capemox/splade-ettin-encoder-17m-gooaq

splade-ettin-encoder-17m trained on GooAQ

This is a SPLADE Sparse Encoder model finetuned from jhu-clsp/ettin-encoder-17m on the gooaq dataset using the sentence-transformers library. It maps sentences & paragraphs to a 50368-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: jhu-clsp/ettin-encoder-17m
• Maximum Sequence Length: 256 tokens
• Output Dimensionality: 50368 dimensions
• Similarity Function: Dot Product
• Supported Modality: Text
• Training Dataset:
  • gooaq
• 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): Transformer({'transformer_task': 'fill-mask', 'modality_config': {'text': {'method': 'forward', 'method_output_name': 'logits'}}, 'module_output_name': 'token_embeddings', 'architecture': 'ModernBertForMaskedLM'})
  (1): SpladePooling({'pooling_strategy': 'max', 'activation_function': 'relu', 'embedding_dimension': 50368})
)

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("capemox/splade-ettin-encoder-17m-gooaq")
# Run inference
sentences = [
    'how close is the closest star?',
    'The two main stars are Alpha Centauri A and Alpha Centauri B, which form a binary pair. They are an average of 4.3 light-years from Earth. The third star is Proxima Centauri. It is about 4.22 light-years from Earth and is the closest star other than the sun.',
    'One gallon can of paint will cover up to 400 square feet, which is enough to cover a small room like a bathroom. Two gallon cans of paint cover up to 800 square feet, which is enough to cover an average size room. This is the most common amount needed, especially when considering second coat coverage.',
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# [3, 50368]

# Get the similarity scores for the embeddings
similarities = model.similarity(embeddings, embeddings)
print(similarities)
# tensor([[144.9772, 158.3885, 140.6785],
#         [158.3885, 496.0056, 344.2902],
#         [140.6785, 344.2902, 528.3322]])

Evaluation

Metrics

Sparse Information Retrieval

• Datasets: NanoMSMARCO_256, NanoNFCorpus_256, NanoNQ_256, NanoClimateFEVER_256, NanoDBPedia_256, NanoFEVER_256, NanoFiQA2018_256, NanoHotpotQA_256, NanoMSMARCO_256, NanoNFCorpus_256, NanoNQ_256, NanoQuoraRetrieval_256, NanoSCIDOCS_256, NanoArguAna_256, NanoSciFact_256 and NanoTouche2020_256
• Evaluated with SparseInformationRetrievalEvaluator with these parameters:

  {
      "max_active_dims": 256
  }
  

Metric	NanoMSMARCO_256	NanoNFCorpus_256	NanoNQ_256	NanoClimateFEVER_256	NanoDBPedia_256	NanoFEVER_256	NanoFiQA2018_256	NanoHotpotQA_256	NanoQuoraRetrieval_256	NanoSCIDOCS_256	NanoArguAna_256	NanoSciFact_256	NanoTouche2020_256
dot_accuracy@1	0.2	0.22	0.1	0.2	0.4	0.34	0.18	0.58	0.16	0.34	0.04	0.3	0.4898
dot_accuracy@3	0.34	0.34	0.28	0.38	0.66	0.7	0.3	0.7	0.38	0.5	0.32	0.44	0.7755
dot_accuracy@5	0.42	0.38	0.4	0.42	0.74	0.78	0.4	0.78	0.4	0.58	0.42	0.64	0.8776
dot_accuracy@10	0.6	0.52	0.54	0.48	0.84	0.92	0.44	0.84	0.46	0.72	0.48	0.7	0.9796
dot_precision@1	0.2	0.22	0.1	0.2	0.4	0.34	0.18	0.58	0.16	0.34	0.04	0.3	0.4898
dot_precision@3	0.1133	0.1933	0.0933	0.1267	0.3733	0.2333	0.14	0.32	0.1267	0.22	0.1067	0.1533	0.4218
dot_precision@5	0.084	0.18	0.08	0.092	0.348	0.156	0.124	0.216	0.08	0.18	0.084	0.14	0.4
dot_precision@10	0.06	0.14	0.058	0.066	0.324	0.094	0.074	0.126	0.048	0.138	0.048	0.078	0.3469
dot_recall@1	0.2	0.0069	0.09	0.085	0.0438	0.3067	0.0822	0.29	0.16	0.0727	0.04	0.275	0.034
dot_recall@3	0.34	0.0187	0.25	0.1783	0.0853	0.6567	0.1734	0.48	0.354	0.1377	0.32	0.42	0.0824
dot_recall@5	0.42	0.0306	0.36	0.2	0.1356	0.7367	0.2406	0.54	0.374	0.1857	0.42	0.61	0.1281
dot_recall@10	0.6	0.0458	0.52	0.26	0.2097	0.8767	0.2906	0.63	0.444	0.2837	0.48	0.68	0.2193
dot_ndcg@10	0.3714	0.1571	0.2925	0.2151	0.3792	0.5965	0.2237	0.5601	0.3147	0.2673	0.2567	0.4731	0.389
dot_mrr@10	0.302	0.2964	0.2323	0.2989	0.5519	0.5313	0.2532	0.6587	0.277	0.4424	0.1845	0.4177	0.6561
dot_map@100	0.3209	0.0466	0.2305	0.1742	0.2587	0.4997	0.1863	0.494	0.2826	0.1904	0.1898	0.4079	0.27
query_active_dims	252.1	255.62	256.0	256.0	253.96	256.0	240.22	256.0	229.0	256.0	256.0	256.0	216.8571
query_sparsity_ratio	0.995	0.9949	0.9949	0.9949	0.995	0.9949	0.9952	0.9949	0.9955	0.9949	0.9949	0.9949	0.9957
corpus_active_dims	256.0	256.0	256.0	256.0	256.0	256.0	256.0	256.0	234.9956	256.0	256.0	256.0	255.9241
corpus_sparsity_ratio	0.9949	0.9949	0.9949	0.9949	0.9949	0.9949	0.9949	0.9949	0.9953	0.9949	0.9949	0.9949	0.9949
avg_flops	104.6667	101.2659	114.127	129.3998	108.9664	120.7999	116.0173	114.6017	119.7198	123.3701	150.7188	129.6654	86.4117

Sparse Nano BEIR

• Dataset: NanoBEIR_mean_256
• Evaluated with SparseNanoBEIREvaluator with these parameters:

  {
      "dataset_names": [
          "msmarco",
          "nfcorpus",
          "nq"
      ],
      "dataset_id": "sentence-transformers/NanoBEIR-en",
      "max_active_dims": 256
  }
  

Metric	Value
dot_accuracy@1	0.18
dot_accuracy@3	0.3267
dot_accuracy@5	0.42
dot_accuracy@10	0.54
dot_precision@1	0.18
dot_precision@3	0.1356
dot_precision@5	0.1173
dot_precision@10	0.0867
dot_recall@1	0.0991
dot_recall@3	0.2158
dot_recall@5	0.2848
dot_recall@10	0.379
dot_ndcg@10	0.273
dot_mrr@10	0.2783
dot_map@100	0.1991
query_active_dims	254.54
query_sparsity_ratio	0.9949
corpus_active_dims	256.0
corpus_sparsity_ratio	0.9949
avg_flops	105.0486

Sparse Nano BEIR

• Dataset: NanoBEIR_mean_256
• Evaluated with SparseNanoBEIREvaluator with these parameters:

  {
      "dataset_names": [
          "climatefever",
          "dbpedia",
          "fever",
          "fiqa2018",
          "hotpotqa",
          "msmarco",
          "nfcorpus",
          "nq",
          "quoraretrieval",
          "scidocs",
          "arguana",
          "scifact",
          "touche2020"
      ],
      "dataset_id": "sentence-transformers/NanoBEIR-en",
      "max_active_dims": 256
  }
  

Metric	Value
dot_accuracy@1	0.2731
dot_accuracy@3	0.4704
dot_accuracy@5	0.5567
dot_accuracy@10	0.6554
dot_precision@1	0.2731
dot_precision@3	0.2017
dot_precision@5	0.1665
dot_precision@10	0.1231
dot_recall@1	0.1297
dot_recall@3	0.269
dot_recall@5	0.337
dot_recall@10	0.4261
dot_ndcg@10	0.3459
dot_mrr@10	0.3925
dot_map@100	0.2732
query_active_dims	249.2619
query_sparsity_ratio	0.9951
corpus_active_dims	254.1114
corpus_sparsity_ratio	0.995
avg_flops	109.026

Training Details

Training Dataset

gooaq

• Dataset: gooaq at b089f72
• Size: 99,000 training samples
• Columns: question and answer
• Approximate statistics based on the first 1000 samples:

  	question	answer
  type	string	string
  details	min: 8 tokens mean: 11.93 tokens max: 21 tokens	min: 15 tokens mean: 58.27 tokens max: 137 tokens

• Samples:

  question	answer
  what are the 5 characteristics of a star?	Key Concept: Characteristics used to classify stars include color, temperature, size, composition, and brightness.
  are copic markers alcohol ink?	Copic Ink is alcohol-based and flammable. Keep away from direct sunlight and extreme temperatures.
  what is the difference between appellate term and appellate division?	Appellate terms An appellate term is an intermediate appellate court that hears appeals from the inferior courts within their designated counties or judicial districts, and are intended to ease the workload on the Appellate Division and provide a less expensive forum closer to the people.

• Loss: SpladeLoss with these parameters:

  {
      "loss": "SparseMultipleNegativesRankingLoss(scale=1.0, similarity_fct='dot_score', gather_across_devices=False, directions=('query_to_doc',), partition_mode='joint', hardness_mode=None, hardness_strength=0.0)",
      "document_regularizer_weight": 3e-05,
      "query_regularizer_weight": 5e-05
  }
  

Evaluation Dataset

gooaq

• Dataset: gooaq at b089f72
• Size: 1,000 evaluation samples
• Columns: question and answer
• Approximate statistics based on the first 1000 samples:

  	question	answer
  type	string	string
  details	min: 8 tokens mean: 12.05 tokens max: 23 tokens	min: 15 tokens mean: 58.98 tokens max: 186 tokens

• Samples:

  question	answer
  should you take ibuprofen with high blood pressure?	In general, people with high blood pressure should use acetaminophen or possibly aspirin for over-the-counter pain relief. Unless your health care provider has said it's OK, you should not use ibuprofen, ketoprofen, or naproxen sodium. If aspirin or acetaminophen doesn't help with your pain, call your doctor.
  how old do you have to be to work in sc?	The general minimum age of employment for South Carolina youth is 14, although the state allows younger children who are performers to work in show business. If their families are agricultural workers, children younger than age 14 may also participate in farm labor.
  how to write a topic proposal for a research paper?	['Write down the main topic of your paper. ... ', 'Write two or three short sentences under the main topic that explain why you chose that topic. ... ', 'Write a thesis sentence that states the angle and purpose of your research paper. ... ', 'List the items you will cover in the body of the paper that support your thesis statement.']

• Loss: SpladeLoss with these parameters:

  {
      "loss": "SparseMultipleNegativesRankingLoss(scale=1.0, similarity_fct='dot_score', gather_across_devices=False, directions=('query_to_doc',), partition_mode='joint', hardness_mode=None, hardness_strength=0.0)",
      "document_regularizer_weight": 3e-05,
      "query_regularizer_weight": 5e-05
  }
  

Training Hyperparameters

Non-Default Hyperparameters

• per_device_train_batch_size: 32
• num_train_epochs: 1
• learning_rate: 2e-05
• bf16: True
• per_device_eval_batch_size: 32
• load_best_model_at_end: True
• batch_sampler: no_duplicates

All Hyperparameters

Click to expand

• per_device_train_batch_size: 32
• num_train_epochs: 1
• max_steps: -1
• learning_rate: 2e-05
• lr_scheduler_type: linear
• lr_scheduler_kwargs: None
• warmup_steps: 0
• optim: adamw_torch_fused
• optim_args: None
• weight_decay: 0.0
• adam_beta1: 0.9
• adam_beta2: 0.999
• adam_epsilon: 1e-08
• optim_target_modules: None
• gradient_accumulation_steps: 1
• average_tokens_across_devices: True
• max_grad_norm: 1.0
• label_smoothing_factor: 0.0
• bf16: True
• fp16: False
• bf16_full_eval: False
• fp16_full_eval: False
• tf32: None
• gradient_checkpointing: False
• gradient_checkpointing_kwargs: None
• torch_compile: False
• torch_compile_backend: None
• torch_compile_mode: None
• use_liger_kernel: False
• liger_kernel_config: None
• use_cache: False
• neftune_noise_alpha: None
• torch_empty_cache_steps: None
• auto_find_batch_size: False
• log_on_each_node: True
• logging_nan_inf_filter: True
• include_num_input_tokens_seen: no
• log_level: passive
• log_level_replica: warning
• disable_tqdm: False
• project: huggingface
• trackio_space_id: trackio
• per_device_eval_batch_size: 32
• prediction_loss_only: True
• eval_on_start: False
• eval_do_concat_batches: True
• eval_use_gather_object: False
• eval_accumulation_steps: None
• include_for_metrics: []
• batch_eval_metrics: False
• save_only_model: False
• save_on_each_node: False
• enable_jit_checkpoint: False
• push_to_hub: False
• hub_private_repo: None
• hub_model_id: None
• hub_strategy: every_save
• hub_always_push: False
• hub_revision: None
• load_best_model_at_end: True
• ignore_data_skip: False
• restore_callback_states_from_checkpoint: False
• full_determinism: False
• seed: 42
• data_seed: None
• use_cpu: False
• accelerator_config: {'split_batches': False, 'dispatch_batches': None, 'even_batches': True, 'use_seedable_sampler': True, 'non_blocking': False, 'gradient_accumulation_kwargs': None}
• parallelism_config: None
• dataloader_drop_last: False
• dataloader_num_workers: 0
• dataloader_pin_memory: True
• dataloader_persistent_workers: False
• dataloader_prefetch_factor: None
• remove_unused_columns: True
• label_names: None
• train_sampling_strategy: random
• length_column_name: length
• ddp_find_unused_parameters: None
• ddp_bucket_cap_mb: None
• ddp_broadcast_buffers: False
• ddp_backend: None
• ddp_timeout: 1800
• fsdp: []
• fsdp_config: {'min_num_params': 0, 'xla': False, 'xla_fsdp_v2': False, 'xla_fsdp_grad_ckpt': False}
• deepspeed: None
• debug: []
• skip_memory_metrics: True
• do_predict: False
• resume_from_checkpoint: None
• warmup_ratio: None
• local_rank: -1
• 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_256_dot_ndcg@10	NanoNFCorpus_256_dot_ndcg@10	NanoNQ_256_dot_ndcg@10	NanoBEIR_mean_256_dot_ndcg@10	NanoClimateFEVER_256_dot_ndcg@10	NanoDBPedia_256_dot_ndcg@10	NanoFEVER_256_dot_ndcg@10	NanoFiQA2018_256_dot_ndcg@10	NanoHotpotQA_256_dot_ndcg@10	NanoQuoraRetrieval_256_dot_ndcg@10	NanoSCIDOCS_256_dot_ndcg@10	NanoArguAna_256_dot_ndcg@10	NanoSciFact_256_dot_ndcg@10	NanoTouche2020_256_dot_ndcg@10
0.0323	100	954.7106	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.0646	200	17.8241	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.0970	300	4.5136	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.1293	400	2.5427	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.1616	500	1.4733	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.1939	600	1.0940	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.1972	610	-	0.8522	0.1440	0.0453	0.0771	0.0888	-	-	-	-	-	-	-	-	-	-
0.2262	700	0.7541	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.2586	800	0.7425	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.2909	900	0.5966	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.3232	1000	0.5606	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.3555	1100	0.5440	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.3878	1200	0.4032	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.3943	1220	-	0.4165	0.2494	0.0613	0.2065	0.1724	-	-	-	-	-	-	-	-	-	-
0.4202	1300	0.3995	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.4525	1400	0.2976	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.4848	1500	0.2971	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.5171	1600	0.2716	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.5495	1700	0.2577	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.5818	1800	0.2370	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.5915	1830	-	0.2104	0.3406	0.1173	0.2414	0.2331	-	-	-	-	-	-	-	-	-	-
0.6141	1900	0.2360	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.6464	2000	0.2238	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.6787	2100	0.2237	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.7111	2200	0.2162	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.7434	2300	0.2044	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.7757	2400	0.2202	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.7886	2440	-	0.1736	0.3932	0.1545	0.2717	0.2731	-	-	-	-	-	-	-	-	-	-
0.8080	2500	0.1672	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.8403	2600	0.2122	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.8727	2700	0.1704	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.9050	2800	0.1870	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.9373	2900	0.1671	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.9696	3000	0.1386	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.9858	3050	-	0.17	0.3714	0.1571	0.2925	0.2737	-	-	-	-	-	-	-	-	-	-
1.0	3094	-	0.1706	0.3675	0.1624	0.2892	0.2730	-	-	-	-	-	-	-	-	-	-
-1	-1	-	-	0.3714	0.1571	0.2925	0.3459	0.2151	0.3792	0.5965	0.2237	0.5601	0.3147	0.2673	0.2567	0.4731	0.3890

• The bold row denotes the saved checkpoint.

Training Time

• Training: 7.7 minutes
• Evaluation: 2.0 minutes
• Total: 9.7 minutes

Framework Versions

• Python: 3.12.3
• Sentence Transformers: 5.4.1
• Transformers: 5.5.4
• PyTorch: 2.11.0+cu130
• Accelerate: 1.13.0
• Datasets: 4.8.4
• Tokenizers: 0.22.2

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},
}

SparseMultipleNegativesRankingLoss

@misc{oord2019representationlearningcontrastivepredictive,
      title={Representation Learning with Contrastive Predictive Coding},
      author={Aaron van den Oord and Yazhe Li and Oriol Vinyals},
      year={2019},
      eprint={1807.03748},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/1807.03748},
}

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}
}