tomaarsen/splade-distilbert-base-uncased-gooaq-peft-r512

splade-distilbert-base-uncased trained on GooAQ

This is a SPLADE Sparse Encoder model finetuned from distilbert/distilbert-base-uncased on the gooaq dataset 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: distilbert/distilbert-base-uncased
• Maximum Sequence Length: 256 tokens
• Output Dimensionality: 30522 dimensions
• Similarity Function: Dot Product
• Training Dataset:
  • gooaq
• 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': 256, 'do_lower_case': False, 'architecture': 'DistilBertForMaskedLM'})
  (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-distilbert-base-uncased-gooaq-peft-r512")
# Run inference
queries = [
    "how many days for doxycycline to work on sinus infection?",
]
documents = [
    'Treatment of suspected bacterial infection is with antibiotics, such as amoxicillin/clavulanate or doxycycline, given for 5 to 7 days for acute sinusitis and for up to 6 weeks for chronic sinusitis.',
    'Most engagements typically have a cocktail dress code, calling for dresses at, or slightly above, knee-length and high heels. If your party states a different dress code, however, such as semi-formal or dressy-casual, you may need to dress up or down accordingly.',
    'The average service life of a gas furnace is about 15 years, but the actual life span of an individual unit can vary greatly. There are a number of contributing factors that determine the age a furnace reaches: The quality of the equipment.',
]
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([[95.1919, 19.8228, 34.4896]])

Evaluation

Metrics

Sparse Information Retrieval

• Datasets: NanoMSMARCO, NanoNFCorpus, NanoNQ, NanoClimateFEVER, NanoDBPedia, NanoFEVER, NanoFiQA2018, NanoHotpotQA, NanoMSMARCO, NanoNFCorpus, NanoNQ, NanoQuoraRetrieval, NanoSCIDOCS, NanoArguAna, NanoSciFact and NanoTouche2020
• Evaluated with SparseInformationRetrievalEvaluator

Metric	NanoMSMARCO	NanoNFCorpus	NanoNQ	NanoClimateFEVER	NanoDBPedia	NanoFEVER	NanoFiQA2018	NanoHotpotQA	NanoQuoraRetrieval	NanoSCIDOCS	NanoArguAna	NanoSciFact	NanoTouche2020
dot_accuracy@1	0.32	0.3	0.3	0.2	0.58	0.38	0.32	0.64	0.32	0.32	0.08	0.36	0.6327
dot_accuracy@3	0.52	0.54	0.54	0.32	0.76	0.56	0.48	0.76	0.56	0.6	0.26	0.48	0.7959
dot_accuracy@5	0.62	0.56	0.6	0.44	0.84	0.66	0.56	0.86	0.64	0.68	0.28	0.54	0.9592
dot_accuracy@10	0.7	0.62	0.76	0.52	0.92	0.8	0.6	0.92	0.72	0.8	0.34	0.62	1.0
dot_precision@1	0.32	0.3	0.3	0.2	0.58	0.38	0.32	0.64	0.32	0.32	0.08	0.36	0.6327
dot_precision@3	0.1733	0.32	0.18	0.1067	0.4667	0.1867	0.22	0.3867	0.1867	0.2533	0.0867	0.1667	0.5238
dot_precision@5	0.124	0.288	0.12	0.088	0.448	0.132	0.172	0.264	0.136	0.208	0.056	0.124	0.5184
dot_precision@10	0.07	0.234	0.076	0.06	0.38	0.082	0.1	0.144	0.076	0.146	0.034	0.072	0.4551
dot_recall@1	0.32	0.0209	0.3	0.0883	0.057	0.3667	0.1669	0.32	0.3067	0.0687	0.08	0.335	0.0412
dot_recall@3	0.52	0.0574	0.52	0.1383	0.1262	0.5467	0.2887	0.58	0.5107	0.1577	0.26	0.44	0.106
dot_recall@5	0.62	0.0743	0.57	0.189	0.1934	0.6267	0.3915	0.66	0.6107	0.2147	0.28	0.525	0.1742
dot_recall@10	0.7	0.0975	0.7	0.2473	0.2774	0.7667	0.453	0.72	0.6907	0.3007	0.34	0.61	0.2958
dot_ndcg@10	0.5083	0.2686	0.4982	0.199	0.4762	0.5583	0.3614	0.644	0.5024	0.2837	0.2133	0.474	0.5017
dot_mrr@10	0.4472	0.4113	0.4434	0.2888	0.6898	0.5038	0.4162	0.7276	0.4542	0.4724	0.1726	0.4374	0.7473
dot_map@100	0.463	0.1047	0.4411	0.157	0.3606	0.4951	0.2968	0.5729	0.4558	0.207	0.1878	0.4391	0.3595
query_active_dims	94.44	117.86	103.14	322.94	127.16	341.5	83.8	181.72	68.72	210.06	664.0	371.0	45.4082
query_sparsity_ratio	0.9969	0.9961	0.9966	0.9894	0.9958	0.9888	0.9973	0.994	0.9977	0.9931	0.9782	0.9878	0.9985
corpus_active_dims	430.5701	741.1876	552.1694	559.9041	502.0566	675.0518	415.0838	580.4144	79.2295	610.8639	613.4025	683.7307	522.2585
corpus_sparsity_ratio	0.9859	0.9757	0.9819	0.9817	0.9836	0.9779	0.9864	0.981	0.9974	0.98	0.9799	0.9776	0.9829

Sparse Nano BEIR

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

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

Metric	Value
dot_accuracy@1	0.3067
dot_accuracy@3	0.5333
dot_accuracy@5	0.5933
dot_accuracy@10	0.6933
dot_precision@1	0.3067
dot_precision@3	0.2244
dot_precision@5	0.1773
dot_precision@10	0.1267
dot_recall@1	0.2136
dot_recall@3	0.3658
dot_recall@5	0.4214
dot_recall@10	0.4992
dot_ndcg@10	0.425
dot_mrr@10	0.434
dot_map@100	0.3363
query_active_dims	105.1467
query_sparsity_ratio	0.9966
corpus_active_dims	547.9445
corpus_sparsity_ratio	0.982

Sparse Nano BEIR

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

  {
      "dataset_names": [
          "climatefever",
          "dbpedia",
          "fever",
          "fiqa2018",
          "hotpotqa",
          "msmarco",
          "nfcorpus",
          "nq",
          "quoraretrieval",
          "scidocs",
          "arguana",
          "scifact",
          "touche2020"
      ]
  }
  

Metric	Value
dot_accuracy@1	0.3656
dot_accuracy@3	0.552
dot_accuracy@5	0.6338
dot_accuracy@10	0.7169
dot_precision@1	0.3656
dot_precision@3	0.2505
dot_precision@5	0.206
dot_precision@10	0.1484
dot_recall@1	0.1901
dot_recall@3	0.327
dot_recall@5	0.3946
dot_recall@10	0.4769
dot_ndcg@10	0.4222
dot_mrr@10	0.4778
dot_map@100	0.3493
query_active_dims	210.3883
query_sparsity_ratio	0.9931
corpus_active_dims	515.9202
corpus_sparsity_ratio	0.9831

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: 11.79 tokens max: 24 tokens	min: 14 tokens mean: 60.02 tokens max: 153 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')",
      "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: 11.93 tokens max: 25 tokens	min: 14 tokens mean: 60.84 tokens max: 127 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')",
      "document_regularizer_weight": 3e-05,
      "query_regularizer_weight": 5e-05
  }
  

Training Hyperparameters

Non-Default Hyperparameters

• eval_strategy: steps
• per_device_train_batch_size: 32
• per_device_eval_batch_size: 32
• learning_rate: 2e-05
• num_train_epochs: 1
• bf16: True
• load_best_model_at_end: 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: 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: 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: 0.0
• 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: True
• fp16: False
• 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
• 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	NanoClimateFEVER_dot_ndcg@10	NanoDBPedia_dot_ndcg@10	NanoFEVER_dot_ndcg@10	NanoFiQA2018_dot_ndcg@10	NanoHotpotQA_dot_ndcg@10	NanoQuoraRetrieval_dot_ndcg@10	NanoSCIDOCS_dot_ndcg@10	NanoArguAna_dot_ndcg@10	NanoSciFact_dot_ndcg@10	NanoTouche2020_dot_ndcg@10
0.0323	100	28.28	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.0646	200	0.4861	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.0970	300	0.2953	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.1293	400	0.2281	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.1616	500	0.2219	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.1939	600	0.1677	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.1972	610	-	0.1561	0.5037	0.2269	0.4528	0.3945	-	-	-	-	-	-	-	-	-	-
0.2262	700	0.1841	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.2586	800	0.1578	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.2909	900	0.1403	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.3232	1000	0.1738	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.3555	1100	0.1453	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.3878	1200	0.138	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.3943	1220	-	0.1417	0.5052	0.2372	0.4614	0.4013	-	-	-	-	-	-	-	-	-	-
0.4202	1300	0.1276	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.4525	1400	0.1376	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.4848	1500	0.1364	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.5171	1600	0.1174	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.5495	1700	0.1107	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.5818	1800	0.1219	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.5915	1830	-	0.0817	0.5298	0.2767	0.4649	0.4238	-	-	-	-	-	-	-	-	-	-
0.6141	1900	0.1012	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.6464	2000	0.1279	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.6787	2100	0.1057	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.7111	2200	0.1276	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.7434	2300	0.1154	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.7757	2400	0.0964	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.7886	2440	-	0.0926	0.5132	0.2587	0.4951	0.4223	-	-	-	-	-	-	-	-	-	-
0.8080	2500	0.0963	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.8403	2600	0.1056	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.8727	2700	0.0969	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.9050	2800	0.0988	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.9373	2900	0.0927	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.9696	3000	0.0856	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.9858	3050	-	0.0783	0.5083	0.2686	0.4982	0.425	-	-	-	-	-	-	-	-	-	-
-1	-1	-	-	0.5083	0.2686	0.4982	0.4222	0.1990	0.4762	0.5583	0.3614	0.6440	0.5024	0.2837	0.2133	0.4740	0.5017

• The bold row denotes the saved checkpoint.

Environmental Impact

Carbon emissions were measured using CodeCarbon.

• Energy Consumed: 0.045 kWh
• Carbon Emitted: 0.017 kg of CO2
• Hours Used: 0.193 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: 2.7.1+cu126
• 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},
}

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}
}