rasyosef/splade-small

SPLADE-BERT-Small-Distil

This is a SPLADE sparse retrieval model based on BERT-Small (29M) that was trained by distilling a Cross-Encoder on the MSMARCO dataset. The cross-encoder used was ms-marco-MiniLM-L6-v2.

This SPLADE model is 2x smaller than Naver's official splade-v3-distilbert while having 91% of it's performance on the MSMARCO benchmark. 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
rasyosef/splade-small	28.8M	35.2
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-small")
# Run inference
queries = [
    "is cognac whisky",
]
documents = [
    'Cognac vs Whiskey. • Whiskey is the alcoholic drink made from grains whereas Cognac is the alcoholic drink made from grapes. • Cognac is a type of brandy. In fact, many label it as the finest of brandies. • Cognac is the brandy originating from a wine producing region of France called Cognac. • While a cognac is considered an after dinner beverage that is intended to digest food, there is no such stereotyping of whiskey that can be consumed anytime of the day.',
    '10 hilariously insightful foreign words. One of the most obvious differences between cognac and whiskey is that cognac makers use grapes, and whiskey makers use grains. Although both processes use fermentation to create the liquors, cognac makers use a double distillation process.',
    'The word whisky (or whiskey) is an anglicisation of the Classical Gaelic word uisce / uisge meaning water (now written as uisce in Irish Gaelic, and uisge in Scottish Gaelic). Distilled alcohol was known in Latin as aqua vitae (water of life).',
]
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([[22.4589, 20.5905, 10.0662]])

Model Details

Model Description

• Model Type: SPLADE Sparse Encoder
• Base model: prajjwal1/bert-small
• 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})
)

More

Click to expand

Evaluation

Metrics

Sparse Information Retrieval

• Evaluated with SparseInformationRetrievalEvaluator

Metric	Value
dot_accuracy@1	0.5172
dot_accuracy@3	0.8368
dot_accuracy@5	0.9232
dot_accuracy@10	0.9762
dot_precision@1	0.5172
dot_precision@3	0.2867
dot_precision@5	0.1924
dot_precision@10	0.1027
dot_recall@1	0.5006
dot_recall@3	0.8238
dot_recall@5	0.9153
dot_recall@10	0.9723
dot_ndcg@10	0.7554
dot_mrr@10	0.6877
dot_map@100	0.6829
query_active_dims	29.7198
query_sparsity_ratio	0.999
corpus_active_dims	168.3538
corpus_sparsity_ratio	0.9945

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: 9.04 tokens max: 41 tokens	min: 18 tokens mean: 81.11 tokens max: 215 tokens	min: 18 tokens mean: 77.81 tokens max: 247 tokens	min: 22 tokens mean: 76.2 tokens max: 217 tokens	size: 2 elements

• Samples:

  query	positive	negative_1	negative_2	label
  The _____________________ is a body system which consists of glands that produce hormones that act throughout the body.	Endocrine System. The endocrine system is made up of a group of glands that produce the body's long-distance messengers, or hormones. Hormones are chemicals that control body functions, such as metabolism, growth, and sexual development.t is made up of a group of organs that transport blood throughout the body. The heart pumps the blood and the arteries and veins transport it. Oxygen-rich blood leaves the left side of the heart and enters the biggest artery, called the aorta.	The endocrine system is a control system of ductless glands that secrete hormones within specific organs. Hormones act as messengers, and are carried by the bloodstream to different cells in the body, which interpret these messages and act on them.he pancreas is unusual among the body's glands in that it also has a very important endocrine function. Small groups of special cells called islet cells throughout the organ make the hormones of insulin and glucagon.	These glands produce different types of hormones that evoke a specific response in other cells, tissues, and/or organs located throughout the body. The hormones reach these faraway targets using the blood stream. Like the nervous system, the endocrine system is one of your body’s main communicators.he Endocrine System Essentials. 1 The endocrine system is made up of a network of glands. 2 These glands secrete hormones to regulate many bodily functions, including growth and metabolism.	[2.3722684383392334, 5.211579322814941]
  causes of low body temperature in adults	Hypothermia is defined as a body temperature (core, or internal body temperature) of less than about 95 F (35 C). Usually, hypothermia occurs when the body's temperature regulation is overwhelmed by a cold environment. However, in the medical and lay literature there are essentially two major classifications, accidental hypothermia and intentional hypothermia.	In general, a baby has a fever when their body temperature exceeds 100.4°F, or 38°C. A child has a fever when their temperature exceeds 99.5°F, or 37.5°C. An adult has a fever when their temperature exceeds 99 to 99.5°F, or 37.2 to 37.5°C.	Consequently, an accurate measurement of body temperature (best is rectal core temperature) of 100.4 F (38 C) or above is considered to be a fever.. A newer option includes a temperature-sensitive infrared device that measures the temperature in the skin by simply rubbing the sensor on the body.	[1.3747079372406006, 8.096447944641113]
  who is laila gifty akita	Lailah Gifty Akita is a Ghanaian and founder of Smart Youth Volunteers Foundation. She obtained a BSc in Renewable Natural Resources Management at Kwame Nkrumah University of Science and Technology, Kumasi-Ghana. She also had MPhil in Oceanography at the University of Ghana. She obtained a doctorate in Geosciences at International Max Planck Research School for Global Biogeochemical Cycles-Friedrich Schiller University of Jena, Germany ( June 2011 to March 2016). Lailah is an influential lady with the passion of empowering the mind of young people to make a great difference.	She is a PhD-student, studying Geosciences at the University of Jena, Germany. She is an enthusiastic inspirational writer. She wishes to challenge and inspire people from all walks of life to dare a greater life. You can capable of heroic deeds. Think well of yourself and act positively. You can correspond with Lailah via an email:[email protected]. https://www.goodreads.com/author/show/8297615.Lailah_Gifty_Akita/blog.	Also in the Talmud, the interpretation is found of rabbi Hanina ben Pappa (3rd century AD), that Lailah is an angel in charge of conception who takes a drop of semen and places it before God, saying: For R. Hanina b.	[2.6488447189331055, 15.058775901794434]

• Loss: SpladeLoss with these parameters:

  {
      "loss": "SparseMarginMSELoss",
      "document_regularizer_weight": 0.12,
      "query_regularizer_weight": 0.2
  }
  

Training Hyperparameters

Non-Default Hyperparameters

• eval_strategy: epoch
• per_device_train_batch_size: 64
• per_device_eval_batch_size: 64
• learning_rate: 4e-05
• num_train_epochs: 4
• lr_scheduler_type: cosine
• warmup_ratio: 0.025
• 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: 64
• per_device_eval_batch_size: 64
• 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: 4e-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: 4
• max_steps: -1
• lr_scheduler_type: cosine
• lr_scheduler_kwargs: {}
• warmup_ratio: 0.025
• 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	18750	7.806	0.7439
2.0	37500	5.7509	0.7520
3.0	56250	4.5026	0.7554
4.0	75000	3.909	0.7534
-1	-1	-	0.7554

• The bold row denotes the saved checkpoint.

Framework Versions

• Python: 3.11.13
• Sentence Transformers: 5.1.0
• Transformers: 4.55.2
• PyTorch: 2.6.0+cu124
• Accelerate: 1.10.0
• Datasets: 4.0.0
• Tokenizers: 0.21.4

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}
}