metadata

tags:
  - sentence-transformers
  - sparse-encoder
  - sparse
  - splade
  - generated_from_trainer
  - dataset_size:1200000
  - loss:SpladeLoss
  - loss:SparseMarginMSELoss
  - loss:FlopsLoss
base_model:
  - prajjwal1/bert-tiny
widget:
  - text: >-
      Most Referenced:report - Return to the USDOJ/OIG Home Page - US Department
      of JusticeReturn to the USDOJ/OIG Home Page - US Department of Justice.
      Opinion:Roberts: Feds to stop using private prisons.
  - text: >-
      Paul O'Neill, the founder of the Trans-Siberian Orchestra (pictured) has
      died at age 61. Paul O'Neill, the founder of the popular Christmas-themed
      rock ensemble Trans-Siberian Orchestra has died. A statement on the
      group's Facebook page reads: The entire Trans-Siberian Orchestra family,
      past and present, is heartbroken to share the devastating news that Paul
      O’Neill has passed away from chronic illness.
  - text: meaning for concern
  - text: >-
      Additional Tips. 1  Do not rub the ink stains as it can spread the stains
      further. 2  Make sure you test the cleaning solution on a small, hidden
      area to check if it is suitable for the material. 3  In case an ink stain
      has become old and dried, the above mentioned home remedies may not be
      effective.arpet: For ink stained spots on a carpet, you may apply a paste
      of cornstarch and milk. Leave it for a few hours before brushing it off.
      Finally, clean the residue with a vacuum cleaner. Leather: Try using a
      leather shampoo or a leather ink remover for removing ink stains from
      leather items.
  - text: >-
      See below: 1. Get your marriage license. Before you can change your name,
      you'll need the original (or certified) marriage license with the raised
      seal and your new last name on it. Call the clerk's office where your
      license was filed to get copies if one wasn't automatically sent to you.
      2. Change your Social Security card.
pipeline_tag: feature-extraction
library_name: sentence-transformers
metrics:
  - dot_accuracy@1
  - dot_accuracy@3
  - dot_accuracy@5
  - dot_accuracy@10
  - dot_precision@1
  - dot_precision@3
  - dot_precision@5
  - dot_precision@10
  - dot_recall@1
  - dot_recall@3
  - dot_recall@5
  - dot_recall@10
  - dot_ndcg@10
  - dot_mrr@10
  - dot_map@100
  - query_active_dims
  - query_sparsity_ratio
  - corpus_active_dims
  - corpus_sparsity_ratio
model-index:
  - name: SPLADE Sparse Encoder
    results:
      - task:
          type: sparse-information-retrieval
          name: Sparse Information Retrieval
        dataset:
          name: Unknown
          type: unknown
        metrics:
          - type: dot_accuracy@1
            value: 0.4772
            name: Dot Accuracy@1
          - type: dot_accuracy@3
            value: 0.793
            name: Dot Accuracy@3
          - type: dot_accuracy@5
            value: 0.8964
            name: Dot Accuracy@5
          - type: dot_accuracy@10
            value: 0.96
            name: Dot Accuracy@10
          - type: dot_precision@1
            value: 0.4772
            name: Dot Precision@1
          - type: dot_precision@3
            value: 0.2713333333333333
            name: Dot Precision@3
          - type: dot_precision@5
            value: 0.18644000000000002
            name: Dot Precision@5
          - type: dot_precision@10
            value: 0.10094000000000002
            name: Dot Precision@10
          - type: dot_recall@1
            value: 0.4616666666666666
            name: Dot Recall@1
          - type: dot_recall@3
            value: 0.7798833333333334
            name: Dot Recall@3
          - type: dot_recall@5
            value: 0.8874
            name: Dot Recall@5
          - type: dot_recall@10
            value: 0.95595
            name: Dot Recall@10
          - type: dot_ndcg@10
            value: 0.721747648718731
            name: Dot Ndcg@10
          - type: dot_mrr@10
            value: 0.6489996031746051
            name: Dot Mrr@10
          - type: dot_map@100
            value: 0.6446961471449598
            name: Dot Map@100
          - type: query_active_dims
            value: 18.334199905395508
            name: Query Active Dims
          - type: query_sparsity_ratio
            value: 0.9993993119747921
            name: Query Sparsity Ratio
          - type: corpus_active_dims
            value: 121.65303042911474
            name: Corpus Active Dims
          - type: corpus_sparsity_ratio
            value: 0.9960142510179831
            name: Corpus Sparsity Ratio
datasets:
  - microsoft/ms_marco
language:
  - en

SPLADE Sparse Encoder

This is a SPLADE sparse retrieval model based on BERT-Tiny (4M) that was trained by distilling a Cross-Encoder on the MSMARCO dataset. The cross-encoder used was ms-marco-MiniLM-L6-v2.

This Tiny SPLADE model beats BM25 by 65.6% on the MSMARCO benchmark. While this model is 15x smaller than Naver's official splade-v3-distilbert, is posesses 80% of it's performance on MSMARCO. This model is small enough to be used without a GPU on a dataset of a few thousand documents.

Collection: https://huggingface.co/collections/rasyosef/splade-tiny-msmarco-687c548c0691d95babf65b70
Distillation Dataset: https://huggingface.co/datasets/yosefw/msmarco-train-distil-v2
Code: https://github.com/rasyosef/splade-tiny-msmarco

Performance

The splade models were evaluated on 55 thousand queries and 8.84 million documents from the MSMARCO dataset.

	Size (# Params)	MRR@10 (MS MARCO dev)
`BM25`	-	18.0
`rasyosef/splade-tiny`	4.4M	30.9
`rasyosef/splade-mini`	11.2M	33.2
`naver/splade-v3-distilbert`	67.0M	38.7

Usage

Direct Usage (Sentence Transformers)

First install the Sentence Transformers library:

pip install -U sentence-transformers

Then you can load this model and run inference.

from sentence_transformers import SparseEncoder

# Download from the 🤗 Hub
model = SparseEncoder("rasyosef/splade-tiny")
# Run inference
queries = [
    "what do i need to change my name on my license in ma",
]
documents = [
    'Change your name on MA state-issued ID such as driver’s license or MA ID card. All documents you bring to RMV need to be originals or certified copies by the issuing agency. PAPERWORK NEEDED: Proof of legal name change — A court order showing your legal name change. Your Social Security Card with your new legal name change',
    "See below: 1. Get your marriage license. Before you can change your name, you'll need the original (or certified) marriage license with the raised seal and your new last name on it. Call the clerk's office where your license was filed to get copies if one wasn't automatically sent to you. 2. Change your Social Security card.",
    "You'll keep the same number—just your name will be different. Mail in your application to the local Social Security Administration office. You should get your new card within 10 business days. 3. Change your license at the DMV. Take a trip to the local Department of Motor Vehicles office to get a new license with your new last name. Bring every form of identification you can get your hands on—your old license, your certified marriage certificate and, most importantly, your new Social Security card.",
]
query_embeddings = model.encode_query(queries)
document_embeddings = model.encode_document(documents)
print(query_embeddings.shape, document_embeddings.shape)
# [1, 30522] [3, 30522]

# Get the similarity scores for the embeddings
similarities = model.similarity(query_embeddings, document_embeddings)
print(similarities)
# tensor([[16.6297, 13.4552, 10.1923]])

Model Details

Model Description

Model Type: SPLADE Sparse Encoder
Base model: prajjwal1/bert-tiny
Maximum Sequence Length: 512 tokens
Output Dimensionality: 30522 dimensions
Similarity Function: Dot Product

Model Sources

Documentation: Sentence Transformers Documentation
Documentation: Sparse Encoder Documentation
Repository: Sentence Transformers on GitHub
Hugging Face: Sparse Encoders on Hugging Face

Full Model Architecture

SparseEncoder(
  (0): MLMTransformer({'max_seq_length': 512, 'do_lower_case': False, 'architecture': 'BertForMaskedLM'})
  (1): SpladePooling({'pooling_strategy': 'max', 'activation_function': 'relu', 'word_embedding_dimension': 30522})
)

Click to expand

Evaluation

Metrics

Sparse Information Retrieval

Evaluated with SparseInformationRetrievalEvaluator

Metric	Value
dot_accuracy@1	0.4772
dot_accuracy@3	0.793
dot_accuracy@5	0.8964
dot_accuracy@10	0.96
dot_precision@1	0.4772
dot_precision@3	0.2713
dot_precision@5	0.1864
dot_precision@10	0.1009
dot_recall@1	0.4617
dot_recall@3	0.7799
dot_recall@5	0.8874
dot_recall@10	0.9559
dot_ndcg@10	0.7217
dot_mrr@10	0.649
dot_map@100	0.6447
query_active_dims	18.3342
query_sparsity_ratio	0.9994
corpus_active_dims	121.653
corpus_sparsity_ratio	0.996

Training Details

Training Dataset

Unnamed Dataset

Size: 1,200,000 training samples
Columns: query, positive, negative_1, negative_2, and label

Approximate statistics based on the first 1000 samples:

	query	positive	negative_1	negative_2	label
type	string	string	string	string	list
details	min: 4 tokens mean: 9.08 tokens max: 35 tokens	min: 23 tokens mean: 79.02 tokens max: 192 tokens	min: 18 tokens mean: 78.24 tokens max: 230 tokens	min: 13 tokens mean: 75.26 tokens max: 230 tokens	size: 2 elements

Samples:

query	positive	negative_1	negative_2	label
`does alzheimer's affect sleep`	`People with Alzheimer’s disease go through many changes, and sleep problems are often some of the most noticeable. Most adults have changes in their sleep patterns as they age. But the problems are more severe and happen more often for people with Alzheimer’s.`	`Could the position you SLEEP in affect your risk of Alzheimer's? People who sleep on their side enable their brain to 'detox' better while they rest. While asleep, brain is hard at work removing toxins that build up in the day. If left to build up, these toxins can cause Alzheimer's and Parkinson's.`	`The Scary Connection Between Snoring and Dementia. For more, visit TIME Health. If you don't snore, you likely know someone who does. Between 19% and 40% of adults snore when they sleep, and that percentage climbs even higher, particularly for men, as we age.`	`[1.407266616821289, 10.169305801391602]`
`what is fy in steel design`	Since the yield strength of the steel is quite clearly defined and controlled, this establishes a very precise reference in structural investigations. An early design decision is that for the yield strength (specified by the Grade of steel used) that is to be used in the design work.Several different grades of steel may be used for large projects, with a minimum grade for ordinary tasks and higher grades for more demanding ones.ost steel used for reinforcement is highly ductile in nature. Its usable strength is its yield strength, as this stress condition initiates such a magnitude of deformation (into the plastic yielding range of the steel), that major cracking will occur in the concrete.	`fy is the yield point of the material. E is the symbol for Young's Modulus of the material. E can be measured by dividing the elastic stress by the elastic strain.That is, this measurement must be made before the yield point of the material is reached.y is the yield point of the material. E is the symbol for Young's Modulus of the material. E can be measured by dividing the elastic stress by the elastic strain.`	The longest dimension of the cant. WT is 13'. Using ASTM A992 carbon steel, a WT9x35.5 is at full bending stress and deflection limits. (Fy = 50 ksi). The only information I've found about using stainless for structural design is that type 304 is usually used.This yield strength (Fy) is only equal to 39 or 42ksi.sing ASTM A992 carbon steel, a WT9x35.5 is at full bending stress and deflection limits. (Fy = 50 ksi). The only information I've found about using stainless for structural design is that type 304 is usually used.	`[0.5, 0.5]`
`most common nutritional deficiencies for teenagers`	: Appendix B: Vitamin and Mineral Deficiencies in the U.S. Some American adults get too little vitamin D, vitamin E, magnesium, calcium, vitamin A and vitamin C (Table B1). More than 40 percent of adults have dietary intakes of vitamin A, C, D and E, calcium and magnesium below the average requirement for their age and gender. Inadequate intake of vitamins and minerals is most common among 14-to-18-year-old teenagers. Adolescent girls have lower nutrient intake than boys (Berner 2014; Fulgoni 2011). But nutrient deficiencies are rare among younger American children; the exceptions are dietary vitamin D and E, for which intake is low for all Americans, and calcium.	`Common Nutritional Deficiencies. 10 Most Common Nutritional Deficiencies.. Calcium. Calcium is one of the most abundant minerals in your body, yet most people still manage to have a calcium deficiency. Calcium is best know for adding strength to your bones and teeth.`	`1) Vitamin D–Vitamin D deficiency is common in infants born to mothers with low levels of Vitamin D. Severe deficiency of this nutrient in infancy and early childhood can lead to the development of Rickets, a disease that affects bone formation and causes bow-legs.`	`[3.182860851287842, 7.834665775299072]`

Loss: SpladeLoss with these parameters:

{
    "loss": "SparseMarginMSELoss",
    "document_regularizer_weight": 0.2,
    "query_regularizer_weight": 0.3
}

Training Hyperparameters

Non-Default Hyperparameters

eval_strategy: epoch
per_device_train_batch_size: 32
per_device_eval_batch_size: 32
num_train_epochs: 5
lr_scheduler_type: cosine
warmup_ratio: 0.05
fp16: True
load_best_model_at_end: True
optim: adamw_torch_fused
push_to_hub: True

All Hyperparameters

Click to expand

overwrite_output_dir: False
do_predict: False
eval_strategy: epoch
prediction_loss_only: True
per_device_train_batch_size: 32
per_device_eval_batch_size: 32
per_gpu_train_batch_size: None
per_gpu_eval_batch_size: None
gradient_accumulation_steps: 1
eval_accumulation_steps: None
torch_empty_cache_steps: None
learning_rate: 5e-05
weight_decay: 0.0
adam_beta1: 0.9
adam_beta2: 0.999
adam_epsilon: 1e-08
max_grad_norm: 1.0
num_train_epochs: 5
max_steps: -1
lr_scheduler_type: cosine
lr_scheduler_kwargs: {}
warmup_ratio: 0.05
warmup_steps: 0
log_level: passive
log_level_replica: warning
log_on_each_node: True
logging_nan_inf_filter: True
save_safetensors: True
save_on_each_node: False
save_only_model: False
restore_callback_states_from_checkpoint: False
no_cuda: False
use_cpu: False
use_mps_device: False
seed: 42
data_seed: None
jit_mode_eval: False
use_ipex: False
bf16: False
fp16: True
fp16_opt_level: O1
half_precision_backend: auto
bf16_full_eval: False
fp16_full_eval: False
tf32: None
local_rank: 0
ddp_backend: None
tpu_num_cores: None
tpu_metrics_debug: False
debug: []
dataloader_drop_last: False
dataloader_num_workers: 0
dataloader_prefetch_factor: None
past_index: -1
disable_tqdm: False
remove_unused_columns: True
label_names: None
load_best_model_at_end: True
ignore_data_skip: False
fsdp: []
fsdp_min_num_params: 0
fsdp_config: {'min_num_params': 0, 'xla': False, 'xla_fsdp_v2': False, 'xla_fsdp_grad_ckpt': False}
fsdp_transformer_layer_cls_to_wrap: None
accelerator_config: {'split_batches': False, 'dispatch_batches': None, 'even_batches': True, 'use_seedable_sampler': True, 'non_blocking': False, 'gradient_accumulation_kwargs': None}
deepspeed: None
label_smoothing_factor: 0.0
optim: adamw_torch_fused
optim_args: None
adafactor: False
group_by_length: False
length_column_name: length
ddp_find_unused_parameters: None
ddp_bucket_cap_mb: None
ddp_broadcast_buffers: False
dataloader_pin_memory: True
dataloader_persistent_workers: False
skip_memory_metrics: True
use_legacy_prediction_loop: False
push_to_hub: True
resume_from_checkpoint: None
hub_model_id: None
hub_strategy: every_save
hub_private_repo: None
hub_always_push: False
hub_revision: None
gradient_checkpointing: False
gradient_checkpointing_kwargs: None
include_inputs_for_metrics: False
include_for_metrics: []
eval_do_concat_batches: True
fp16_backend: auto
push_to_hub_model_id: None
push_to_hub_organization: None
mp_parameters:
auto_find_batch_size: False
full_determinism: False
torchdynamo: None
ray_scope: last
ddp_timeout: 1800
torch_compile: False
torch_compile_backend: None
torch_compile_mode: None
include_tokens_per_second: False
include_num_input_tokens_seen: False
neftune_noise_alpha: None
optim_target_modules: None
batch_eval_metrics: False
eval_on_start: False
use_liger_kernel: False
liger_kernel_config: None
eval_use_gather_object: False
average_tokens_across_devices: False
prompts: None
batch_sampler: batch_sampler
multi_dataset_batch_sampler: proportional
router_mapping: {}
learning_rate_mapping: {}

Training Logs

Epoch	Step	Training Loss	dot_ndcg@10
1.0	37500	11.4095	0.7103
2.0	75000	10.5305	0.7139
3.0	112500	9.5368	0.7197
4.0	150000	8.717	0.7216
5.0	187500	8.3094	0.7217

The bold row denotes the saved checkpoint.

Framework Versions

Python: 3.11.13
Sentence Transformers: 5.0.0
Transformers: 4.54.0
PyTorch: 2.6.0+cu124
Accelerate: 1.8.1
Datasets: 4.0.0
Tokenizers: 0.21.2

Citation

BibTeX

Sentence Transformers

@inproceedings{reimers-2019-sentence-bert,
    title = "Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks",
    author = "Reimers, Nils and Gurevych, Iryna",
    booktitle = "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing",
    month = "11",
    year = "2019",
    publisher = "Association for Computational Linguistics",
    url = "https://arxiv.org/abs/1908.10084",
}

SpladeLoss

@misc{formal2022distillationhardnegativesampling,
      title={From Distillation to Hard Negative Sampling: Making Sparse Neural IR Models More Effective},
      author={Thibault Formal and Carlos Lassance and Benjamin Piwowarski and Stéphane Clinchant},
      year={2022},
      eprint={2205.04733},
      archivePrefix={arXiv},
      primaryClass={cs.IR},
      url={https://arxiv.org/abs/2205.04733},
}

SparseMarginMSELoss

@misc{hofstätter2021improving,
    title={Improving Efficient Neural Ranking Models with Cross-Architecture Knowledge Distillation},
    author={Sebastian Hofstätter and Sophia Althammer and Michael Schröder and Mete Sertkan and Allan Hanbury},
    year={2021},
    eprint={2010.02666},
    archivePrefix={arXiv},
    primaryClass={cs.IR}
}

FlopsLoss

@article{paria2020minimizing,
    title={Minimizing flops to learn efficient sparse representations},
    author={Paria, Biswajit and Yeh, Chih-Kuan and Yen, Ian EH and Xu, Ning and Ravikumar, Pradeep and P{'o}czos, Barnab{'a}s},
    journal={arXiv preprint arXiv:2004.05665},
    year={2020}
}