tomaarsen/splade-cocondenser-selfdistil-msmarco-kldiv-marginmse-minilm-temp-2

Splade++ SelfDistil finetuned on MS MARCO

This is a SPLADE Sparse Encoder model finetuned from naver/splade-cocondenser-selfdistil 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: naver/splade-cocondenser-selfdistil
• 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, 'architecture': '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-selfdistil-msmarco-kldiv-marginmse-minilm-temp-2")
# Run inference
queries = [
    "who started gladiator lacrosse",
]
documents = [
    'Weed Eater was a string trimmer company founded in 1971 in Houston, Texas by George C. Ballas, Sr. , the inventor of the device. The idea for the Weed Eater trimmer came to him from the spinning nylon bristles of an automatic car wash.He thought that he could come up with a similar technique to protect the bark on trees that he was trimming around. His company was eventually bought by Emerson Electric and merged with Poulan.Poulan/Weed Eater was later purchased by Electrolux, which spun off the outdoors division as Husqvarna AB in 2006.Inventor Ballas was the father of champion ballroom dancer Corky Ballas and the grandfather of Dancing with the Stars dancer Mark Ballas.George Ballas died on June 25, 2011.he idea for the Weed Eater trimmer came to him from the spinning nylon bristles of an automatic car wash. He thought that he could come up with a similar technique to protect the bark on trees that he was trimming around. His company was eventually bought by Emerson Electric and merged with Poulan.',
    "The earliest types of gladiator were named after Rome's enemies of that time: the Samnite, Thracian and Gaul. The Samnite, heavily armed, elegantly helmed and probably the most popular type, was renamed Secutor and the Gaul renamed Murmillo, once these former enemies had been conquered then absorbed into Rome's Empire.",
    'Summit Hill, PA. Sponsored Topics. Summit Hill is a borough in Carbon County, Pennsylvania, United States. The population was 2,974 at the 2000 census. Summit Hill is located at 40°49â\x80²39â\x80³N 75°51â\x80²57â\x80³W / 40.8275°N 75.86583°W / 40.8275; -75.86583 (40.827420, -75.865892).',
]
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([[11.6863, 18.5272,  0.5780]])

Evaluation

Metrics

Sparse Information Retrieval

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

Metric	NanoMSMARCO	NanoNFCorpus	NanoNQ
dot_accuracy@1	0.48	0.44	0.5
dot_accuracy@3	0.72	0.6	0.74
dot_accuracy@5	0.76	0.64	0.86
dot_accuracy@10	0.88	0.66	0.9
dot_precision@1	0.48	0.44	0.5
dot_precision@3	0.24	0.42	0.2533
dot_precision@5	0.152	0.368	0.18
dot_precision@10	0.088	0.268	0.096
dot_recall@1	0.48	0.0439	0.48
dot_recall@3	0.72	0.1003	0.7
dot_recall@5	0.76	0.1206	0.81
dot_recall@10	0.88	0.14	0.86
dot_ndcg@10	0.6789	0.3503	0.6881
dot_mrr@10	0.6152	0.5247	0.6457
dot_map@100	0.6237	0.1644	0.6259
query_active_dims	51.3	47.88	56.98
query_sparsity_ratio	0.9983	0.9984	0.9981
corpus_active_dims	457.3064	851.705	527.3356
corpus_sparsity_ratio	0.985	0.9721	0.9827

Sparse Nano BEIR

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

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

Metric	Value
dot_accuracy@1	0.4733
dot_accuracy@3	0.6867
dot_accuracy@5	0.7533
dot_accuracy@10	0.8133
dot_precision@1	0.4733
dot_precision@3	0.3044
dot_precision@5	0.2333
dot_precision@10	0.1507
dot_recall@1	0.3346
dot_recall@3	0.5068
dot_recall@5	0.5635
dot_recall@10	0.6267
dot_ndcg@10	0.5724
dot_mrr@10	0.5952
dot_map@100	0.4714
query_active_dims	52.0533
query_sparsity_ratio	0.9983
corpus_active_dims	573.7408
corpus_sparsity_ratio	0.9812

Training Details

Training Dataset

Unnamed Dataset

• Size: 99,000 training samples
• Columns: query, positive, negative, and label
• Approximate statistics based on the first 1000 samples:

  	query	positive	negative	label
  type	string	string	string	list
  details	min: 4 tokens mean: 9.2 tokens max: 34 tokens	min: 18 tokens mean: 79.86 tokens max: 219 tokens	min: 18 tokens mean: 79.96 tokens max: 270 tokens	size: 2 elements

• Samples:

  query	positive	negative	label
  rtn tv network	Home Shopping Network. Home Shopping Network (HSN) is an American broadcast, basic cable and satellite television network that is owned by HSN, Inc. (NASDAQ: HSNI), which also owns catalog company Cornerstone Brands. Based in St. Petersburg, Florida, United States, the home shopping channel has former and current sister channels in several other countries.	The Public Switched Telephone Network - The public switched telephone network (PSTN) is the international network of circuit-switched telephones. Learn more about PSTN at HowStuffWorks. x	[-1.0804121494293213, -5.908488750457764]
  how did president nixon react to the watergate investigation?	The Watergate scandal was a major political scandal that occurred in the United States during the early 1970s, following a break-in by five men at the Democratic National Committee headquarters at the Watergate office complex in Washington, D.C. on June 17, 1972, and President Richard Nixon's administration's subsequent attempt to cover up its involvement. After the five burglars were caught and the conspiracy was discovered, Watergate was investigated by the United States Congress. Meanwhile, N	The release of the tape was ordered by the Supreme Court on July 24, 1974, in a case known as United States v. Nixon. The court’s decision was unanimous. President Nixon released the tape on August 5. It was one of three conversations he had with Haldeman six days after the Watergate break-in. The tapes prove that he ordered a cover-up of the Watergate burglary. The Smoking Gun tape reveals that Nixon ordered the FBI to abandon its investigation of the break-in. [Read more…]	[4.117279052734375, 3.191757917404175]
  what is a summary offense in pennsylvania	We provide cost effective house arrest and electronic monitoring services to magisterial district court systems throughout Pennsylvania including York, Harrisburg, Philadelphia and Allentown.In addition, we also serve the York County, Lancaster County and Chester County.e provide cost effective house arrest and electronic monitoring services to magisterial district court systems throughout Pennsylvania including York, Harrisburg, Philadelphia and Allentown.	In order to be convicted of Simple Assault, one must cause bodily injury. To be convicted of Aggravated Assault, one must cause serious bodily injury. From my research, Pennsylvania law defines bodily injury as the impairment of physical condition or substantial pain.	[-8.954689025878906, -1.3361705541610718]

• Loss: SpladeLoss with these parameters:

  {
      "loss": "SparseDistillKLDivMarginMSELoss",
      "lambda_corpus": 0.0005,
      "lambda_query": 0.0005
  }
  

Evaluation Dataset

Unnamed Dataset

• Size: 1,000 evaluation samples
• Columns: query, positive, negative, and label
• Approximate statistics based on the first 1000 samples:

  	query	positive	negative	label
  type	string	string	string	list
  details	min: 4 tokens mean: 9.12 tokens max: 37 tokens	min: 17 tokens mean: 78.91 tokens max: 239 tokens	min: 25 tokens mean: 81.25 tokens max: 239 tokens	size: 2 elements

• Samples:

  query	positive	negative	label
  how long to cook roast beef for	Roasting times for beef. Preheat your oven to 160°C (325°F) and use these cooking times to prepare a roast that's moist, tender and delicious. Your roast should be covered with foil for the first half of the roasting time to prevent drying the outer layer.3 to 5lb Joint 1½ to 2 hours.reheat your oven to 160°C (325°F) and use these cooking times to prepare a roast that's moist, tender and delicious. Your roast should be covered with foil for the first half of the roasting time to prevent drying the outer layer.	Estimating Cooking Time for Large Beef Roasts. If you roast at a steady 325F (160C), subtract 2 minutes or so per pound. If the roast is refrigerated just before going into the oven, add 2 or 3 minutes per pound. WARNING NOTES: Remember, the rib roast will continue to cook as it sets.	[6.501978874206543, 8.214995384216309]
  definition of fire inspection	Learn how to do a monthly fire extinguisher inspection in your workplace. Departments must assign an individual to inspect monthly the extinguishers in or adjacent to the department's facilities.1 Read Fire Extinguisher Types and Maintenance for more information.earn how to do a monthly fire extinguisher inspection in your workplace. Departments must assign an individual to inspect monthly the extinguishers in or adjacent to the department's facilities.	reconnaissance by fire-a method of reconnaissance in which fire is placed on a suspected enemy position in order to cause the enemy to disclose his presence by moving or returning fire. reconnaissance in force-an offensive operation designed to discover or test the enemy's strength (or to obtain other information). mission undertaken to obtain, by visual observation or other detection methods, information about the activities and resources of an enemy or potential enemy, or to secure data concerning the meteorological, hydrographic, or geographic characteristics of a particular area.	[-0.38299351930618286, -0.9372650384902954]
  how many stores does family dollar have	Property Spotlight: New Retail Center at Hamilton & Warner - Outlots Available!! Family Dollar is closing stores following a disappointing second quarter. Family Dollar Stores Inc. won’t just be cutting prices in an attempt to boost its business – it’ll be closing stores as well. The Matthews, N.C.-based discount retailer plans to shutter 370 under-performing shops, according to the Charlotte Business Journal.	Glassdoor has 1,976 Family Dollar Stores reviews submitted anonymously by Family Dollar Stores employees. Read employee reviews and ratings on Glassdoor to decide if Family Dollar Stores is right for you.	[4.726407527923584, 8.284608840942383]

• Loss: SpladeLoss with these parameters:

  {
      "loss": "SparseDistillKLDivMarginMSELoss",
      "lambda_corpus": 0.0005,
      "lambda_query": 0.0005
  }
  

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
-1	-1	-	-	0.6592	0.3737	0.6949	0.5759
0.0162	100	0.5304	-	-	-	-	-
0.0323	200	0.4704	-	-	-	-	-
0.0485	300	0.516	-	-	-	-	-
0.0646	400	0.462	-	-	-	-	-
0.0808	500	0.5311	0.5003	0.6586	0.3597	0.7063	0.5749
0.0970	600	0.5254	-	-	-	-	-
0.1131	700	0.5476	-	-	-	-	-
0.1293	800	0.5316	-	-	-	-	-
0.1454	900	0.5691	-	-	-	-	-
0.1616	1000	0.573	0.5246	0.6721	0.3488	0.6984	0.5731
0.1778	1100	0.5744	-	-	-	-	-
0.1939	1200	0.5241	-	-	-	-	-
0.2101	1300	0.5625	-	-	-	-	-
0.2262	1400	0.5586	-	-	-	-	-
0.2424	1500	0.548	0.5440	0.6709	0.3470	0.6945	0.5708
0.2586	1600	0.5765	-	-	-	-	-
0.2747	1700	0.5324	-	-	-	-	-
0.2909	1800	0.5751	-	-	-	-	-
0.3070	1900	0.5612	-	-	-	-	-
0.3232	2000	0.5056	0.5179	0.6785	0.3605	0.6882	0.5757
0.3394	2100	0.559	-	-	-	-	-
0.3555	2200	0.5247	-	-	-	-	-
0.3717	2300	0.5248	-	-	-	-	-
0.3878	2400	0.5025	-	-	-	-	-
0.4040	2500	0.5469	0.5251	0.6981	0.3575	0.7008	0.5855
0.4202	2600	0.5467	-	-	-	-	-
0.4363	2700	0.5039	-	-	-	-	-
0.4525	2800	0.5484	-	-	-	-	-
0.4686	2900	0.5004	-	-	-	-	-
0.4848	3000	0.4987	0.5011	0.6710	0.3578	0.6970	0.5753
0.5010	3100	0.4981	-	-	-	-	-
0.5171	3200	0.4859	-	-	-	-	-
0.5333	3300	0.496	-	-	-	-	-
0.5495	3400	0.4926	-	-	-	-	-
0.5656	3500	0.5231	0.4793	0.7063	0.3626	0.6875	0.5855
0.5818	3600	0.4681	-	-	-	-	-
0.5979	3700	0.4892	-	-	-	-	-
0.6141	3800	0.4765	-	-	-	-	-
0.6303	3900	0.4622	-	-	-	-	-
0.6464	4000	0.4893	0.4840	0.7026	0.3620	0.6879	0.5842
0.6626	4100	0.4902	-	-	-	-	-
0.6787	4200	0.4777	-	-	-	-	-
0.6949	4300	0.4819	-	-	-	-	-
0.7111	4400	0.4678	-	-	-	-	-
0.7272	4500	0.4751	0.4619	0.6915	0.3523	0.6919	0.5786
0.7434	4600	0.4796	-	-	-	-	-
0.7595	4700	0.4492	-	-	-	-	-
0.7757	4800	0.4547	-	-	-	-	-
0.7919	4900	0.4538	-	-	-	-	-
0.8080	5000	0.4418	0.4440	0.6851	0.3566	0.6955	0.5791
0.8242	5100	0.4634	-	-	-	-	-
0.8403	5200	0.445	-	-	-	-	-
0.8565	5300	0.4832	-	-	-	-	-
0.8727	5400	0.4617	-	-	-	-	-
0.8888	5500	0.4279	0.4578	0.6793	0.3579	0.6941	0.5771
0.9050	5600	0.4515	-	-	-	-	-
0.9211	5700	0.4579	-	-	-	-	-
0.9373	5800	0.4834	-	-	-	-	-
0.9535	5900	0.4676	-	-	-	-	-
0.9696	6000	0.4482	0.4371	0.6809	0.3525	0.6874	0.5736
0.9858	6100	0.4829	-	-	-	-	-
-1	-1	-	-	0.6789	0.3503	0.6881	0.5724

Environmental Impact

Carbon emissions were measured using CodeCarbon.

• Energy Consumed: 0.191 kWh
• Carbon Emitted: 0.074 kg of CO2
• Hours Used: 0.538 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},
}

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}
}