tomaarsen/splade-cocondenser-msmarco-margin-mse-minilm-small-old

CoCondenser trained on MS MARCO

This is a SPLADE Sparse Encoder model finetuned from Luyu/co-condenser-marco 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: Luyu/co-condenser-marco
• Maximum Sequence Length: 512 tokens
• Output Dimensionality: 30522 dimensions
• Similarity Function: Dot Product
• Language: en
• License: apache-2.0

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}) with MLMTransformer model: 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("tomaarsen/splade-cocondenser-msmarco-margin-mse-minilm")
# Run inference
queries = [
    "what causes aging fast",
]
documents = [
    'UV-A light, specifically, is what mainly causes tanning, skin aging, and cataracts, UV-B causes sunburn, skin aging and skin cancer, and UV-C is the strongest, and therefore most effective at killing microorganisms. Again â\x80\x93 single words and multiple bullets.',
    "Answers from Ronald Petersen, M.D. Yes, Alzheimer's disease usually worsens slowly. But its speed of progression varies, depending on a person's genetic makeup, environmental factors, age at diagnosis and other medical conditions. Still, anyone diagnosed with Alzheimer's whose symptoms seem to be progressing quickly â\x80\x94 or who experiences a sudden decline â\x80\x94 should see his or her doctor.",
    "Bell's palsy and Extreme tiredness and Extreme fatigue (2 causes) Bell's palsy and Extreme tiredness and Hepatitis (2 causes) Bell's palsy and Extreme tiredness and Liver pain (2 causes) Bell's palsy and Extreme tiredness and Lymph node swelling in children (2 causes)",
]
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([[11.2444, 10.6804,  4.3465]])

Evaluation

Metrics

Sparse Information Retrieval

• Datasets: NanoMSMARCO, NanoNFCorpus and NanoNQ
• Evaluated with SparseInformationRetrievalEvaluator

Metric	NanoMSMARCO	NanoNFCorpus	NanoNQ
dot_accuracy@1	0.42	0.44	0.48
dot_accuracy@3	0.66	0.64	0.74
dot_accuracy@5	0.76	0.64	0.8
dot_accuracy@10	0.84	0.68	0.88
dot_precision@1	0.42	0.44	0.48
dot_precision@3	0.22	0.3933	0.2533
dot_precision@5	0.152	0.336	0.168
dot_precision@10	0.084	0.27	0.094
dot_recall@1	0.42	0.0439	0.46
dot_recall@3	0.66	0.0987	0.7
dot_recall@5	0.76	0.1141	0.76
dot_recall@10	0.84	0.1401	0.84
dot_ndcg@10	0.6312	0.3445	0.664
dot_mrr@10	0.5637	0.5322	0.6205
dot_map@100	0.5721	0.1566	0.6042
query_active_dims	21.1	17.92	25.1
query_sparsity_ratio	0.9993	0.9994	0.9992
corpus_active_dims	157.6907	311.426	194.1861
corpus_sparsity_ratio	0.9948	0.9898	0.9936

Sparse Nano BEIR

• Dataset: NanoBEIR_mean
• Evaluated with SparseNanoBEIREvaluator with these parameters:

  {
      "dataset_names": [
          "msmarco",
          "nfcorpus",
          "nq"
      ]
  }
  

Metric	Value
dot_accuracy@1	0.4467
dot_accuracy@3	0.68
dot_accuracy@5	0.7333
dot_accuracy@10	0.8
dot_precision@1	0.4467
dot_precision@3	0.2889
dot_precision@5	0.2187
dot_precision@10	0.1493
dot_recall@1	0.308
dot_recall@3	0.4862
dot_recall@5	0.5447
dot_recall@10	0.6067
dot_ndcg@10	0.5466
dot_mrr@10	0.5721
dot_map@100	0.4443
query_active_dims	21.3733
query_sparsity_ratio	0.9993
corpus_active_dims	206.6305
corpus_sparsity_ratio	0.9932

Training Details

Training Dataset

Unnamed Dataset

• Size: 90,000 training samples
• Columns: query, positive, negative, and score
• Approximate statistics based on the first 1000 samples:

  	query	positive	negative	score
  type	string	string	string	float
  details	min: 4 tokens mean: 9.22 tokens max: 36 tokens	min: 15 tokens mean: 79.27 tokens max: 247 tokens	min: 16 tokens mean: 81.15 tokens max: 201 tokens	min: -14.32 mean: 4.62 max: 21.72

• Samples:

  query	positive	negative	score
  most powerful army in the world	U.S. Army Reserve Command You may be asking yourself, “What is the Army Reserve?” The Army is the most powerful and sophisticated military force in the world.	The British Royal Navy was the most powerful sea-going force by the time of World War 1 (1914-1918) and this was well-underst...	2.919867515563965
  define vasomotor	Define peripheral neuropathy: a disease or degenerative state of the peripheral nerves in which motor, sensory, or vasomotor nerve fibers may be… a disease or degenerative state of the peripheral nerves in which motor, sensory, or vasomotor nerve fibers may be affected and which is marked…	Vairāgya (Devanagari: वैराग्य, also spelt Vairagya) is a Sanskrit term used in Hindu philosophy that roughly translates as dispassion, detachment, or renunciation, in particular renunciation from the pains and pleasures in the material world (Maya).	3.0037026405334473
  nitrates definition biology	In Botany or Plant Biology. By Photosynthesis, the palisade cells make glucose which has many uses including: storage as starch, to make fat, to make cellulose and to make protein. Glucose is converted w…ith mineral slat nitrates to make the protein. Nitrates provide the essential nitrogen to make protein. The Ribosome, an organelle of the plant cell, manufactures most of the cell's protein.	Almost all inorganic nitrate salts are soluble in water at standard temperature and pressure. A common example of an inorganic nitrate salt is potassium nitrate (saltpeter). A rich source of inorganic nitrate in the human body comes from diets rich in leafy green foods, such as spinach and arugula.It is now believed that dietary nitrate in the form of plant-based foods is converted in the body to nitrite.itrate is a polyatomic ion with the molecular formula NO 3 − and a molecular mass of 62.0049 g/mol.	-1.6804794073104858

• Loss: SpladeLoss with these parameters:

  {
      "loss": "SparseMarginMSELoss",
      "lambda_corpus": 0.08,
      "lambda_query": 0.1
  }
  

Evaluation Dataset

Unnamed Dataset

• Size: 10,000 evaluation samples
• Columns: query, positive, negative, and score
• Approximate statistics based on the first 1000 samples:

  	query	positive	negative	score
  type	string	string	string	float
  details	min: 4 tokens mean: 9.01 tokens max: 35 tokens	min: 17 tokens mean: 79.8 tokens max: 336 tokens	min: 18 tokens mean: 81.3 tokens max: 273 tokens	min: -15.9 mean: 4.91 max: 21.67

• Samples:

  query	positive	negative	score
  femoral artery definition	medical Definition of circumflex artery : any of several paired curving arteries: as a: either of two arteries that branch from the deep femoral artery or from the femoral artery itself:	Femoral vein. The femoral vein is located in the upper thigh and pelvic region of the human body. It travels in close proximity to the femoral artery. This vein is one of the larger vessels in the venous system. Instead of draining deoxygenated blood from specific parts of the body, it receives blood from several significant branches. These include popliteal, the profunda femoris, and the great sapheneous veins.	-0.1968388557434082
  what causes mastitis and how do you treat it	Mastitis is an infection of the tissue of the breast that occurs most frequently during the time of breastfeeding. This infection causes pain, swelling, redness, and increased temperature of the breast. It can occur when bacteria, often from the infant's mouth, enter a milk duct through a crack in the nipple. This causes an infection and painful inflammation of the breast.	Common causes of mastitis include bacteria from the baby’s mouth, bacteria entering via breast injuries (bruising, fissures, cracks in the nipple), milk stasis (milk pooling in the breast), and bacteria from the hands of the mother or health care provider.	-0.8143405914306641
  what is a buck moth	Buck moth caterpillars that have a light background color can be confused with both the Nevada buck moth, Hemileuca nevadensis Stretch, and the New England buck moth, Hemileuca lucina Henry Edwards. The larvae of these three species can best be distinguished based on the preferred host plants (Wagner 2005).hey rely on resources that are acquired by the caterpillars (larvae). The caterpillars are robust and can exceed four inches (10 cm) in North America. Figure 4. Adult cecropia moth, Hyalophora cecropia (Linnaeus). Photograph by Pennsylvania Department of Conservation and Natural Resources-Forestry Archive, Bugwood.org.	bucktail that gets talked about quietly in the . privacy of remote cabins. The “Musky-Teer” is a big fish bait that anglers treasure in their collection. You won’t find these at your local bait shop but we’ve been stocking these highly prized baits in all colors for years.	11.004357814788818

• Loss: SpladeLoss with these parameters:

  {
      "loss": "SparseMarginMSELoss",
      "lambda_corpus": 0.08,
      "lambda_query": 0.1
  }
  

Training Hyperparameters

Non-Default Hyperparameters

• eval_strategy: steps
• per_device_train_batch_size: 16
• per_device_eval_batch_size: 16
• learning_rate: 2e-05
• num_train_epochs: 1
• warmup_ratio: 0.1
• fp16: True
• batch_sampler: no_duplicates

All Hyperparameters

Click to expand

• overwrite_output_dir: False
• do_predict: False
• eval_strategy: steps
• prediction_loss_only: True
• per_device_train_batch_size: 16
• per_device_eval_batch_size: 16
• 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: 2e-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: 1
• max_steps: -1
• lr_scheduler_type: linear
• lr_scheduler_kwargs: {}
• warmup_ratio: 0.1
• 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: False
• 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
• 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: False
• resume_from_checkpoint: None
• hub_model_id: None
• hub_strategy: every_save
• hub_private_repo: None
• hub_always_push: False
• 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
• 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	Validation Loss	NanoMSMARCO_dot_ndcg@10	NanoNFCorpus_dot_ndcg@10	NanoNQ_dot_ndcg@10	NanoBEIR_mean_dot_ndcg@10
0.0178	100	501776.8	-	-	-	-	-
0.0356	200	9740.8356	-	-	-	-	-
0.0533	300	61.9771	-	-	-	-	-
0.0711	400	37.6145	-	-	-	-	-
0.0889	500	28.8887	24.4953	0.4878	0.3047	0.5425	0.4450
0.1067	600	24.7991	-	-	-	-	-
0.1244	700	22.1517	-	-	-	-	-
0.1422	800	22.0889	-	-	-	-	-
0.16	900	20.7825	-	-	-	-	-
0.1778	1000	20.0856	18.6383	0.5751	0.3303	0.6100	0.5051
0.1956	1100	18.6968	-	-	-	-	-
0.2133	1200	20.5069	-	-	-	-	-
0.2311	1300	19.8162	-	-	-	-	-
0.2489	1400	19.1892	-	-	-	-	-
0.2667	1500	17.5024	18.0698	0.5750	0.3281	0.6222	0.5084
0.2844	1600	17.7801	-	-	-	-	-
0.3022	1700	17.9045	-	-	-	-	-
0.32	1800	16.3731	-	-	-	-	-
0.3378	1900	16.293	-	-	-	-	-
0.3556	2000	16.1167	14.5428	0.5696	0.3422	0.6232	0.5116
0.3733	2100	16.561	-	-	-	-	-
0.3911	2200	16.5533	-	-	-	-	-
0.4089	2300	14.9371	-	-	-	-	-
0.4267	2400	15.565	-	-	-	-	-
0.4444	2500	14.2143	15.2027	0.6071	0.3376	0.6600	0.5349
0.4622	2600	13.7188	-	-	-	-	-
0.48	2700	14.8554	-	-	-	-	-
0.4978	2800	15.1021	-	-	-	-	-
0.5156	2900	13.3032	-	-	-	-	-
0.5333	3000	13.8999	12.9609	0.5874	0.3423	0.6562	0.5286
0.5511	3100	12.7418	-	-	-	-	-
0.5689	3200	12.9422	-	-	-	-	-
0.5867	3300	13.6937	-	-	-	-	-
0.6044	3400	13.1183	-	-	-	-	-
0.6222	3500	12.7998	12.2024	0.6262	0.3424	0.6771	0.5486
0.64	3600	12.7799	-	-	-	-	-
0.6578	3700	12.2294	-	-	-	-	-
0.6756	3800	13.6836	-	-	-	-	-
0.6933	3900	13.579	-	-	-	-	-
0.7111	4000	12.6337	13.9878	0.6156	0.3435	0.6526	0.5372
0.7289	4100	12.682	-	-	-	-	-
0.7467	4200	12.2157	-	-	-	-	-
0.7644	4300	12.3127	-	-	-	-	-
0.7822	4400	11.7435	-	-	-	-	-
0.8	4500	12.086	12.3685	0.6262	0.3386	0.6782	0.5477
0.8178	4600	12.5455	-	-	-	-	-
0.8356	4700	11.7477	-	-	-	-	-
0.8533	4800	11.9948	-	-	-	-	-
0.8711	4900	11.8997	-	-	-	-	-
0.8889	5000	12.1624	12.8277	0.6241	0.3515	0.6740	0.5499
0.9067	5100	11.4352	-	-	-	-	-
0.9244	5200	10.9171	-	-	-	-	-
0.9422	5300	11.3242	-	-	-	-	-
0.96	5400	11.437	-	-	-	-	-
0.9778	5500	11.3141	11.6410	0.6366	0.3441	0.6605	0.5471
0.9956	5600	11.8683	-	-	-	-	-
-1	-1	-	-	0.6312	0.3445	0.6640	0.5466

Environmental Impact

Carbon emissions were measured using CodeCarbon.

• Energy Consumed: 0.225 kWh
• Carbon Emitted: 0.088 kg of CO2
• Hours Used: 0.653 hours

Training Hardware

• On Cloud: No
• GPU Model: 1 x NVIDIA GeForce RTX 3090
• CPU Model: 13th Gen Intel(R) Core(TM) i7-13700K
• RAM Size: 31.78 GB

Framework Versions

• Python: 3.11.6
• Sentence Transformers: 4.2.0.dev0
• Transformers: 4.52.4
• PyTorch: 2.6.0+cu124
• Accelerate: 1.5.1
• Datasets: 2.21.0
• Tokenizers: 0.21.1

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}
    }