tomaarsen/inference-free-splade-distilbert-base-uncased-nq-3e-4-lc

Inference-free SPLADE distilbert-base-uncased 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): Router(
    (query_0_IDF): IDF ({'frozen': False}, dim:30522, tokenizer: DistilBertTokenizerFast)
    (document_0_MLMTransformer): MLMTransformer({'max_seq_length': 512, 'do_lower_case': False}) with MLMTransformer model: DistilBertForMaskedLM 
    (document_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-distilbert-base-uncased-nq-3e-4-lc")
# Run inference
queries = [
    "is send in the clowns from a musical",
]
documents = [
    'Send In the Clowns "Send In the Clowns" is a song written by Stephen Sondheim for the 1973 musical A Little Night Music, an adaptation of Ingmar Bergman\'s film Smiles of a Summer Night. It is a ballad from Act Two, in which the character Desirée reflects on the ironies and disappointments of her life. Among other things, she looks back on an affair years earlier with the lawyer Fredrik, who was deeply in love with her but whose marriage proposals she had rejected. Meeting him after so long, she realizes she is in love with him and finally ready to marry him, but now it is he who rejects her: he is in an unconsummated marriage with a much younger woman. Desirée proposes marriage to rescue him from this situation, but he declines, citing his dedication to his bride. Reacting to his rejection, Desirée sings this song. The song is later reprised as a coda after Fredrik\'s young wife runs away with his son, and Fredrik is finally free to accept Desirée\'s offer.[1]',
    'The Suite Life on Deck The Suite Life on Deck is an American sitcom that aired on Disney Channel from September 26, 2008 to May 6, 2011. It is a sequel/spin-off of the Disney Channel Original Series The Suite Life of Zack & Cody. The series follows twin brothers Zack and Cody Martin and hotel heiress London Tipton in a new setting, the SS Tipton, where they attend classes at "Seven Seas High School" and meet Bailey Pickett while Mr. Moseby manages the ship. The ship travels around the world to nations such as Italy, France, Greece, India, Sweden and the United Kingdom where the characters experience different cultures, adventures, and situations.[1]',
    'Money in the Bank ladder match The first match was contested in 2005 at WrestleMania 21, after being invented (in kayfabe) by Chris Jericho.[1] At the time, it was exclusive to wrestlers of the Raw brand, and Edge won the inaugural match.[1] From then until 2010, the Money in the Bank ladder match, now open to all WWE brands, became a WrestleMania mainstay. 2010 saw a second and third Money in the Bank ladder match when the Money in the Bank pay-per-view debuted in July. Unlike the matches at WrestleMania, this new event featured two such ladder matches – one each for a contract for the WWE Championship and World Heavyweight Championship, respectively.',
]
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([[11.2000,  0.0000,  0.6561]])

Evaluation

Metrics

Sparse Information Retrieval

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

Metric	NanoMSMARCO	NanoNFCorpus	NanoNQ
dot_accuracy@1	0.26	0.4	0.4
dot_accuracy@3	0.54	0.52	0.62
dot_accuracy@5	0.6	0.54	0.74
dot_accuracy@10	0.84	0.64	0.8
dot_precision@1	0.26	0.4	0.4
dot_precision@3	0.18	0.38	0.2133
dot_precision@5	0.12	0.32	0.152
dot_precision@10	0.084	0.272	0.084
dot_recall@1	0.26	0.0225	0.38
dot_recall@3	0.54	0.0774	0.58
dot_recall@5	0.6	0.0924	0.69
dot_recall@10	0.84	0.1237	0.76
dot_ndcg@10	0.5301	0.3246	0.5795
dot_mrr@10	0.4347	0.4722	0.5381
dot_map@100	0.4418	0.1354	0.5206
query_active_dims	7.22	5.66	10.32
query_sparsity_ratio	0.9998	0.9998	0.9997
corpus_active_dims	108.2734	131.0975	91.8959
corpus_sparsity_ratio	0.9965	0.9957	0.997

Sparse Nano BEIR

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

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

Metric	Value
dot_accuracy@1	0.3533
dot_accuracy@3	0.56
dot_accuracy@5	0.6267
dot_accuracy@10	0.76
dot_precision@1	0.3533
dot_precision@3	0.2578
dot_precision@5	0.1973
dot_precision@10	0.1467
dot_recall@1	0.2208
dot_recall@3	0.3991
dot_recall@5	0.4608
dot_recall@10	0.5746
dot_ndcg@10	0.4781
dot_mrr@10	0.4816
dot_map@100	0.3659
query_active_dims	7.7333
query_sparsity_ratio	0.9997
corpus_active_dims	107.1176
corpus_sparsity_ratio	0.9965

Training Details

Training Dataset

natural-questions

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

  	query	answer
  type	string	string
  details	min: 10 tokens mean: 11.71 tokens max: 26 tokens	min: 4 tokens mean: 131.81 tokens max: 450 tokens

• Samples:

  query	answer
  who played the father in papa don't preach	Alex McArthur Alex McArthur (born March 6, 1957) is an American actor.
  where was the location of the battle of hastings	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.
  how many puppies can a dog give birth to	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(scale=1.0, similarity_fct='dot_score')",
      "lambda_corpus": 0.0003,
      "lambda_query": 0
  }
  

Evaluation Dataset

natural-questions

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

  	query	answer
  type	string	string
  details	min: 10 tokens mean: 11.69 tokens max: 23 tokens	min: 15 tokens mean: 134.01 tokens max: 512 tokens

• Samples:

  query	answer
  where is the tiber river located in italy	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 mi) 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 sq mi). The river has achieved lasting fame as the main watercourse of the city of Rome, founded on its eastern banks.
  what kind of car does jay gatsby drive	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.
  who sings if i can dream about you	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(scale=1.0, similarity_fct='dot_score')",
      "lambda_corpus": 0.0003,
      "lambda_query": 0
  }
  

Training Hyperparameters

Non-Default Hyperparameters

• eval_strategy: steps
• per_device_train_batch_size: 16
• per_device_eval_batch_size: 16
• learning_rate: 2e-05
• warmup_ratio: 0.1
• fp16: True
• batch_sampler: no_duplicates
• router_mapping: {'query': 'query', 'answer': 'document'}
• learning_rate_mapping: {'IDF\.weight': 0.001}

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: 3
• 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: {'query': 'query', 'answer': 'document'}
• learning_rate_mapping: {'IDF\.weight': 0.001}

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
0.0323	200	0.3026	-	-	-	-	-
0.0646	400	0.1058	-	-	-	-	-
0.0970	600	0.0605	-	-	-	-	-
0.1293	800	0.0443	-	-	-	-	-
0.1616	1000	0.0385	0.0598	0.6070	0.3234	0.5837	0.5047
0.1939	1200	0.0391	-	-	-	-	-
0.2262	1400	0.0432	-	-	-	-	-
0.2586	1600	0.0448	-	-	-	-	-
0.2909	1800	0.0363	-	-	-	-	-
0.3232	2000	0.0348	0.0578	0.5471	0.3245	0.5586	0.4768
0.3555	2200	0.0383	-	-	-	-	-
0.3878	2400	0.0378	-	-	-	-	-
0.4202	2600	0.0354	-	-	-	-	-
0.4525	2800	0.0348	-	-	-	-	-
0.4848	3000	0.0276	0.0474	0.5432	0.3288	0.5474	0.4731
0.5171	3200	0.0341	-	-	-	-	-
0.5495	3400	0.0352	-	-	-	-	-
0.5818	3600	0.0312	-	-	-	-	-
0.6141	3800	0.036	-	-	-	-	-
0.6464	4000	0.0295	0.0543	0.5880	0.3372	0.5392	0.4881
0.6787	4200	0.032	-	-	-	-	-
0.7111	4400	0.0321	-	-	-	-	-
0.7434	4600	0.0322	-	-	-	-	-
0.7757	4800	0.0303	-	-	-	-	-
0.8080	5000	0.0328	0.0476	0.5662	0.3225	0.5796	0.4894
0.8403	5200	0.0322	-	-	-	-	-
0.8727	5400	0.03	-	-	-	-	-
0.9050	5600	0.0306	-	-	-	-	-
0.9373	5800	0.0273	-	-	-	-	-
0.9696	6000	0.0327	0.0474	0.5383	0.3334	0.5084	0.4600
1.0019	6200	0.0342	-	-	-	-	-
1.0343	6400	0.0201	-	-	-	-	-
1.0666	6600	0.0212	-	-	-	-	-
1.0989	6800	0.0209	-	-	-	-	-
1.1312	7000	0.0319	0.0484	0.5389	0.3125	0.5364	0.4626
1.1635	7200	0.0243	-	-	-	-	-
1.1959	7400	0.0219	-	-	-	-	-
1.2282	7600	0.022	-	-	-	-	-
1.2605	7800	0.0237	-	-	-	-	-
1.2928	8000	0.0257	0.0461	0.5594	0.3314	0.5339	0.4749
1.3251	8200	0.0152	-	-	-	-	-
1.3575	8400	0.0177	-	-	-	-	-
1.3898	8600	0.0228	-	-	-	-	-
1.4221	8800	0.0197	-	-	-	-	-
1.4544	9000	0.025	0.0416	0.5534	0.3253	0.5838	0.4875
1.4867	9200	0.025	-	-	-	-	-
1.5191	9400	0.0229	-	-	-	-	-
1.5514	9600	0.0198	-	-	-	-	-
1.5837	9800	0.022	-	-	-	-	-
1.6160	10000	0.0279	0.0434	0.5778	0.3390	0.5574	0.4914
1.6484	10200	0.0201	-	-	-	-	-
1.6807	10400	0.0196	-	-	-	-	-
1.7130	10600	0.0188	-	-	-	-	-
1.7453	10800	0.0207	-	-	-	-	-
1.7776	11000	0.0194	0.0446	0.5603	0.3301	0.5776	0.4893
1.8100	11200	0.0166	-	-	-	-	-
1.8423	11400	0.0207	-	-	-	-	-
1.8746	11600	0.0212	-	-	-	-	-
1.9069	11800	0.0172	-	-	-	-	-
1.9392	12000	0.0198	0.0451	0.5653	0.3344	0.5716	0.4904
1.9716	12200	0.0183	-	-	-	-	-
2.0039	12400	0.0212	-	-	-	-	-
2.0362	12600	0.0111	-	-	-	-	-
2.0685	12800	0.0144	-	-	-	-	-
2.1008	13000	0.0124	0.0412	0.5408	0.3278	0.5869	0.4852
2.1332	13200	0.014	-	-	-	-	-
2.1655	13400	0.0162	-	-	-	-	-
2.1978	13600	0.0118	-	-	-	-	-
2.2301	13800	0.0138	-	-	-	-	-
2.2624	14000	0.0137	0.0404	0.5265	0.3313	0.5875	0.4817
2.2948	14200	0.0147	-	-	-	-	-
2.3271	14400	0.0134	-	-	-	-	-
2.3594	14600	0.0137	-	-	-	-	-
2.3917	14800	0.0119	-	-	-	-	-
2.4240	15000	0.0139	0.0409	0.5140	0.3347	0.5665	0.4717
2.4564	15200	0.0135	-	-	-	-	-
2.4887	15400	0.0169	-	-	-	-	-
2.5210	15600	0.0123	-	-	-	-	-
2.5533	15800	0.0147	-	-	-	-	-
2.5856	16000	0.0154	0.0388	0.5210	0.3278	0.5713	0.4734
2.6180	16200	0.0117	-	-	-	-	-
2.6503	16400	0.0125	-	-	-	-	-
2.6826	16600	0.0137	-	-	-	-	-
2.7149	16800	0.0143	-	-	-	-	-
2.7473	17000	0.0118	0.0410	0.5345	0.3306	0.5739	0.4797
2.7796	17200	0.0123	-	-	-	-	-
2.8119	17400	0.0105	-	-	-	-	-
2.8442	17600	0.0139	-	-	-	-	-
2.8765	17800	0.0138	-	-	-	-	-
2.9089	18000	0.0122	0.0396	0.5250	0.3235	0.5880	0.4788
2.9412	18200	0.0114	-	-	-	-	-
2.9735	18400	0.0131	-	-	-	-	-
-1	-1	-	-	0.5301	0.3246	0.5795	0.4781

Environmental Impact

Carbon emissions were measured using CodeCarbon.

• Energy Consumed: 0.249 kWh
• Carbon Emitted: 0.097 kg of CO2
• Hours Used: 0.63 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},
}

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}
    }