yosefw/SPLADE-BERT-Medium-BS384

SPLADE-BERT-Medium

This is a SPLADE Sparse Encoder model finetuned from prajjwal1/bert-medium using the sentence-transformers library. It maps sentences & paragraphs to a 30522-dimensional sparse vector space and can be used for semantic search and sparse retrieval.

Model Details

Model Description

• Model Type: SPLADE Sparse Encoder
• Base model: prajjwal1/bert-medium
• Maximum Sequence Length: 512 tokens
• Output Dimensionality: 30522 dimensions
• Similarity Function: Dot Product
• Language: en
• License: mit

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})
)

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("yosefw/SPLADE-BERT-Medium-BS384")
# Run inference
queries = [
    "how long to bake arm roast",
]
documents = [
    'Line baking dish ... to also cover roast). Place roast ... the roast. Place in preheated 300 degree oven for 2 1/2 to 3 hours. About 50 minutes per pound.rim all excess fat from roast. Place potatoes ... Crockery Pot on top of potatoes and onions. Cover and cook on low setting for 10 to 12 hours (high 5 to 6).',
    'Considerations. The total time it takes to cook an arm roast depends on its size. A 3- to 4-lb. chuck roast takes 5 to 6 hours on high and 10 to 12 hours on low.Chuck roasts usually contain enough marbled fat to cook without water, but most Crock-Pot roast recipes call for a little liquid.Most importantly, resist the temptation to lift the lid while your roast is cooking. 3- to 4-lb. chuck roast takes 5 to 6 hours on high and 10 to 12 hours on low. Chuck roasts usually contain enough marbled fat to cook without water, but most Crock-Pot roast recipes call for a little liquid. Most importantly, resist the temptation to lift the lid while your roast is cooking.',
    'Set your Crock Pot on high to reach a simmer point of 209 degrees F in 3 to 4 hours, or low to reach the same cooking temperature in 7 to 8 hours. The total time it takes to cook an arm roast depends on its size. A 3- to 4-lb. chuck roast takes 5 to 6 hours on high and 10 to 12 hours on low.Chuck roasts usually contain enough marbled fat to cook without water, but most Crock-Pot roast recipes call for a little liquid.Most importantly, resist the temptation to lift the lid while your roast is cooking. 3- to 4-lb. chuck roast takes 5 to 6 hours on high and 10 to 12 hours on low. Chuck roasts usually contain enough marbled fat to cook without water, but most Crock-Pot roast recipes call for a little liquid. Most importantly, resist the temptation to lift the lid while your roast is cooking.',
]
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([[16.1861, 15.3382, 15.6794]])

Evaluation

Metrics

Sparse Information Retrieval

• Evaluated with SparseInformationRetrievalEvaluator

Metric	Value
dot_accuracy@1	0.4716
dot_accuracy@3	0.7802
dot_accuracy@5	0.8684
dot_accuracy@10	0.9396
dot_precision@1	0.4716
dot_precision@3	0.2671
dot_precision@5	0.1806
dot_precision@10	0.0985
dot_recall@1	0.4563
dot_recall@3	0.7666
dot_recall@5	0.8592
dot_recall@10	0.9339
dot_ndcg@10	0.7089
dot_mrr@10	0.6398
dot_map@100	0.636
query_active_dims	23.285
query_sparsity_ratio	0.9992
corpus_active_dims	175.6307
corpus_sparsity_ratio	0.9942

Training Details

Training Dataset

Unnamed Dataset

• Size: 496,123 training samples
• Columns: query, positive, negative_1, and negative_2
• Approximate statistics based on the first 1000 samples:

  	query	positive	negative_1	negative_2
  type	string	string	string	string
  details	min: 4 tokens mean: 8.87 tokens max: 43 tokens	min: 24 tokens mean: 81.23 tokens max: 259 tokens	min: 20 tokens mean: 79.21 tokens max: 197 tokens	min: 20 tokens mean: 77.89 tokens max: 207 tokens

• Samples:

  query	positive	negative_1	negative_2
  heart specialists in ridgeland ms	Dr. George Reynolds Jr, MD is a cardiology specialist in Ridgeland, MS and has been practicing for 35 years. He graduated from Vanderbilt University School Of Medicine in 1977 and specializes in cardiology and internal medicine.	Dr. James Kramer is a Internist in Ridgeland, MS. Find Dr. Kramer's phone number, address and more.	Dr. James Kramer is an internist in Ridgeland, Mississippi. He received his medical degree from Loma Linda University School of Medicine and has been in practice for more than 20 years. Dr. James Kramer's Details
  does baytril otic require a prescription	Baytril Otic Ear Drops-Enrofloxacin/Silver Sulfadiazine-Prices & Information. A prescription is required for this item. A prescription is required for this item. Brand medication is not available at this time.	RX required for this item. Click here for our full Prescription Policy and Form. Baytril Otic (enrofloxacin/silver sulfadiazine) Emulsion from Bayer is the first fluoroquinolone approved by the Food and Drug Administration for the topical treatment of canine otitis externa.	Product Details. Baytril Otic is a highly effective treatment prescribed by many veterinarians when your pet has an ear infection caused by susceptible bacteria or fungus. Baytril Otic is: a liquid emulsion that is used topically directly in the ear or on the skin in order to treat susceptible bacterial and yeast infections.
  what is on a gyro	Report Abuse. Gyros or gyro (giros) (pronounced /ˈjɪəroʊ/ or /ˈdʒaɪroʊ/, Greek: γύρος turn) is a Greek dish consisting of meat (typically lamb and/or beef), tomato, onion, and tzatziki sauce, and is served with pita bread. Chicken and pork meat can be used too.	A gyroscope (from Ancient Greek γῦρος gûros, circle and σκοπέω skopéō, to look) is a spinning wheel or disc in which the axis of rotation is free to assume any orientation by itself. When rotating, the orientation of this axis is unaffected by tilting or rotation of the mounting, according to the conservation of angular momentum.	Diagram of a gyro wheel. Reaction arrows about the output axis (blue) correspond to forces applied about the input axis (green), and vice versa. A gyroscope is a wheel mounted in two or three gimbals, which are a pivoted supports that allow the rotation of the wheel about a single axis.

• Loss: SpladeLoss with these parameters:

  {
      "loss": "SparseMultipleNegativesRankingLoss(scale=1.0, similarity_fct='dot_score', gather_across_devices=False)",
      "document_regularizer_weight": 0.003,
      "query_regularizer_weight": 0.005
  }
  

Training Hyperparameters

Non-Default Hyperparameters

• eval_strategy: epoch
• per_device_train_batch_size: 48
• per_device_eval_batch_size: 48
• gradient_accumulation_steps: 8
• learning_rate: 8e-05
• num_train_epochs: 8
• lr_scheduler_type: cosine
• warmup_ratio: 0.025
• fp16: True
• load_best_model_at_end: True
• push_to_hub: True
• batch_sampler: no_duplicates

All Hyperparameters

Click to expand

• overwrite_output_dir: False
• do_predict: False
• eval_strategy: epoch
• prediction_loss_only: True
• per_device_train_batch_size: 48
• per_device_eval_batch_size: 48
• per_gpu_train_batch_size: None
• per_gpu_eval_batch_size: None
• gradient_accumulation_steps: 8
• eval_accumulation_steps: None
• torch_empty_cache_steps: None
• learning_rate: 8e-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: 8
• 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: no_duplicates
• multi_dataset_batch_sampler: proportional
• router_mapping: {}
• learning_rate_mapping: {}

Training Logs

Epoch	Step	Training Loss	dot_ndcg@10
1.0	1292	42.0325	0.7155
2.0	2584	1.1261	0.7216
3.0	3876	1.049	0.7214
4.0	5168	0.9631	0.7188
5.0	6460	0.8725	0.7120
-1	-1	-	0.7089

Framework Versions

• Python: 3.12.11
• Sentence Transformers: 5.1.0
• Transformers: 4.55.4
• PyTorch: 2.8.0+cu126
• Accelerate: 1.10.1
• 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},
}

SparseMultipleNegativesRankingLoss

@misc{henderson2017efficient,
    title={Efficient Natural Language Response Suggestion for Smart Reply},
    author={Matthew Henderson and Rami Al-Rfou and Brian Strope and Yun-hsuan Sung and Laszlo Lukacs and Ruiqi Guo and Sanjiv Kumar and Balint Miklos and Ray Kurzweil},
    year={2017},
    eprint={1705.00652},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}

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}
}