tomaarsen/inference-free-splade-bert-tiny-nq

Inference-free SPLADE BERT-tiny trained on Natural-Questions tuples

This is a Asymmetric Inference-free SPLADE Sparse Encoder model trained on the natural-questions 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: Asymmetric Inference-free SPLADE Sparse Encoder
• Maximum Sequence Length: 512 tokens
• Output Dimensionality: 30522 dimensions
• Similarity Function: Dot Product
• Training Dataset:
  • natural-questions
• 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): Asym(
    (query_0_IDF): IDF ({'frozen': False}, dim:30522, tokenizer: BertTokenizerFast)
    (corpus_0_MLMTransformer): MLMTransformer({'max_seq_length': 512, 'do_lower_case': False}) with MLMTransformer model: BertForMaskedLM 
    (corpus_1_SpladePooling): 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/inference-free-splade-bert-tiny-nq")
# Run inference
sentences = [
    'when did the american civil rights movement end',
    'African-American civil rights movement (1954–1968) The Civil Rights Movement (also known as the American civil rights movement, African-American civil rights movement, and other terms,[b]) was a human rights movement from 1954–1968 that encompassed strategies, groups, and social movements to accomplish its goal of ending legalized racial segregation and discrimination laws in the United States. The movement secured the legal recognition and federal protection of black Americans in the United States Constitution and federal law.',
    'Paleolithic Paleolithic humans made tools of stone, bone, and wood.[23] The early paleolithic hominins, Australopithecus, were the first users of stone tools. Excavations in Gona, Ethiopia have produced thousands of artifacts, and through radioisotopic dating and magnetostratigraphy, the sites can be firmly dated to 2.6Â\xa0million years ago. Evidence shows these early hominins intentionally selected raw materials with good flaking qualities and chose appropriate sized stones for their needs to produce sharp-edged tools for cutting.[29]',
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# (3, 30522)

# Get the similarity scores for the embeddings
similarities = model.similarity(embeddings, embeddings)
print(similarities.shape)
# [3, 3]

Evaluation

Metrics

Sparse Information Retrieval

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

Metric	NanoMSMARCO	NanoNFCorpus	NanoNQ
dot_accuracy@1	0.26	0.46	0.32
dot_accuracy@3	0.5	0.6	0.58
dot_accuracy@5	0.64	0.6	0.7
dot_accuracy@10	0.72	0.64	0.78
dot_precision@1	0.26	0.46	0.32
dot_precision@3	0.1667	0.3733	0.1933
dot_precision@5	0.128	0.312	0.14
dot_precision@10	0.072	0.248	0.08
dot_recall@1	0.26	0.0433	0.31
dot_recall@3	0.5	0.0769	0.56
dot_recall@5	0.64	0.0911	0.66
dot_recall@10	0.72	0.1147	0.73
dot_ndcg@10	0.4817	0.3172	0.5286
dot_mrr@10	0.4057	0.5278	0.4738
dot_map@100	0.4174	0.1363	0.4673

Sparse Nano BEIR

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

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

Metric	Value
dot_accuracy@1	0.3467
dot_accuracy@3	0.56
dot_accuracy@5	0.6467
dot_accuracy@10	0.7133
dot_precision@1	0.3467
dot_precision@3	0.2444
dot_precision@5	0.1933
dot_precision@10	0.1333
dot_recall@1	0.2044
dot_recall@3	0.379
dot_recall@5	0.4637
dot_recall@10	0.5216
dot_ndcg@10	0.4425
dot_mrr@10	0.4691
dot_map@100	0.3404

Training Details

Training Dataset

natural-questions

• Dataset: natural-questions at f9e894e
• Size: 99,000 training samples
• Columns: query and corpus
• Approximate statistics based on the first 1000 samples:

  	query	corpus
  type	dict	dict
  details

• Samples:

  query	corpus
  {'query': "who played the father in papa don't preach"}	{'corpus': 'Alex McArthur Alex McArthur (born March 6, 1957) is an American actor.'}
  {'query': 'where was the location of the battle of hastings'}	{'corpus': 'Battle of Hastings The Battle of Hastings[a] was fought on 14 October 1066 between the Norman-French army of William, the Duke of Normandy, and an English army under the Anglo-Saxon King Harold Godwinson, beginning the Norman conquest of England. It took place approximately 7 miles (11 kilometres) northwest of Hastings, close to the present-day town of Battle, East Sussex, and was a decisive Norman victory.'}
  {'query': 'how many puppies can a dog give birth to'}	{'corpus': 'Canine reproduction The largest litter size to date was set by a Neapolitan Mastiff in Manea, Cambridgeshire, UK on November 29, 2004; the litter was 24 puppies.[22]'}

• Loss: SpladeLoss with these parameters:

  {'loss': SparseMultipleNegativesRankingLoss(
    (model): SparseEncoder(
      (0): Asym(
        (query_0_IDF): IDF ({'frozen': False}, dim:30522, tokenizer: BertTokenizerFast)
        (corpus_0_MLMTransformer): MLMTransformer({'max_seq_length': 512, 'do_lower_case': False}) with MLMTransformer model: BertForMaskedLM 
        (corpus_1_SpladePooling): SpladePooling({'pooling_strategy': 'max', 'activation_function': 'relu', 'word_embedding_dimension': 30522})
      )
    )
    (cross_entropy_loss): CrossEntropyLoss()
  ), 'lambda_corpus': 3e-05, 'lambda_query': 0, 'corpus_regularizer': FlopsLoss(
    (model): SparseEncoder(
      (0): Asym(
        (query_0_IDF): IDF ({'frozen': False}, dim:30522, tokenizer: BertTokenizerFast)
        (corpus_0_MLMTransformer): MLMTransformer({'max_seq_length': 512, 'do_lower_case': False}) with MLMTransformer model: BertForMaskedLM 
        (corpus_1_SpladePooling): SpladePooling({'pooling_strategy': 'max', 'activation_function': 'relu', 'word_embedding_dimension': 30522})
      )
    )
  ), 'query_regularizer': None}
  

Evaluation Dataset

natural-questions

• Dataset: natural-questions at f9e894e
• Size: 1,000 evaluation samples
• Columns: query and corpus
• Approximate statistics based on the first 1000 samples:

  	query	corpus
  type	dict	dict
  details

• Samples:

  query	corpus
  {'query': 'where is the tiber river located in italy'}	{'corpus': 'Tiber The Tiber (/ˈtaɪbər/, Latin: Tiberis,[1] Italian: Tevere [ˈteːvere])[2] is the third-longest river in Italy, rising in the Apennine Mountains in Emilia-Romagna and flowing 406 kilometres (252\xa0mi) through Tuscany, Umbria and Lazio, where it is joined by the river Aniene, to the Tyrrhenian Sea, between Ostia and Fiumicino.[3] It drains a basin estimated at 17,375 square kilometres (6,709\xa0sq\xa0mi). The river has achieved lasting fame as the main watercourse of the city of Rome, founded on its eastern banks.'}
  {'query': 'what kind of car does jay gatsby drive'}	{'corpus': 'Jay Gatsby At the Buchanan home, Jordan Baker, Nick, Jay, and the Buchanans decide to visit New York City. Tom borrows Gatsby's yellow Rolls Royce to drive up to the city. On the way to New York City, Tom makes a detour at a gas station in "the Valley of Ashes", a run-down part of Long Island. The owner, George Wilson, shares his concern that his wife, Myrtle, may be having an affair. This unnerves Tom, who has been having an affair with Myrtle, and he leaves in a hurry.'}
  {'query': 'who sings if i can dream about you'}	{'corpus': 'I Can Dream About You "I Can Dream About You" is a song performed by American singer Dan Hartman on the soundtrack album of the film Streets of Fire. Released in 1984 as a single from the soundtrack, and included on Hartman's album I Can Dream About You, it reached number 6 on the Billboard Hot 100.[1]'}

• Loss: SpladeLoss with these parameters:

  {'loss': SparseMultipleNegativesRankingLoss(
    (model): SparseEncoder(
      (0): Asym(
        (query_0_IDF): IDF ({'frozen': False}, dim:30522, tokenizer: BertTokenizerFast)
        (corpus_0_MLMTransformer): MLMTransformer({'max_seq_length': 512, 'do_lower_case': False}) with MLMTransformer model: BertForMaskedLM 
        (corpus_1_SpladePooling): SpladePooling({'pooling_strategy': 'max', 'activation_function': 'relu', 'word_embedding_dimension': 30522})
      )
    )
    (cross_entropy_loss): CrossEntropyLoss()
  ), 'lambda_corpus': 3e-05, 'lambda_query': 0, 'corpus_regularizer': FlopsLoss(
    (model): SparseEncoder(
      (0): Asym(
        (query_0_IDF): IDF ({'frozen': False}, dim:30522, tokenizer: BertTokenizerFast)
        (corpus_0_MLMTransformer): MLMTransformer({'max_seq_length': 512, 'do_lower_case': False}) with MLMTransformer model: BertForMaskedLM 
        (corpus_1_SpladePooling): SpladePooling({'pooling_strategy': 'max', 'activation_function': 'relu', 'word_embedding_dimension': 30522})
      )
    )
  ), 'query_regularizer': None}
  

Training Hyperparameters

Non-Default Hyperparameters

• eval_strategy: steps
• per_device_train_batch_size: 64
• per_device_eval_batch_size: 64
• 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: 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: 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
• dispatch_batches: None
• split_batches: 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

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
0.0129	20	1.5412	-	-	-	-	-
0.0259	40	1.4839	-	-	-	-	-
0.0388	60	1.3033	-	-	-	-	-
0.0517	80	0.9905	-	-	-	-	-
0.0646	100	0.6821	-	-	-	-	-
0.0776	120	0.5529	-	-	-	-	-
0.0905	140	0.4829	-	-	-	-	-
0.1034	160	0.4292	-	-	-	-	-
0.1164	180	0.4654	-	-	-	-	-
0.1293	200	0.4169	0.3483	0.4863	0.3131	0.4902	0.4299
0.1422	220	0.3911	-	-	-	-	-
0.1551	240	0.3671	-	-	-	-	-
0.1681	260	0.3954	-	-	-	-	-
0.1810	280	0.355	-	-	-	-	-
0.1939	300	0.3433	-	-	-	-	-
0.2069	320	0.344	-	-	-	-	-
0.2198	340	0.3423	-	-	-	-	-
0.2327	360	0.3445	-	-	-	-	-
0.2456	380	0.356	-	-	-	-	-
0.2586	400	0.3209	0.3060	0.5232	0.3098	0.5000	0.4443
0.2715	420	0.322	-	-	-	-	-
0.2844	440	0.3328	-	-	-	-	-
0.2973	460	0.3181	-	-	-	-	-
0.3103	480	0.3547	-	-	-	-	-
0.3232	500	0.3566	-	-	-	-	-
0.3361	520	0.3332	-	-	-	-	-
0.3491	540	0.3269	-	-	-	-	-
0.3620	560	0.3691	-	-	-	-	-
0.3749	580	0.3167	-	-	-	-	-
0.3878	600	0.3734	0.2931	0.5081	0.3131	0.5168	0.4460
0.4008	620	0.314	-	-	-	-	-
0.4137	640	0.2801	-	-	-	-	-
0.4266	660	0.316	-	-	-	-	-
0.4396	680	0.3174	-	-	-	-	-
0.4525	700	0.3015	-	-	-	-	-
0.4654	720	0.2642	-	-	-	-	-
0.4783	740	0.2889	-	-	-	-	-
0.4913	760	0.3005	-	-	-	-	-
0.5042	780	0.2702	-	-	-	-	-
0.5171	800	0.3116	0.2754	0.4961	0.3178	0.5150	0.4429
0.5301	820	0.2941	-	-	-	-	-
0.5430	840	0.2726	-	-	-	-	-
0.5559	860	0.3007	-	-	-	-	-
0.5688	880	0.282	-	-	-	-	-
0.5818	900	0.27	-	-	-	-	-
0.5947	920	0.2797	-	-	-	-	-
0.6076	940	0.2739	-	-	-	-	-
0.6206	960	0.2905	-	-	-	-	-
0.6335	980	0.2986	-	-	-	-	-
0.6464	1000	0.2332	0.2641	0.4942	0.3144	0.5403	0.4496
0.6593	1020	0.2876	-	-	-	-	-
0.6723	1040	0.2722	-	-	-	-	-
0.6852	1060	0.2712	-	-	-	-	-
0.6981	1080	0.283	-	-	-	-	-
0.7111	1100	0.2835	-	-	-	-	-
0.7240	1120	0.2795	-	-	-	-	-
0.7369	1140	0.2709	-	-	-	-	-
0.7498	1160	0.2855	-	-	-	-	-
0.7628	1180	0.2871	-	-	-	-	-
0.7757	1200	0.283	0.2573	0.4918	0.3159	0.5315	0.4464
0.7886	1220	0.2599	-	-	-	-	-
0.8016	1240	0.2843	-	-	-	-	-
0.8145	1260	0.2552	-	-	-	-	-
0.8274	1280	0.2913	-	-	-	-	-
0.8403	1300	0.279	-	-	-	-	-
0.8533	1320	0.2704	-	-	-	-	-
0.8662	1340	0.2628	-	-	-	-	-
0.8791	1360	0.2579	-	-	-	-	-
0.8920	1380	0.2554	-	-	-	-	-
0.9050	1400	0.259	0.2548	0.4891	0.3174	0.5247	0.4437
0.9179	1420	0.2609	-	-	-	-	-
0.9308	1440	0.2325	-	-	-	-	-
0.9438	1460	0.2817	-	-	-	-	-
0.9567	1480	0.2807	-	-	-	-	-
0.9696	1500	0.2744	-	-	-	-	-
0.9825	1520	0.2795	-	-	-	-	-
0.9955	1540	0.2706	-	-	-	-	-
-1	-1	-	-	0.4817	0.3172	0.5286	0.4425

Environmental Impact

Carbon emissions were measured using CodeCarbon.

• Energy Consumed: 0.031 kWh
• Carbon Emitted: 0.012 kg of CO2
• Hours Used: 0.096 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: 4.49.0
• 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},
}