone

Partial real task (e.g. answering medical questions collected from real people)

WebShop simulates online shopping using real product data and realistic ux, but it operates in a custom environment with a simplified interface and deterministic search engine. So while the core interactions reflect a real-world activity, it doesn’t capture the full complexity or variability of actual web browsing with human properly in the loop or user's behaviour.

Composite phenomenon

No

null

Agents

Web

null

['Real task', 'Crowd-sourced']

['Convenience', 'Targeted', 'Criterion']

['Multiple choice', 'Free response', 'Interaction']

['Reward']

['Contested']

['Yes']

['Comparison made']

['Yes']

['Partial']

['Mean', 'Std']

sanyalRobustLRDiagnosticBenchmark2022

ROBUSTLR: A Diagnostic Benchmark for Evaluating Logical Robustness of Deductive Reasoners

Include

null

Deductive reasoning is an important skill that modern language models should possess. However, small logical perturbations of deductive reasoning problems can lead to inconsistent model responses. To test this consistency, the paper introduces RobustLR a benchmark consisting of logical problems ("theories") and variations thereof that should be consistenly answered correctly by models.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

robustness of deductive reasoning against small shifts in logical operators or rephrasing.

Yes

"We consider a deductive reasoner (language model) to be logically robust if the model behavior is consistent across various logical perturbations."

Comprehensive

Consistency here can be misinterpreted: The perturbations applied to problems cause different conclusions. Consistency is here defined as being accurate across perturbations, i.e. changing the label when the input changes. This is in contrast to many other works that regard consistency as invariance.

The task has 2 levels: The underlying task is conducting deductive reasoning. This is a classification problem: "True", "False" "Unknown". The "meta-task" is being consistent across a set of related problems.

One item in the benchmark is a set: "original problem" + a set of perturbations on the problem. Each problem is a set of facts, rules and deduction.

null

Procedurally-generated task examples (e.g. Creating instances from a template)

null

No

null

Random sample (creators defined a task space and sampled from it), Convenience sample (creators found a set of tasks that was readily accessible)

Multiple choice

Exact Match (accuracy, F1, precision, recall)

null

Academia

Yes

null

The synthetic nature of the benchmark is very much limiting the ecological validity of the benchmark for real user interaction, but the authors are very clear and transparent about it. The lack of ecological validity is compensated by internal validity.

Test

null

yes

Simple Mean

Yes

different kinds of perturbations of the problem.

null

https://github.com/INK-USC/RobustLR

RobustLR

Widely-agreed

Yes

No

No comparisons made

No

Yes

The authors clearly state limitations due to simple composition of rules used for perturbations and the synthetic toy nature of the dataset. They also validate that humans can achieve good scores on the problems while langauge models dont.

mean of weighted-F1 scores

Outputs alone

Representative task (e.g. answering medical licensing exam questions)

null

Composite phenomenon

Yes

null

Reasoning

Logical

Robustness

['Procedurally-generated']

['Random', 'Convenience']

['Multiple choice']

['Exact match']

['Widely-agreed']

['Yes']

['No comparison made']

['Yes']

['Representative']

['Mean']

albalakFETABenchmarkFewSample2022

FETA: A Benchmark for Few-Sample Task Transfer in Open-Domain Dialogue

Include

null

Examines few-sample task transfer across 17 subtasks (e.g., utterance-level classification, dialogue-level classification, span extraction, multiple-choice) in open-domain dialogue with diverse properties (dyadic vs. multi-party, anonymized vs. recurring speaker, varying dialogue lengths).

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Task transfer, transferring knowledge contained in related tasks, in few-sample settings (10% of original instance set)

Yes

Task transfer, transferring knowledge contained in related tasks. Definition 3 (Task Transfer). Given a source task TS = {YS, fS(XS)} and target task TT = {YT , fT (XT )}, task transfer is the use of a learning algorithm, A, to improve the learning of fT by using the knowledge in TS. They also define Few-Sample: For this reason, we focus on the fewsample setting, defined in FETA as 10% of the original instance set. Out of 10%, 5%, and 1%, 10% was empirically determined to be the smallest percentage that retains labels from all label sets in both the train and development partitions.

Subset

They define seperately: (1) Cross-dataset task transfer, when XS ≠ XT , we also have P(XS) ≠ P(XT ) and DS ≠ DT ; domain shift; vs (2) intra-dataset task transfer, when XS = XT , there is no domain shift.

The tasks are classic NLP tasks subsumed in dialog - e.g., emotional recognition during chit-chat conversations, or character identification from a TV transcript.

Input = a dialogue (from DailyDialog); Subtask = Emotion Recognition; Output = Happiness; OR Input = a transcript from a TV Show (from Friends); Subtask = QA, question = How long did Rachael train for?; Output = 2 weeks.

They focus on intra-dataset transfer but not cross-domain transfer.

Modified from another benchmark (e.g. translation into another language), Human TV show; Human chitchat dialogues

71,212

Yes

They provide the datasource (dialog, friends), the task name (e.g., emotion recognition, or QA), and the a categorisation of task type (e.g., utterance classification vs span extraction)

Convenience sample (creators found a set of tasks that was readily accessible)

Depends on the subtask category (Utterance Classification, Dialogue Classification, Multiple Choice, Span Extraction)

Exact Match (accuracy, F1, precision, recall)

null

Mix (multiple authors from industry and academia)

Yes

null

Test, Train, Validation

Train=28,261, Dev = 5,132

null

Simple Mean

Yes

They provide results over the task categories - Utterance Classification, Dialogue Classification, Multiple Choice, Span Extraction

null

https://alon-albalak.github.io/feta-website/

FETA

Widely-agreed

Partially

Yes

No

No comparisons made

No

null

Mean, and they they show a delta (for change in aggregate sources across all tasks). It is unclear if this is a range or a standard deviation. I think it's a range.

Outputs alone

Partial real task (e.g. answering medical questions collected from real people)

Using the model for various tasks contained in dialogue seems a more general ecologically valid use case, than the Friends transcript understanding but this could also be an applied usecase.

Composite phenomenon

Yes

null

Language Modelling

Adaptability

null

['Another benchmark', 'Author-crafted']

['Convenience']

['Short free response']

['Exact match']

['Widely-agreed']

['Partially']

['No comparison made']

['No']

['Partial']

['Mean']

beanLINGOLYBenchmarkOlympiadLevel2024

LINGOLY: A Benchmark of Olympiad-Level Linguistic Reasoning Puzzles in Low Resource and Extinct Languages

Include

null

The paper introduces LINGOLY, a new benchmark built on Linguistics Olympiad puzzles in low-resource and extinct languages to test genuine reasoning capabilities in LLMs. The benchmark is crafted covering diverse reasoning complexity, linguistic subject areas, instruction types, and high/low resources. The paper uncovers error pattenrs between high and low resource settings and show the ongoing challenges in multi-step, out-of-domain reasoning.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Multi-step, out-of-domain linguistic reasoning, low-resource languages,

Yes

We argue that a benchmark task measures reasoning if the task 1) cannot be done without reasoning (necessity) and 2) can be done via reasoning (sufficiency). However, the combination of these features is difficult to achieve in practice since memorisation and contamination may reduce the necessity of reasoning, and in tasks which draw on background knowledge, as in most ‘commonsense’ benchmarks[7], reasoning itself is insufficient to complete the task.

Subset

No-context baseline -- evaluate if the model performance drops when the context is removed. This concept is to assess the performance if the model has relied on memorization or reasoning from the linguistic clues in the context.

The task is to understand genuine reasoning capabilities of LLMs by providing context of low-resource linguistic information and questions to solve based on the given context (or without context to penalize the memorized knowledge). The expected output is a concise textual answer that can be matched up with ground-truth labels.

Below is a problem sheet… {PREAMBLE} {CONTEXT} {QUESTIONS} {SUBQUESTIONS} Now respond to the following… {REPEAT 1 QUESTION} Format your response as… {FORMAT TEMPLATE}

Compare the model performance with and without contextual information to penalize the memorized knowledge and evaluate the genuine reasoning abilities of LLMs using the linguistic cues from the given knowledge.

Human exam questions (e.g. GRE questions), Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions)

null

Yes

human difficulty, linguistic subjects, task format, language

Convenience sample (creators found a set of tasks that was readily accessible)

Multiple choice, Short free response (e.g. single word or number), Structured response (e.g. valid JSON, API call alone)

Exact Match (accuracy, F1, precision, recall)

null

The task from LINGOLY is adapted from official Linguistics Olympiads puzzle sets rather than everyday language usage scenarios or standard benchmarking corpora.

Academia

Yes

null

One critical point is whether language models provide poor performances due to the unfamiliar format or out-of-domain reasoning -- the mismatch between the puzzle's presentation style and the distribution of model instruction templates may cause certain reasoning failures depending on model types. It would be nice to see how benchmarks have certain patterns with model types.

Test

1,133 questions all for testing.

Free response is existed but excluded from evaluation (The only case where an instance has a missing answer is when the intended answer was a free response, e.g., “explain your reasoning”. These questions are included in the dataset but removed from the scoring as they are not compatible with being machine-scored.)

Simple Mean

Yes

Human difficulty, puzzle format, linguistic subject, language resourcedness

null

The huggingface is working great while the Github zip file requires passcode to get access.

LINGOLY

Widely-agreed

Yes

No

No comparisons made

No

Yes

Across models, performance is consistently higher on problems with easier human difficulty and higher resource languages than those of harder difficulty and lower-resource languages. (LLMs tested have limited reasoning abilities about low-resource languages and do not achieve the multi-step reasoning required in the harder questions, in addition to errors of following instructions alongside core reasoning tasks.)

The authors use a weighted mean in calculating an approximate human performance threshold but not for model performance. They take a weighted average of the annual medal thresholds for ‘Advanced’ problems.

Outputs alone

Representative task (e.g. answering medical licensing exam questions)

While the benchmark comes from authentic Linguistic Olympiad puzzles, they are still competition-style questions rather than real world usage scenarios. Hence it can be categorized as representative task of a speciflized exam setting.

Single cohesive phenomenon

No

null

Reasoning

Logical

null

['Human exams', 'Author-crafted']

['Convenience']

['Multiple choice', 'Short free response', 'Structured']

['Exact match']

['Widely-agreed']

['Yes']

['No comparison made']

['Yes']

['Representative']

['Mean']

nasirGameTraversalBenchmarkEvaluatingPlanning2024

GameTraversalBenchmark: Evaluating Planning Abilities Of Large Language Models Through Traversing 2D Game Maps

Include

null

The paper investigates the planning capabilities of LLMs by proposing GameTraversalBenchmark (GTB), a benchmark consisting of diverse 2D grid-based game maps. The paper also provide metrics to give insights towards planning abilities in LLMs.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Planning abilities of LLMs

No

null

Subset

null

The task is a game based on 2D maps. They consider a generated map as one data point for the benchmark. The map’s generated objective coordinates are the points where the LLM agent needs to traverse to attain the most rewards.

Each item is a 2D grid-based map if alphanumeric characters.

null

LLM-generated task examples (e.g. Filtered from responses to a prompt)

150

No

null

Unknown

Structured response (e.g. valid JSON, API call alone)

Exact Match (accuracy, F1, precision, recall), The paper defines a reward score

null

Academia

Yes

null

Test

null

Simple Mean

No

null

https://github.com/umair-nasir14/Game-Traversal-Benchmark/

GameTraversalBenchmark (GTB)

Not defined

Yes

No

No comparisons made

No

null

simple mean and STD

Outputs alone

null

Single cohesive phenomenon

Not applicable

null

Reasoning

Planning

null

['LLM-generated']

['Unknown']

['Structured']

['Exact match', 'Reward']

['No definition']

['Yes']

['No comparison made']

['No']

['']

['Mean', 'Std']

feiLawBenchBenchmarkingLegal2024

LawBench: Benchmarking Legal Knowledge of Large Language Models

Include

null

LawBench tests 21 models on 20 Chinese legal tasks (500 instances each), which are classified along Bloom's taxonomy into knowledge memorization, understanding, and application. It is the first benchmark for the Chinese legal domain, and the first for civil law (vs. common law) jurisdictions.

null

Specific Application (A single use case, where the benchmark is likely to be examples of that use case)

legal knowledge memorization, understanding, and application

Yes

LawBench is the first evaluation benchmark developed for the Chinese legal domain. It defines the phenomenon in terms of legal knowledge capabilities mapped to cognitive levels from Bloom’s Taxonomy.

Subset

Bloom's taxonomy for task grouping

Perform 20 specific legal functions using text-based input and return a defined output (of various forms, including classification label, summary, number)

Varies strongly between the 20 tasks, but generally: a legal input (fact description, question, judgement) and a required output of various forms.

null

Real task examples (e.g. GitHub issues), Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions), Modified from another benchmark (e.g. translation into another language), LLM-generated task examples (e.g. Filtered from responses to a prompt)

10000

Yes

Task ID, blooms taxonomy level (used to indicate difficulty), task type, metric

Convenience sample (creators found a set of tasks that was readily accessible), Targeted items (creators defined a task space and chose tasks within it strategically), Specific criteria (items were taken from a larger set based on specified rules)

Multiple choice, Short free response (e.g. single word or number), Free response (e.g. summary paragarph)

Exact Match (accuracy, F1, precision, recall), n-gram (BLEU, ROUGE, chrF), LLM post-processing (extracting answers, reformatting for automated scoring)

null

Most tasks adapted from existing legal datasets: CAIL, JEC_QA, and LEVEN.

Mostly academia, 1 research institute, 1 high school

Yes

null

Test

null

Response format varies by task. Dataset sampling above: mostly "convenience sampled"/rehashed from existing benchmarks.

Simple Mean

Yes

By task (each of 20), by blooms taxonomy level (each of memorization, understanding, application), by zero-shot vs. one-shot

null

https://github.com/open-compass/LawBench

LawBench

Widely-agreed

Yes

No

null

Simple means and macro-averaging (mean across tasks, which is identical here because each task has same # of instances)

Outputs alone

Representative task (e.g. answering medical licensing exam questions)

Validity varies strongly between tasks. Memorization tasks (2/20) do not reflect real-world human work. Most others are taken from benchmarks in QA format. Some are "partial real tasks" eg. answering legal questions scraped from a legal QA site.

Composite phenomenon

Yes

null

Law

null

['Real task', 'Author-crafted', 'Another benchmark', 'LLM-generated']

['Convenience', 'Targeted', 'Criterion']

['Multiple choice', 'Short free response', 'Free response']

['Exact match', 'Soft match', 'LLM post-processing']

['Widely-agreed']

['Yes']

['No comparison made']

['No']

['Representative']

['Mean']

yuksekgonulWhenWhyVisionlanguage2023

When and Why Vision-Language Models Behave like Bags-Of-Words, and What to Do About It?

Include

null

This paper creates the Attribution, Relation, and Order (ARO) benchmark to systematically evaluate the ability of VLMs to understand different types of relationships, attributes, and order information. They demonstrate that VLMs can perform well on image-text retrieval over existing datasets without using the composition and order information.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Compositional understanding in VLMs

No

null

Subset

null

ARO consists of Visual Genome Attribution, to test the understanding of objects’ properties; Visual Genome Relation, to test for relational understanding; and COCO-Order & Flickr30k-Order, to test for order sensitivity in VLMs.

A sample would be an image, 1 true and 1 false statement about the image, the two objects presented in the image, the attributes of the objects

null

Modified from another benchmark (e.g. translation into another language)

28,700

No

null

Specific criteria (items were taken from a larger set based on specified rules)

Multiple choice, Short free response (e.g. single word or number)

Exact Match (accuracy, F1, precision, recall)

null

Academia

Yes

null

Test

null

Simple Mean

No

Stratification based on the four introduced tasks: 1) Visual Genome Attributions, 2) Visual Genome Relations, 3) COCO Order and 4) Flickr30k Order

null

https://huggingface.co/datasets/gowitheflow/ARO-Visual-Attribution

ARO

Not defined

Yes

No

null

macro-accuracy

Outputs alone

Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

null

Composite phenomenon

Yes

null

Reasoning

Compositional

null

['Another benchmark']

['Criterion']

['Multiple choice', 'Short free response']

['Exact match']

['No definition']

['Yes']

['No comparison made']

['No']

['Constructed']

['Mean']

xieWhodunitBenchEvaluatingLarge2024

WhodunitBench: Evaluating Large Multimodal Agents via Murder Mystery Games

Include

null

The paper evaluates LLMs ability to participate in (and answers questions about) murder mystery games. In the arena component (agents play as either detective or murderer in a multi-agent setting), the agents are tested on win rate against the other models. The QA component is split based on capability categories (Perception, Role-Play, Decision-making and Cognition)

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

The authors evaluate four distinct capabilities: multi-modal perception, interaction, reasoning and goal achievement.

Yes

• Multi-modal Perception is the most basic ability for LMAs, which requires LMAs to perceive information from the multimodal environment (e.g., vision and language). • Interaction requires LMAs, whether through role-playing or direct engagement, to communicate with the environment or other agents to gather essential information for task completion. • Reasoning requires LMAs to combine their internal knowledge with newly gathered information to perform long-chain, multi-step reasoning. • Decision Making and Goal Achievement requires LMAs to establish clear goals and make independent decisions in response to environmental changes. This autonomous decision-making is crucial for effectively navigating and completing tasks in dynamic settings.

Subset

Since the benchmarks evaluates many things, the level of detail differs between the constructs.

The agent arena component is based on "winning" in a murder mystery game, whereas the Chain-of-Evaluation component is based on a QA format.

In the arena setting, each task item is a single murder mystery game with a winner. In the CoE, each task is a multiple-choice question.

null

Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions), Crowd-sourced task examples (e.g. Prolific-created tasks)

3000

No

null

Convenience sample (creators found a set of tasks that was readily accessible)

Multiple choice, Extended interaction (e.g. conversation, calling an API and processing the response)

Exact Match (accuracy, F1, precision, recall), LLM-as-a-Judge (text quality, preferences, NOT extracting answers for other metrics), Win rate

null

The arena is based on a script, and the questions are manually annotated. The murder game scripts ccome from real sources.

Academia

Repo without any code is provided.

null

Test

Only reported approximately

CoE is multiple choice, arena is extended interaction

Simple Mean

No

null

https://github.com/jun0wanan/WhodunitBench-Murder_Mystery_Games

WhodunitBench

Contested

Yes

No

No comparisons made

No

null

Simple mean (no variance or standard reported)

Outputs alone

Partial real task (e.g. answering medical questions collected from real people)

It is based on a pure "fictional" game, with the hope that capabilities are general enough to transfer.

Composite phenomenon

Yes

null

Agents

null

['Author-crafted', 'Crowd-sourced']

['Convenience']

['Multiple choice', 'Interaction']

['Exact match', 'LLM-as-a-Judge', 'Reward']

['Contested']

['Yes']

['No comparison made']

['No']

['Partial']

['Mean']

saparinaAMBROSIABenchmarkParsing2024

AMBROSIA: A Benchmark for Parsing Ambiguous Questions into Database Queries

Include

null

Paper introduces a new benchmark dataset designed to evaluate text-to-SQL parsers' ability to handle ambiguous user requests. The dataset includes questions demonstrating scope ambiguity, attachment ambiguity, and vagueness, along with their interpretations and corresponding SQL queries. The authors highlight that existing large language models (LLMs) struggle with these ambiguities, suggesting a need for improved parser development.

null

Specific Application (A single use case, where the benchmark is likely to be examples of that use case)

text-to-SQL parsing

Yes

Evaluation of text-to-SQL parsers capable of recognizing and interpreting ambiguous requests

Comprehensive

null

text-to-SQL parsing, generate database, validate generated databases

Question, prompt, SQL query, scope/ambiguity/vagueness, generated database, score (human annotation)

null

Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions), LLM-generated task examples (e.g. Filtered from responses to a prompt)

5093

null

Targeted items (creators defined a task space and chose tasks within it strategically)

Structured response (e.g. valid JSON, API call alone)

Exact Match (accuracy, F1, precision, recall), Human ratings (text quality, preference, NOT manual scoring of other metrics)

null

Academia

Yes

null

Test

null

Simple Mean

Yes

null

https://ambrosia-benchmark.github.io/

AMBROSIA

Widely-agreed

Yes

No

null

mean and variance

Outputs alone

Representative task (e.g. answering medical licensing exam questions)

null

Composite phenomenon

Yes

null

Code Generation

Natural Language

null

['Author-crafted', 'LLM-generated']

['Targeted']

['Structured']

['Exact match', 'Human ratings']

['Widely-agreed']

['Yes']

['No comparison made']

['No']

['Representative']

['Mean', 'Std']

augustyniakThisWayDesigning2022

This is the way: designing and compiling LEPISZCZE, a comprehensive NLP benchmark for Polish

Include

null

Authors introduce LEPISZCZE, a new, comprehensive benchmark for Polish NLP with a large variety of tasks and high-quality operationalization of the benchmark. LEPISZCZE was designed with flexibility in mind. Including new models, datasets, and tasks is as simple as possible while still offering data versioning and model tracking. In the first run of the benchmark, 13 experiments (task and dataset pairs) were tested based on the five most recent LMs for Polish. Five datasets from the Polish benchmark are reused and eight novel datasets are added.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

model performance on Polish language across various tasks (13)

null

The ability of language models to understand and process Polish language across a diverse range of NLP tasks, evaluated using 13 task-dataset pairs that include classification, natural language inference, and sequence labeling tasks.

Subset

null

Each task in the LEPISZCZE benchmark is defined as a standard NLP problem—such as classification, sequence labeling, or natural language inference—applied to Polish-language datasets. These tasks test specific linguistic capabilities of models, like sentiment analysis, named entity recognition, part-of-speech tagging, and others.

there are datasets for 13 tasks.

Entailment Classification, Q&A Classification, Sentiment Analysis, Paraphrase Classification, Abusive Clauses Detection, Aspect-based Sentiment Analysis , NER, POS Tagging, Political Advertising Detection, Punctuation Restoration, Punctuation Restoration. Dialogue Acts Classification

Real task examples (e.g. GitHub issues), Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions), Crowd-sourced task examples (e.g. Prolific-created tasks), Modified from another benchmark (e.g. translation into another language)

30,003

No

null

Convenience sample (creators found a set of tasks that was readily accessible), Targeted items (creators defined a task space and chose tasks within it strategically), Specific criteria (items were taken from a larger set based on specified rules)

Short free response (e.g. single word or number), Structured response (e.g. valid JSON, API call alone)

Exact Match (accuracy, F1, precision, recall)

null

Academia

Yes

null

Test, Train, Validation

204,504 and 9,970

null

Simple Mean

No

null

https://huggingface.co/clarin-pl , https://github.com/CLARIN-PL/LEPISZCZE

LEPISZCZE

Contested

Yes

No

No comparisons made

No

null

mean and standard deviation

Outputs alone

Partial real task (e.g. answering medical questions collected from real people)

null

Composite phenomenon

Yes

null

Multilinguality

null

['Real task', 'Author-crafted', 'Crowd-sourced', 'Another benchmark']

['Convenience', 'Targeted', 'Criterion']

['Short free response', 'Structured']

['Exact match']

['Contested']

['Yes']

['No comparison made']

['No']

['Partial']

['Mean', 'Std']

huiUDABenchmarkSuite2024

UDA: A Benchmark Suite for Retrieval Augmented Generation in Real-world Document Analysis

Include

null

The paper introduces the UDA (Unstructured Document Analysis) benchmark. UDA questions are expert-annotated Q&A pairs on PDF and HTML documents, constructed from datasets of academic papers, financial reports, and Wikipedia pages.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Analysing unstructured documents

No

Vague and multifaceted: "we propose a benchmark suite that enables the evaluation of various components of RAG-based unstructured document analysis"

Subset

null

LLMs are given an unstructured document and a factual question about the contents of that document. The correct answer is some extracted text or figure from the document.

An unstructured document might be a financial report in PDF format, containing tabular data. The question might ask for the total value of some column, like "total vested shares during the 2012 fiscal year, in millions," and correct answers might be [1.46, 1.45972].

null

Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions), Modified from another benchmark (e.g. translation into another language)

29,590

Yes

topic area

Convenience sample (creators found a set of tasks that was readily accessible)

Short free response (e.g. single word or number), Free response (e.g. summary paragarph)

Exact Match (accuracy, F1, precision, recall), n-gram (BLEU, ROUGE, chrF)

null

Hand-written answers are "expert annotated" by the authors of six Q&A datasets; the authors curate and filter these without changing the labels.

Academia

Yes

null

Test

null

"Free responses" are intended to be extracted from the provided file's text.

Simple Mean

Yes

Scores by underlying Q&A dataset, context type (whether document chunks are provided by RAG or by human annotators)

null

https://github.com/qinchuanhui/UDA-Benchmark

UDA

Widely-agreed

Yes

No

Yes

No

No comparisons made

No

null

Simple mean/sum; % improvement between contexts

Outputs alone

Representative task (e.g. answering medical licensing exam questions)

null

Composite phenomenon

Yes

null

Retrieval

null

['Author-crafted', 'Another benchmark']

['Convenience']

['Short free response', 'Free response']

['Exact match', 'Soft match']

['Widely-agreed']

['Yes']

['No']

['No comparison made']

['No']

['Representative']

['Mean', 'Other']

xiaFOFOBenchmarkEvaluate2024

FOFO: A Benchmark to Evaluate LLMs’ Format-Following Capability

Include

null

FOFO Is a benchmark for domain-specific format following capabilities. It evaluates a wide array of domains and subdomains across a diverse set of formats from specific medical forms to Maple. The specific examples are generated using GPT-4 and human validation.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Format following

Yes

"precise adherence to specified formats given by humans"

Subset

null

The task is to generate dummy data in a specified format defined by detailed instructions within a given domain.

A single formatting instruction with a domain (e.g., Manufacturing), a subdomain (e.g., Optimization), and a format (e.g., "Standard Operating Procedures") with an example of the format.

null

LLM-generated task examples (e.g. Filtered from responses to a prompt)

494

Yes

domain,subdomain,format

Convenience sample (creators found a set of tasks that was readily accessible)

Structured response (e.g. valid JSON, API call alone)

LLM-as-a-Judge (text quality, preferences, NOT extracting answers for other metrics)

null

Mix (multiple authors from industry and academia)

Yes

null

Test

null

Simple Mean

Yes

null

https://github.com/SalesforceAIResearch/FoFo

FOFO

Widely-agreed

Yes

No

null

Outputs alone

Partial real task (e.g. answering medical questions collected from real people)

While following formatting instructions is real, the data is still dummy.

Composite phenomenon

Yes

null

Instruction Following

null

['LLM-generated']

['Convenience']

['Structured']

['LLM-as-a-Judge']

['Widely-agreed']

['Yes']

['No comparison made']

['No']

['Partial']

null

wangMINTEvaluatingLLMs2024

MINT: EVALUATING LLMS IN MULTI-TURN INTERACTION WITH TOOLS AND LANGUAGE FEEDBACK

Include

null

MINT extends existing benchmark to evaluate the effects of code interpreter usage and multi-turn feedback on LLM performance. It filters benchmark task to ones that benefit from feedback and multi-turn interactions and evaluates different feedback types from "lazy user" to "informative user" and with(out) tools.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Reasoning, coding, and decision-making

No

null

Subset

Each high-level phenomena is measured separately

The task is how performance on existing benchmarks (QA) increases when given access to GPT-4 feedback and/or code interpretor.

The tasks come from different benchmarks. Most are in a QA format.

null

Modified from another benchmark (e.g. translation into another language)

586

Yes

source dataset

Random sample (creators defined a task space and sampled from it)

Short free response (e.g. single word or number), Extended interaction (e.g. conversation, calling an API and processing the response)

Exact Match (accuracy, F1, precision, recall)

null

The tasks are sampled from 8 different benchmarks.

Academia

Yes

null

Test

null

While the expected result is often a short free response, it can be created through interaction.

Simple Mean

Yes

Provided by number of turns of feedback

null

https://github.com/xingyaoww/mint-bench

MINT

Contested

Yes

No

No comparisons made

They do a partial study with actual human feedback on the benchmark tasks.

No

null

Outputs alone

Representative task (e.g. answering medical licensing exam questions)

null

Single cohesive phenomenon

Not applicable

null

Agents

Coding

null

['Another benchmark']

['Random']

['Short free response', 'Interaction']

['Exact match']

['Contested']

['Yes']

['Comparison made']

['No']

['Representative']

null

valmeekamPlanBenchExtensibleBenchmark2023

PlanBench: An Extensible Benchmark for Evaluating Large Language Models on Planning and Reasoning about Change

Include

null

PlanBench introduces a suite of tasks relevant to planning using similar formats to the International Planning Competition. The tasks are taken from either Blocksworld or logistics and also obfuscated to avoid reliance on common-sense knowledge.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Planning

Yes

planning involves coming up with a course of actions (policy) which when executed would take an agent from a certain initial state to a desired world state

Subset

null

The main task (planning) is given a description of a state (e.g., block configuration), rules, and a goal state, come up with a plan that transforms from state the goal state. The sub-tasks are variations of components.

A specified state, actions, and goal state + a query for what the LLM should do (compe up with a plan, predict plan execution) etc.

There are in total 8 different tasks with slightly different goals (e.g., direct planning, replanning, execution prediction)

Procedurally-generated task examples (e.g. Creating instances from a template)

1910

Yes

domain

Random sample (creators defined a task space and sampled from it)

Free response (e.g. summary paragarph), Structured response (e.g. valid JSON, API call alone)

Exact Match (accuracy, F1, precision, recall)

null

Academia

Yes

null

Test

null

The plan is a fairly structured set of actions, but not quite as structured as e.g., an API

Simple Mean

Yes

Domain, Obfuscated (Bool)

null

https://github.com/karthikv792/LLMs-Planning

PlanBench

Contested

Yes

No

No comparisons made

No

null

Outputs alone

Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

The task is based on real competition but which has a level of gaminess

Composite phenomenon

Yes

null

Reasoning

Planning

null

['Procedurally-generated']

['Random']

['Free response', 'Structured']

['Exact match']

['Contested']

['Yes']

['No comparison made']

['No']

['Constructed']

null

zhangMELAMultilingualEvaluation2024

MELA: Multilingual Evaluation of Linguistic Acceptability

Include

null

The paper intorduces a multilingual acceptability judgement benchmark covering a diverse set of 10 languages, all annotated by expert linguists. The acceptability judgment task tests a language model’s ability to distinguish syntactically acceptable sentences from unacceptable ones in a human language. The paper establishes LLM baselines on this benchmark, and investigates cross-lingual transfer in acceptability judgements with XLM-R.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Linguistic Acceptability

Yes

The acceptability judgment task tests a language model’s ability to distinguish syntactically acceptable sentences from unacceptable ones.

Comprehensive

null

The acceptability judgment task tests a language model’s ability to distinguish syntactically acceptable sentences from unacceptable ones.

a sentence

null

hand-written by linguists in respective languages, taken from textbooks, handbooks and journal articles in theoretical syntax + some examples taken from previous benchmarks

46k

No

null

Random sample (creators defined a task space and sampled from it)

Multiple choice

Exact Match (accuracy, F1, precision, recall), Matthews Correlation Coefficient (MCC, Matthews), which is a measure of similarity between binary distributions taking values from -1 to 1 and always yielding 0 for any two uncorrelated distributions, regardless of class imbalance.

null

Academia

Yes

null

Test, Train, Validation

train set: 33'293, validation:3'970

null

Simple Mean

No

null

https://github.com/sjtu-compling/MELA

MELA

Widely-agreed

Yes

No

null

No

null

simple mean and standard deviation

Outputs alone

Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

null

Single cohesive phenomenon

Not applicable

null

Multilinguality

null

['Expert-crafted']

['Random']

['Multiple choice']

['Exact match', 'Correlation']

['Widely-agreed']

['Yes']

['No comparison made']

['No']

['Constructed']

['Mean', 'Std']

etxanizLatxaOpenLanguage2024

Latxa: An Open Language Model and Evaluation Suite for Basque

Include

null

The paper introduces 4 multiple-choice evaluation datasets for Basque: EusProfi-ciency, comprising 5,169 questions from official language proficiency exams; EusReading, comprising 352 reading comprehension questions; EusTrivia, comprising 1,715 trivia questions from 5 knowledge areas; and EusExams, comprising 16,774 questions from public examinations.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

language proficiency, knowledge and reasoning

No

null

Subset

The benchmark includes 4 different tasks

There are 4 tasks: reading comprehension, language proficency, mcq questions on Basque language and culture, and mcq questions on Basque government

an mcq question

null

Human exam questions (e.g. GRE questions)

~7.5k

No

null

Unknown

Multiple choice

Exact Match (accuracy, F1, precision, recall)

null

Academia

Yes

null

Test

null

Simple Mean

No

null

https://github.com/hitz-zentroa/latxa?tab=readme-ov-file

null

Widely-agreed

Yes

The benchmark is itself realistic

No

null

accuracy, F1, standard deviation

Outputs alone

Representative task (e.g. answering medical licensing exam questions)

null

Single cohesive phenomenon

Not applicable

null

Multilinguality

null

['Human exams']

['Unknown']

['Multiple choice']

['Exact match']

['Widely-agreed']

['Yes']

['Realistic']

['No']

['Representative']

['Mean', 'Std', 'Other']

tangStrucbenchAreLarge2024

Struc-Bench: Are Large Language Models Good at Generating Complex Structured Tabular Data?

Include

null

The paper introduces a new benchmark to assess LLMs’ proficiency in structuring tables and introduces a novel fine-tuning method, cognizant of data structures, to bolster their performance.

null

Specific Application (A single use case, where the benchmark is likely to be examples of that use case)

Generating structured tabular data

Yes

LLMs are tasked with generating complex struc- tured tables, a process that involves understanding both the content and the specific format require- ments, such as LaTeX syntax. This task extends beyond simple text generation as it demands preci- sion not just in content creation but also in adhering to a detailed and precise structural format.

Comprehensive

null

The task is generating structured tabular data.

text tables, HTML tables, and LaTeX tables and their description

null

Modified from another benchmark (e.g. translation into another language)

~16k

No

null

Random sample (creators defined a task space and sampled from it)

Structured response (e.g. valid JSON, API call alone)

P-Score (Prompting Score) and H-Score (Heuristical Score)

null

Academia

Yes

null

Test, Train

Train: 14.1k, Test:1700

null

Simple Mean

No

null

https://github.com/gersteinlab/Struc-Bench?tab=readme-ov-file

Struc-Bench

Widely-agreed

Yes

No

No comparisons made

No

null

simple mean

Outputs alone

Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

null

Single cohesive phenomenon

No

null

Code Generation

null

['Another benchmark']

['Random']

['Structured']

['LLM-as-a-Judge']

['Widely-agreed']

['Yes']

['No comparison made']

['No']

['Constructed']

['Mean']

riemenschneiderExploringLargeLanguage2023

Exploring Large Language Models for Classical Philology

Include

null

They define two probing tasks to investigate the knowledge acquired by models pre-trained on Classical texts. The experiments provide the first benchmarking analysis of existing models of Ancient Greek.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

null

No

The tasks are supposed to assess semantic and world knowledge in LLMs.

Comprehensive

null

Measuring semantic and world knowledge in LLMs

A sentence

null

Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions)

~550

No

null

Unknown

Multiple choice

Exact Match (accuracy, F1, precision, recall)

null

Academia

Link is provided but the data is not there

null

Test, Train

null

No

null

https://github.com/Heidelberg-NLP/ancient-language-models/tree/main

null

Not defined

null

Yes

No

null

No

null

Outputs alone

Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

null

Multilinguality

null

['Author-crafted']

['Unknown']

['Multiple choice']

['Exact match']

['No definition']

['']

['Yes']

['No comparison made']

['No']

['Constructed']

null

qiPreservingKnowledgeInvariance2023

Preserving Knowledge Invariance: Rethinking Robustness Evaluation of Open Information Extraction

Include

null

The paper introduces ROBUST, a benchmark designed to evaluate open information extraction models by measuring their ability to generalize knowledge extraction across syntactically diverse sentences that share the same semantic content.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

the generalization of open information extraction

Yes

[...] each example is a knowledge-invariant clique that consists of sentences with structured knowledge of the same meaning but with different syntactic and expressive forms. [...] a model is judged to be robust if its performance is consistently accurate on the overall cliques.

Comprehensive

null

Open Information Extraction (OpenIE) aims to extract n-ary knowledge tuples {(a1,p,a2,...,an)} consisting of n arguments and one predicate from the natural text.

Sentences with arguments and one predicate form a set (clique), where sentences are semantically invariant.

The base task is OpenIE. Each tuple of sentence+arguments+predicate within a clique is analyzed. The "meta-task" is doing well on the worst tuple within one clique.

Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions), Modified from another benchmark (e.g. translation into another language), Procedurally-generated task examples (e.g. Creating instances from a template)

1272 cliques, 4971 sentences

No

null

Random sample (creators defined a task space and sampled from it), Convenience sample (creators found a set of tasks that was readily accessible)

Structured response (e.g. valid JSON, API call alone)

Exact Match (accuracy, F1, precision, recall)

null

Academia

Yes

null

Test

null

n-tuples of text are extracted from the resonse.

Simple Mean

No

null

https://github.com/qijimrc/ROBUST

ROBUST

Widely-agreed

Yes

No

null

For each tuple, the F1 is computed, then across a clique the minimum is computed and aggregated across the dataset as mean.

Outputs alone

Representative task (e.g. answering medical licensing exam questions)

null

Single cohesive phenomenon

Not applicable

null

NLP

Extraction

null

['Author-crafted', 'Another benchmark', 'Procedurally-generated']

['Random', 'Convenience']

['Structured']

['Exact match']

['Widely-agreed']

['Yes']

['No comparison made']

['No']

['Representative']

['Mean']

shahWhenFLUEMeets2022

WHEN FLUE MEETS FLANG: Benchmarks and Large Pre-trained Language Model for Financial Domain

Include

null

the Financial Language Understanding Evaluation (FLUE), an open-source comprehensive suite of benchmarks for the financial domain. These include new benchmarks across 5 NLP tasks in financial domain as well as common benchmarks used in the previous research.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

natural language understanding in the financial domain

Yes

The ability of LLMs to perform across 5 financial tasks such as financial sentiment analysis, news headline classification, named entity recognition, structure boundary detection, and question answering.

Subset

null

The task is defined as evaluating language models on a suite of five financial domain NLP tasks: financial sentiment analysis, news headline classification, named entity recognition, structure boundary detection, and question answering.

N/A, for every task there will be a respective item

null

Real task examples (e.g. GitHub issues), Modified from another benchmark (e.g. translation into another language)

969, 234, 2282, 302, 131, 333

No

null

Convenience sample (creators found a set of tasks that was readily accessible), Targeted items (creators defined a task space and chose tasks within it strategically), Specific criteria (items were taken from a larger set based on specified rules)

Short free response (e.g. single word or number), Structured response (e.g. valid JSON, API call alone)

Exact Match (accuracy, F1, precision, recall)

null

Mix (multiple authors from industry and academia)

Yes

null

Test, Train, Validation

for all 5 tasks: 19,367 and 2,674

null

Simple Mean

No

null

https://salt-nlp.github.io/FLANG/

FLUE

Contested

Yes

No

null

Simple mean: F1 scores and accuracy. MSE. nDCG and MRR. Perplexity

Outputs alone

Partial real task (e.g. answering medical questions collected from real people)

null

Composite phenomenon

Yes

null

Finance

null

['Real task', 'Another benchmark']

['Convenience', 'Targeted', 'Criterion']

['Short free response', 'Structured']

['Exact match']

['Contested']

['Yes']

['No comparison made']

['No']

['Partial']

['Mean', 'Other']

kalyanWikiDONewBenchmark2024

WikiDO: A New Benchmark Evaluating Cross-Modal Retrieval for Vision-Language Models

Include

null

The authors argue that current VLM benchmarks are insufficient to assess the OOD generalization capability of models due to high visual and linguistic similarity between the evaluation and finetuning datasets. The propose WIKIDO which consists of image-text data derived from Wikipedia Diversity Observatory, a diverse source of Wikipedia articles spanning several diversity axes including geography, gender, ethnicity and domains/topics.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Generalization / OOD performance

No

null

Subset

null

The proposed dataset can be used for both image-to-text, i.e. retrieve the most relevant textual description(s) from a set, and text-to-image retrieval, i.e. retrieve the most relevant image(s) from a dataset.

A single row in the dataset will have the path of the image, the Wiki ID of the image, the reference text from Wikipedia, the title of the wikipedia article, the topic label from Wikipedia Diversity Observatory and the generated caption of the image

null

Modified from another benchmark (e.g. translation into another language), Procedurally-generated task examples (e.g. Creating instances from a template), LLM-generated task examples (e.g. Filtered from responses to a prompt)

train: 384K pairs, 2 test sets (ID and OOD) of size 3K each.

Yes

topic

Targeted items (creators defined a task space and chose tasks within it strategically)

Retrieval

Exact Match (accuracy, F1, precision, recall)

null

Mix (multiple authors from industry and academia)

Yes

null

Test, Train

train: 384K pairs, 2 test sets (ID and OOD) of size 3K each.

null

Simple Mean

Yes

In-distribution vs Out-of-distribution

null

https://huggingface.co/datasets/Pavankalyan/WikiDO

WikiDO

Widely-agreed

Yes

No

No comparisons made

No

Yes

The authors show that across various settings, nearly all models perform better on in-distribution (ID) data than on out-of-distribution (OOD) data, except for CLIP, which performs equally well in both settings.

simple mean

Outputs alone

Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

null

Single cohesive phenomenon

No

null

Retrieval

null

['Another benchmark', 'Procedurally-generated', 'LLM-generated']

['Targeted']

['Short free response']

['Exact match']

['Widely-agreed']

['Yes']

['No comparison made']

['Yes']

['Constructed']

['Mean']

marchisioUnderstandingMitigatingLanguage2024

Understanding and Mitigating Language Confusion in LLMs

Include

null

The paper introduces a benchmark to measure language confusion in LLMs. They investigate language confusion on the line and word level in two practical settings: a) Monolingual generation, where a user queries the LLM in a given language, implicitly requesting an answer in the same language; and b) cross-lingual generation, where a user explicitly instructs a model to generate text in a different language.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Language Confusion

Yes

LLMs are often unable to consistently generate text in the user’s desired language, or the appropriate language given the context. They call this category of error “language confusion”.

Subset

null

They investigate language confusion on the line and word level in two practical settings: a) Monolingual generation, where a user queries the LLM in a given language, implicitly requesting an answer in the same language; and b) cross-lingual generation, where a user explicitly instructs a model to generate text in a different language.

A sentence (prompt)

null

Modified from another benchmark (e.g. translation into another language), For some part of the data they include human generated prompts

7100

Yes

Language of the prompt and the original data source

Random sample (creators defined a task space and sampled from it)

Free response (e.g. summary paragarph)

The paper introduces 2 new metrics for language confusion. Line-level pass rate (LPR) and Word-level pass rate (WPR).

null

Industry

Yes

null

Test

null

Simple Mean

No

null

https://github.com/for-ai/language-confusion

LCB

Contested

Yes

No

null

The benchmark is itself realistic

No

null

simple mean

Outputs alone

Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

null

Single cohesive phenomenon

Not applicable

null

Multilinguality

null

['Another benchmark', 'Author-crafted']

['Random']

['Free response']

['Exact match']

['Contested']

['Yes']

['Realistic']

['No']

['Constructed']

['Mean']

itoGeneralizationCapacityNeural2024

On the generalization capacity of neural networks during generic multimodal reasoning

Include

null

The paper introduces gCOG, a multimodal reasoning dataset designed to measure various types of OOD generalisation (distractor generalisation, systematic compositional, and productive compositional). The authors train various encoder architectures from scratch and compare their performances. Transformers can systematically generalise at scale, but no architectures can productively generalise.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Multimodal generalisation

Yes

"OOD generalization – the ability to perform tasks beyond the training distribution" (1)

Comprehensive

null

Models are given an 8x8 grid containing multicoloured letters at different indices, and must follow a binary tree of "if-then-else" instructions to answer a question like "Get the position of the orange 't'".

A query in natural language, an image of an 8x8 grid in some .jpg-like format, and a correct answer, which is either a shape ("d") a colour ("orange") or a location ((5, 4)).

The concrete dataset used for their evaluation is not provided, only a generator object in python is given.

Modified from another benchmark (e.g. translation into another language), Procedurally-generated task examples (e.g. Creating instances from a template)

null

Yes

task tree depth, num distractors

Random sample (creators defined a task space and sampled from it), Specific criteria (items were taken from a larger set based on specified rules)

Short free response (e.g. single word or number)

Exact Match (accuracy, F1, precision, recall)

null

Industry

Yes

null

Simple Mean

Yes

IID and OOD accuracy on varying numbers of distractors and tree depths

null

https://github.com/IBM/gcog

Generic COG (gCOG)

Contested

Yes

No

No comparisons made

No

Yes

"Identifying neural architectures that can robustly generalize OOD is a central goal in artificial intelligence. Compositional generalization benchmarks, which explicitly evaluate for generalization, provide a good testbed for measuring these capabilities" (9)

simple mean/sum

Outputs alone

Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

null

Composite phenomenon

Yes

null

Language Modelling

Adaptability

null

['Another benchmark', 'Procedurally-generated']

['Random', 'Criterion']

['Short free response']

['Exact match']

['Contested']

['Yes']

['No comparison made']

['Yes']

['Constructed']

['Mean']

liMultimodalArXivDataset2024

Multimodal ArXiv: A Dataset for Improving Scientific Comprehension of Large Vision-Language Models

Include

null

Multimodal ArXiv consists of ArXivCap, a figure-caption dataset sourced from scientific papers, and ArXivQA, a QA dataset generated by prompting GPT-4V for QA pairs on ArXivCap entries. Results show that fine-tuning on these datasets boosts performance on the MathVista benchmark, and that evaluation results for various scientific plot comprehension subtasks are poor.

null

Specific Application (A single use case, where the benchmark is likely to be examples of that use case)

comprehending scientific plots

No

null

Subset

The phenomenon is vaguely defined but the tasks are precisely defined

Vision-to-text subtasks: caption a single (or multiple) scientific figure(s), including an in-context learning subtask, and generate paper titles given figures and captions.

A ground truth paper title and a list of scientific figures and corresponding captions

null

Real task examples (e.g. GitHub issues), LLM-generated task examples (e.g. Filtered from responses to a prompt)

100,000

Yes

arXiv domain, arXiv DOI

Targeted items (creators defined a task space and chose tasks within it strategically)

Short free response (e.g. single word or number), Free response (e.g. summary paragarph)

n-gram (BLEU, ROUGE, chrF), LLM post-processing (extracting answers, reformatting for automated scoring)

null

Academia

Yes

null

Test

null

Simple Mean

No

null

https://huggingface.co/datasets/MMInstruction/ArxivQA; https://huggingface.co/datasets/MMInstruction/ArxivCap

Multimodal ArXiv

Not defined

Yes

null

Yes

No

The benchmark is itself realistic

Yes

"after training the model on QA pairs from each domain... Most domains hurt the Figure QA task. This suggests that synthetic Figure QA might not be the best benchmark for assessing realistic reasoning ability." (14373-4) "our Multimodal ArXiv dataset sources from ArXiv papers due to their accessibility and open-source licenses. This approach may overlook the diversity of disciplines and data modalities present in the broader scientific literature." (14378)

simple mean/sum

Outputs alone

Partial real task (e.g. answering medical questions collected from real people)

null

Composite phenomenon

Yes

null

VQA

Understanding

null

['Real task', 'LLM-generated']

['Targeted']

['Short free response', 'Free response']

['Soft match', 'LLM post-processing']

['No definition']

['Yes']

['']

['Realistic']

['Yes']

['Partial']

['Mean']

zouVGBenchEvaluatingLarge2024

VGBench: Evaluating Large Language Models on Vector Graphics Understanding and Generation

Include

null

The paper introduces VGBench, a comprehensive benchmark for vector graphics images that tests both visual understanding and generation. Formats like SVG, TikZ, and Graphviz are included, and performance is generally strong, though LLMs do worse with the lower-level SVG format.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

processing vector graphics

No

null

Comprehensive

null

For the QA task (VGQA), models are given a vector graphics representation (in textual format) and a multiple choice question about a high-level feature of the image, like the colour of a depicted entity. For the generation task (VGen), models must generate vector graphics code from a textual description.

For VGQA: a snippet of vector graphics code, a question with multiple choice answers, and a correct answer. For VGen: a textual description, the desired output format (e.g. SVG), and some ground truth vector graphics code.

null

Real task examples (e.g. GitHub issues), Modified from another benchmark (e.g. translation into another language), LLM-generated task examples (e.g. Filtered from responses to a prompt)

10,124

Yes

vector graphic format

Convenience sample (creators found a set of tasks that was readily accessible)

Multiple choice, Structured response (e.g. valid JSON, API call alone)

Exact Match (accuracy, F1, precision, recall), n-gram (BLEU, ROUGE, chrF), LLM post-processing (extracting answers, reformatting for automated scoring)

null

Academia

Yes

null

Test

4,279 examples in VGQA, 5,845 examples in VGen

null

Simple Mean

Yes

vector graphics format and question subtype (e.g. "Domain", "Layout", "Relation" questions)

null

https://huggingface.co/datasets/vgbench/VGen; https://huggingface.co/datasets/vgbench/VGQA

VGBench

Widely-agreed

Yes

No

No comparisons made

No

null

simple mean/sum

Outputs alone

Partial real task (e.g. answering medical questions collected from real people)

null

Composite phenomenon

Yes

null

Instruction Following

null

['Real task', 'Another benchmark', 'LLM-generated']

['Convenience']

['Multiple choice', 'Structured']

['Exact match', 'Soft match', 'LLM post-processing']

['Widely-agreed']

['Yes']

['No comparison made']

['No']

['Partial']

['Mean']

zhangXSemPLRCrosslingualSemantic2023

XSemPLR: Cross-Lingual Semantic Parsing in Multiple Natural Languages and Meaning Representations

Include

null

The paper introduces XSEMPLR, a unified benchmark for cross-lingual semantic parsing featuring 22 natural languages and 8 meaning representations by examining and selecting 9 existing datasets to cover 5 tasks and 164 domains. They use XSEMPLR to conduct a benchmark study on a wide range of multilingual language models, including encoder-based models (mBERT, XLM-R), encoder-decoder models (mBART, mT5), and decoder-based models (Codex, BLOOM). The findings show that large multilingual language models are still inadequate for performing CLSP tasks. They also find that the performance gap between monolingual training and cross-lingual transfer learning is still significant for multilingual models, though it can be mitigated by cross-lingual few-shot training.

null

Specific Application (A single use case, where the benchmark is likely to be examples of that use case)

cross-lingual semantic parsing

Yes

Cross-Lingual Semantic Parsing (CLSP) aims to translate queries in multiple natural languages (NLs) into meaning representations (MRs).

Comprehensive

null

The task is to train a model to convert a sentence in natural language into a meaning representation (e.g., SQL, programming code, Prolog, Functional Query Language, etc.).

A pair of input and output where input is a text in natural language and output is a text of input's meaning representation

null

Modified from another benchmark (e.g. translation into another language)

Train set: ~42k, test set: ~7500, Dev set: ~5500

No

null

Random sample (creators defined a task space and sampled from it)

Free response (e.g. summary paragarph)

Exact Match (accuracy, F1, precision, recall), n-gram (BLEU, ROUGE, chrF)

null

Mix (multiple authors from industry and academia)

Yes

null

Test, Train, Validation

null

Simple Mean

No

null

https://github.com/psunlpgroup/XSemPLR

XSEMPLR

Widely-agreed

Yes

No

No comparisons made

No

null

Simple mean

Outputs alone

Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

null

Single cohesive phenomenon

Not applicable

null

Multilinguality

null

['Another benchmark']

['Random']

['Free response']

['Exact match', 'Soft match']

['Widely-agreed']

['Yes']

['No comparison made']

['No']

['Constructed']

['Mean']

sunInformalLanguageProcessing2024

Toward Informal Language Processing: Knowledge of Slang in Large Language Models

Include

null

Using movie subtitles, the authors construct a dataset that supports evaluation on a diverse set of tasks pertaining to the automatic processing of slang. For both evaluation and finetuning, they show the effectiveness of their dataset on two core applications: 1) slang detection, and 2) identification of regional and historical sources of slang from natural sentences.

null

Specific Application (A single use case, where the benchmark is likely to be examples of that use case)

informal language processing (Knowledge of slang in LLMs)

No

They focus on two core tasks for informal language processing. First, they evaluate the extent to which LLMs can reliably detect slang usages in natural sentences. Second, they assess whether LLMs can be used to identify regional-historical sources of slang via a text classification task.

Subset

null

Task1: Given a set of sentences, they evaluate slang detection at both sentence-level and word-level. Task2: Given a sentence containing a slang usage, they ask the model to classify its source (e.g. region and age).

a sentence of natural language

null

Crowd-sourced task examples (e.g. Prolific-created tasks)

25,000

Yes

Annotator confidence, Movie ID, Region, Year

Random sample (creators defined a task space and sampled from it)

Multiple choice

Exact Match (accuracy, F1, precision, recall), They also report two metrics to compare an LLM’s predictive confidence in slang usages relative to their literal counterparts.

null

The benchmark is build on top of OpenSubtitles corpus.

Mix (multiple authors from industry and academia)

Yes

null

Test, Train

null

Simple Mean

No

null

https://github.com/amazon-science/slang-llm-benchmark

null

Contested

Yes

No

null

simple mean

Outputs alone

Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

null

Composite phenomenon

Yes

null

Multilinguality

null

['Crowd-sourced']

['Random']

['Multiple choice']

['Exact match', 'Correlation']

['Contested']

['Yes']

['No comparison made']

['No']

['Constructed']

['Mean']

wangPretrainingLanguageModel2023

ON PRE-TRAINED LANGUAGE MODELS FOR ANTIBODY

Include

null

This paper introduces the AnTibody Understanding Evaluation (ATUE) benchmark to systematically assess the representation capabilities of general and antibody-specific pre-trained language models across a range of antibody-related tasks. It also explores how incorporating biological mechanisms into pre-training can enhance model performance and evaluates the transferability of learned representations to real-world applications such as drug discovery and immune system analysis.

null

Specific Application (A single use case, where the benchmark is likely to be examples of that use case)

LLMs capability to do antibody representation learning and biological reasoning with sequence specificity

Yes

how LLMs perform in antibody tasks with different specificity and how introducing specific biological mechanisms to the pre-training process can benefit the model. Additionally, authors evaluate if the learned antibody pre-trained representations can be applied to real-world antibody problems, like drug discovery and immune process understanding.

Subset

null

Evaluate the ability of pre-trained language models to perform on four supervised antibody-related prediction tasks—antigen binding, paratope prediction, B cell maturation classification, and SARS-CoV-2 antibody discovery—each varying in antibody specificity. These tasks assess whether the models can capture biologically meaningful information from antibody sequences.

N/A there are four tasks

null

Real task examples (e.g. GitHub issues)

3242, 1662, 88094, 22000

No

null

Convenience sample (creators found a set of tasks that was readily accessible), Targeted items (creators defined a task space and chose tasks within it strategically), Specific criteria (items were taken from a larger set based on specified rules)

Structured response (e.g. valid JSON, API call alone)

Exact Match (accuracy, F1, precision, recall), Matthews Correlation Coefficient (MCC), and AUC (Area Under the ROC Curve)

null

Mix (multiple authors from industry and academia)

Yes

null

Test, Train, Validation

15,128/3,242 , N/A

null

Simple Mean

No

null

https://github.com/dqwang122/EATLM

ATUE

Contested

Yes

No

No comparisons made

No

null

Outputs alone

Partial real task (e.g. answering medical questions collected from real people)

null

Composite phenomenon

Yes

null

Biology

null

['Real task']

['Convenience', 'Targeted', 'Criterion']

['Structured']

['Exact match', 'Correlation']

['Contested']

['Yes']

['No comparison made']

['No']

['Partial']

null

bajpaiCanLLMsReplace2024

Can LLMs replace Neil deGrasse Tyson? Evaluating the Reliability of LLMs as Science Communicators

Include

null

This paper focuses on evaluating the reliability of current LLMs as science communicators. They introduce a dataset, SCiPS-QA, comprising 742 Yes/No queries embedded in complex scientific concepts, along with a benchmarking suite that evaluates LLMs for correctness and consistency across various criteria. They also benchmark three proprietary LLMs from the OpenAI GPT family and 13 open-access LLMs from the Meta Llama-2, Llama-3, and Mistral families.

null

Specific Application (A single use case, where the benchmark is likely to be examples of that use case)

Reliability of LLMs as Science Communicators

No

Can existing LLMs answer scientific reasoning questions successfully and faithfully that require understanding the nuances of scientific knowledge?

Comprehensive

null

A binary (yes/No) classification task where the model is asked to answer a scientific question.

A question in science

null

Not explained

742

Yes

topic, date

Unknown

Multiple choice

Exact Match (accuracy, F1, precision, recall)

null

Academia

Yes

null

Test

null

Simple Mean

No

null

https://github.com/Prasoon1207/llm-science-miscommunication/blob/main/data/data.csv

SCiPS-QA

Not defined

Yes

No

No comparisons made

No

Yes

No

null

Simple mean and standard deviation

Outputs alone

Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

null

Single cohesive phenomenon

Not applicable

null

General Science

null

['Unknown']

['Multiple choice']

['Exact match']

['No definition']

['Yes']

['No comparison made']

['No']

['Constructed']

['Mean', 'Std']

hauserLargeLanguageModelsExpertlevel2024

Large Language Models' Expert-level Global History Knowledge Benchmark (HiST-LLM)

Include

null

The paper introduces the History Seshat Test for LLMs (HiST-LLM), based on a subset of the Seshat Global History Databank, which provides a structured representation of human historical knowledge, containing 36,000 data points across 600 historical societies and over 2,700 scholarly references. Using this dataset, they benchmark a total of seven models from the Gemini, OpenAI, and Llama families.

null

Specific Application (A single use case, where the benchmark is likely to be examples of that use case)

LLM's Expert-level Global History Knowledge

No

The ability of the model to answer expert-level history questions.

Comprehensive

null

The ask is to ask the model a multiple-choice question about history.

A multiple-choice question

null

Human expert created the examples

36000

No

null

Random sample (creators defined a task space and sampled from it)

Multiple choice

Exact Match (accuracy, F1, precision, recall)

null

Academia

Yes

null

Test

null

Simple Mean

No

null

https://github.com/seshat-db/HiST-LLM

HiST-LLM

Contested

Yes

No

No comparisons made

No

null

Mean and standard deviation

Outputs alone

Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

null

Single cohesive phenomenon

Not applicable

null

History

null

['Expert-crafted']

['Random']

['Multiple choice']

['Exact match']

['Contested']

['Yes']

['No comparison made']

['No']

['Constructed']

['Mean', 'Std']

sadatMSciNLIDiverseBenchmark2024

MSciNLI: A Diverse Benchmark for Scientific Natural Language Inference

Include

null

This paper introduces MSCINLI, a new dataset comprising 132,320 sentence pairs from five diverse scientific domains to enhance the study of scientific Natural Language Inference (NLI). Baseline models, including fine-tuned and prompted LLMs, reveal the dataset's challenging nature, as well as performance degradation due to domain shifts, highlighting the unique characteristics of each domain. Additionally, employing both scientific NLI datasets in intermediate task transfer learning showcases improvements in downstream scientific tasks.

null

Specific Application (A single use case, where the benchmark is likely to be examples of that use case)

Natural language inference (semantic relationship between two sentences), scientific domains

Yes

predicting the semantic relation between two sentences extracted from research articles

Comprehensive

null

sentence pairs, multiple choice on semantic relation between sentences

null

question, prompt, domain, class, difficulty, response correct/score

Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions), Modified from another benchmark (e.g. translation into another language)

127,320

Yes

difficulty, domain

Targeted items (creators defined a task space and chose tasks within it strategically), Specific criteria (items were taken from a larger set based on specified rules)

Multiple choice

Exact Match (accuracy, F1, precision, recall)

null

Academia

Yes

null

Test, Train

null

Simple Mean

Yes

difficulty

null

GitHub, huggingface

MSciNLI

Widely-agreed

Yes

No

Yes

No

null

mean and variance, t-tests

Outputs alone

Representative task (e.g. answering medical licensing exam questions)

null

Single cohesive phenomenon

Not applicable

null

General Science

null

['Author-crafted', 'Another benchmark']

['Targeted', 'Criterion']

['Multiple choice']

['Exact match']

['Widely-agreed']

['Yes']

['Comparison made']

['No']

['Representative']

['Mean', 'Std', 'Tests']

dengNewTermBenchmarkingRealtime2024

NewTerm: Benchmarking Real-Time New Terms for Large Language Models with Annual Updates

Include

null

This paper introduces NewTerm, an adaptive benchmark designed for the real-time evaluation of new terms in large language models (LLMs) to address their struggle with real-time information due to knowledge cutoffs. The benchmark is constructed using a highly automated method allowing flexible and minimal human effort updates, revealing a performance reduction of over 20% on various LLMs with new terms and highlighting difficulties in generalizing to distant new terms. Annual updates to NewTerm, starting with 2022 and 2023, are planned to continuously assess and analyze the evolving challenge of new terms in LLMs.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

Updating of knowledge, real-time evaluation of new terms introduced after knowledge cutoff

Yes

flexible updates for real-time information

Comprehensive

null

Answer questions about new terms from dictionary, introduced after knowledge cutoff

Question, multiple choice answers, response, correct

null

Real task examples (e.g. GitHub issues), Procedurally-generated task examples (e.g. Creating instances from a template)

null

Yes

Domains: The Choice of Multiple Alter (COMA), The Choice of Similar Terms (COST), Common Sense Judgement (CSJ)

Specific criteria (items were taken from a larger set based on specified rules)

Multiple choice

Exact Match (accuracy, F1, precision, recall)

null

Academia

Yes

null

Test

null

Simple Mean

Yes

Domains: The Choice of Multiple Alter (COMA), The Choice of Similar Terms (COST), Common Sense Judgement (CSJ)

null

GitHub

NewTerm

Widely-agreed

Yes

No

The benchmark is itself realistic

No

null

simple mean/sum

Outputs alone

Representative task (e.g. answering medical licensing exam questions)

null

Single cohesive phenomenon

Not applicable

null

Language Modelling

Updating

null

['Real task', 'Procedurally-generated']

['Criterion']

['Multiple choice']

['Exact match']

['Widely-agreed']

['Yes']

['Realistic']

['No']

['Representative']

['Mean']

yeRoTBenchMultilevelBenchmark2024

RoTBench: A Multi-Level Benchmark for Evaluating the Robustness of Large Language Models in Tool Learning

Include

null

LLMs are increasingly deployedin settings where they can use tools, e.g. call functions to retrieve real-time information on weather. This paper proposes benchmark measuring the robustness of LLMs in selecting tools when these are specified under noise (e.g. the function name is perturbed).

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

tool use when tool names or arguments are mislabeled

No

LLMs should exhibit consistent tool use when tools or their arguments are mislabeled.

Subset

null

Prompt + List of availabe tools + ground truth tool + ground truth arguments

null

Procedurally-generated task examples (e.g. Creating instances from a template)

735

No

null

Random sample (creators defined a task space and sampled from it), Convenience sample (creators found a set of tasks that was readily accessible)

Free response (e.g. summary paragarph), Structured response (e.g. valid JSON, API call alone)

Exact Match (accuracy, F1, precision, recall)

null

existing benchmark + small perturbations

Academia

Yes

null

A) The noise induced in the benchmark significantly alters the *expected behaviour* of the model. For instance, imagine "Get_GPS_COORDINATES : This tool is used for fetching information weather for specified location." is a perturbation of "Get_WEATHER: This tool is used for fetching infromation weather for specified location." Clearly, the inconsistent information provided to the LLM between the function name and its docstring changes the expected behaviour of the model and hence "consistent" behaviour is not necessarily a sign of robustness. This casts doubt on the construct validity of “Robust Tool Use”. A positive note: The authors test human perofrmance and humans get scores between 69% and 89%, showing the task is still somewhat possible to humans. B) The authors built their dataset by perturbing an existing dataset. their explanations of the existing dataset are negligle. It should be best practice to at least explain what the task of the original dataset is exactly, its size and limitations.

Test, Train

null

Simple Mean

Yes

different intermediate stages to a full sucess.

null

https://github.com/Junjie-Ye/RoTBench

RoTBench

Contested

Yes

No

No comparisons made

No

Yes

No

null

Outputs alone

Constructed task (e.g. predicting medical diagnoses from clinicians' notes)

null

Composite phenomenon

Yes

null

Agents

Tool Use

null

['Procedurally-generated']

['Random', 'Convenience']

['Free response', 'Structured']

['Exact match']

['Contested']

['Yes']

['No comparison made']

['No']

['Constructed']

null

maMMLONGBENCHDOCBenchmarkingLongcontext2024

MMLONGBENCH-DOC: Benchmarking Long-context Document Understanding with Visualizations

Include

null

The paper presents a long-context multimodal benchmark dataset of more than 1k expert annotated questions over long PDFs which require aggregating evidence across multiple locations and evidence formats (text, image, charts, etc.) to answer. MMLongBench-Doc presents a challenge for strong models such as GPT-4o and other large vision language models (LVLMs), demonstrating the need for improved long-context LVLM capabilities.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

long-context document understanding

Yes

"the automatic understanding of [long-context] documents. The understanding of these lengthy documents brings new challenges for LVLMs", including localization and cross-page comprehension

Comprehensive

null

Give a document to a model and have it answer a question regarding information in the document.

Documents are PDF files. Questions are stored in json format with the following attributes: document ID, document type, question, answer, evidence pages, evidence sources, and answer format.

null

Author-crafted task examples (e.g. hand-written examples, manual transformation of existing data into questions), Modified from another benchmark (e.g. translation into another language), Procedurally-generated task examples (e.g. Creating instances from a template)

1082

Yes

evidence source, answer format, question length statistics, answer length statistics, document length statistics

Targeted items (creators defined a task space and chose tasks within it strategically), Specific criteria (items were taken from a larger set based on specified rules)

Short free response (e.g. single word or number)

Exact Match (accuracy, F1, precision, recall)

null

Mix (multiple authors from industry and academia)

Yes

null

Test

null

Simple Mean

Yes

type of evidence source, number of evidence pages involved in answering the question, document type

null

https://github.com/mayubo2333/MMLongBench-Doc

MMLongBench-Doc

Widely-agreed

Yes

No

No comparisons made

No

null

Outputs alone

Partial real task (e.g. answering medical questions collected from real people)

null

Single cohesive phenomenon

Not applicable

null

NLP

Long Context

null

['Author-crafted', 'Another benchmark', 'Procedurally-generated']

['Targeted', 'Criterion']

['Short free response']

['Exact match']

['Widely-agreed']

['Yes']

['No comparison made']

['No']

['Partial']

null

kuratovBABILongTestingLimits2024

BABILong: Testing the Limits of LLMs with Long Context Reasoning-in-a-Haystack

Include

null

The BABILong benchmark tests language models’ ability to reason across facts distributed in extremely long documents in the reasoning setting, scattering relevant facts among less relevant natural text. The paper finds LLMs only effectively use less than 20% of the context in such settings, with reasoning complexity negatively impacting performance. Multiple methods including in-context reasoning, retrieval augmented generation, and context extension are applied to profile model capabilities in these long-context tasks.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

language models’ ability to reason across facts distributed in extremely long documents

Yes

"language models’ ability to reason across facts distributed in extremely long documents"

Comprehensive

null

Perform one of 20 reasoning tasks (e.g., fact chaining, simple induction, deduction, counting, and handling lists/sets), generally presented in question format, given a long context with relevant and distracting articles.

A long-context input text, question, and the question's answer based on the input

null

Modified from another benchmark (e.g. translation into another language), Procedurally-generated task examples (e.g. Creating instances from a template)

null

Yes

facts per task, relevant facts per task, reasoning task type

Targeted items (creators defined a task space and chose tasks within it strategically), Specific criteria (items were taken from a larger set based on specified rules)

Short free response (e.g. single word or number)

Exact Match (accuracy, F1, precision, recall)

null

Mix (multiple authors from industry and academia)

Yes

null

Test

null

Simple Mean

No

input length, task type, context size

null

https://github.com/booydar/babilong

BABILong

Widely-agreed

Yes

No

Yes

Advantages of the benchmark are compared versus existing related benchmarks based on design and correlation study, and the content of the benchmark and the relation between model performance and capability are analyzed.

simple mean

Outputs alone

Partial real task (e.g. answering medical questions collected from real people)

null

Single cohesive phenomenon

Not applicable

null

NLP

Long Context

null

['Another benchmark', 'Procedurally-generated']

['Targeted', 'Criterion']

['Short free response']

['Exact match']

['Widely-agreed']

['Yes']

['No comparison made']

['Yes']

['Partial']

['Mean']

wangAdaLEvalEvaluatingLongcontext2024

Ada-LEval: Evaluating long-context LLMs with length-adaptable benchmarks

Include

null

Ada-LEval presents a length-adaptable benchmark for long-context understanding capabilities of LLMs, involving challenging questions for reliable evaluation and context lengths extending to the ultra-long setting. SOTA open and closed models are evaluated to demonstrate current limitations of LLMs in such settings.

null

General Capability (A broadly useful ability, which could be relevant to multiple applications)

null

No

Context window is a notable factor in LLM performance and is critical to handling long texts. The effectiveness of LLMs in managing long text is still open for exploration and assessment.

Comprehensive

null

1. Take in a long text and arrange the text segments in the correct order. 2. Choose the best answer from multiple candidate answers to a question based on a given long text.

Not provided, but generally the task samples consist of either a question and many sample answers, or a series of texts to be rearranged (per the task definition).

null

Real task examples (e.g. GitHub issues), Modified from another benchmark (e.g. translation into another language), Procedurally-generated task examples (e.g. Creating instances from a template)

over 80k

Yes

total samples per context length, max tokens, average number of tokens

Targeted items (creators defined a task space and chose tasks within it strategically), Specific criteria (items were taken from a larger set based on specified rules)

Multiple choice, Free response (e.g. summary paragarph)

Exact Match (accuracy, F1, precision, recall), Distribution (perplexity, calibration, correlation), instruction following rate

null

Mix (multiple authors from industry and academia)

Yes

null

Test

null

Simple Mean

Yes

context lengths from 2k to 16k

null

https://github.com/open-compass/Ada-LEval

Ada-LEval

Widely-agreed

Yes

No

Yes

Comparison with traditional long-context benchmarks such as GovReport demonstrate Ada-LEval requires more overall text understanding to complete.

simple mean

Outputs alone

Partial real task (e.g. answering medical questions collected from real people)

null

Single cohesive phenomenon

Not applicable

null

NLP

Long Context

null

['Real task', 'Another benchmark', 'Procedurally-generated']

['Targeted', 'Criterion']

['Multiple choice', 'Free response']

['Exact match', 'Distribution', 'Exact match']

['Widely-agreed']

['Yes']

['No comparison made']

['Yes']

['Partial']

['Mean']