tomaarsen/splade-cocondenser-msmarco-margin-mse-small

CoCondenser trained on Natural-Questions tuples

This is a SPLADE Sparse Encoder model finetuned from Luyu/co-condenser-marco on the msmarco 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: Luyu/co-condenser-marco
• Maximum Sequence Length: 512 tokens
• Output Dimensionality: 30522 dimensions
• Similarity Function: Dot Product
• Training Dataset:
  • msmarco
• 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}) with MLMTransformer model: 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-msmarco-margin-mse")
# Run inference
queries = [
    "when did shanghai disneyland open",
]
documents = [
    "Shanghai Disney officially opens: A peek inside. June 17, 2016, 6 p.m. After five years of construction, $5.5 billion in spending and a month of testing to work out the kinks, Shanghai Disney Resort opened to the public just before noon, Shanghai time, on Thursday, June 16 (which was 9 p.m. Wednesday in Anaheim, home of the original Disney park). Shanghai Disneyland features six themed areas, and the resort contains two hotels, a shopping district and 99 acres of gardens, lakes and parkland. We'll keep you updated throughout the week with new details and peeks inside the resort.",
    'Map of the Old City of Shanghai. By the early 1400s, Shanghai had become important enough for Ming dynasty engineers to begin dredging the Huangpu River (also known as Shen). In 1553, a city wall was built around the Old Town (Nanshi) as a defense against the depredations of the Wokou (Japanese pirates).',
    'The conflict is referred to in China as the War of Resistance against Japanese Aggression (1937-45) and the Anti-Fascist War. Japanâ\x80\x99s expansionist policy of the 1930s, driven by the military, was to set up what it called the Greater East Asia Co-Prosperity Sphere. Marco Polo Bridge, Beijing.A sphere.e are marking the anniversary of Germany and Japanâ\x80\x99s surrender in 1945, but it is legitimate to suggest that the incident that sparked the conflict that became WWII occurred not in Poland in 1939 but in China, near this eleven-arched bridge on the outskirts of Beijing, in July 1937. Letâ\x80\x99s look at the undisputed facts.',
]
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([[31.8057, 19.5344, 12.4372]])

Evaluation

Metrics

Sparse Information Retrieval

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

Metric	NanoMSMARCO	NanoNFCorpus	NanoNQ
dot_accuracy@1	0.46	0.38	0.5
dot_accuracy@3	0.64	0.58	0.76
dot_accuracy@5	0.72	0.62	0.8
dot_accuracy@10	0.82	0.74	0.88
dot_precision@1	0.46	0.38	0.5
dot_precision@3	0.2133	0.36	0.26
dot_precision@5	0.144	0.316	0.168
dot_precision@10	0.082	0.27	0.096
dot_recall@1	0.46	0.0397	0.48
dot_recall@3	0.64	0.0752	0.71
dot_recall@5	0.72	0.0936	0.75
dot_recall@10	0.82	0.1467	0.85
dot_ndcg@10	0.6289	0.3304	0.6772
dot_mrr@10	0.5689	0.4958	0.6329
dot_map@100	0.5779	0.1478	0.6167
query_active_dims	56.1	53.68	55.94
query_sparsity_ratio	0.9982	0.9982	0.9982
corpus_active_dims	192.4087	367.5432	228.8362
corpus_sparsity_ratio	0.9937	0.988	0.9925

Sparse Nano BEIR

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

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

Metric	Value
dot_accuracy@1	0.4467
dot_accuracy@3	0.66
dot_accuracy@5	0.7133
dot_accuracy@10	0.8133
dot_precision@1	0.4467
dot_precision@3	0.2778
dot_precision@5	0.2093
dot_precision@10	0.1493
dot_recall@1	0.3266
dot_recall@3	0.4751
dot_recall@5	0.5212
dot_recall@10	0.6056
dot_ndcg@10	0.5455
dot_mrr@10	0.5658
dot_map@100	0.4475
query_active_dims	55.24
query_sparsity_ratio	0.9982
corpus_active_dims	246.1716
corpus_sparsity_ratio	0.9919

Training Details

Training Dataset

msmarco

• Dataset: msmarco at 9e329ed
• Size: 90,000 training samples
• Columns: score, query, positive, and negative
• Approximate statistics based on the first 1000 samples:

  	score	query	positive	negative
  type	float	string	string	string
  details	min: -2.22 mean: 13.59 max: 22.53	min: 4 tokens mean: 9.05 tokens max: 40 tokens	min: 19 tokens mean: 81.18 tokens max: 203 tokens	min: 15 tokens mean: 77.08 tokens max: 249 tokens

• Samples:

  score	query	positive	negative
  4.470368494590124	where does the bile duct carry its secretions	The function of the common bile duct is to carry bile from the liver and the gallbladder into the duodenum, the top of the small intestine directly after the stomach. The bile it carries interacts with ingested fats and fat-soluble vitamins to enable them to be absorbed by the intestine.	The gall bladder is a pouch-shaped organ that stores the bile produced by the liver. The gall bladder shares a vessel, called the common bile duct, with the liver. When bile is needed, it moves through the common bile duct into the first part of the small intestine, the duodenum. It is here that the bile breaks down fat.
  9.550037781397503	definition of reverse auction	Reverse auction. A reverse auction is a type of auction in which the roles of buyer and seller are reversed. In an ordinary auction (also known as a 'forward auction'), buyers compete to obtain goods or services by offering increasingly higher prices. In a reverse auction, the sellers compete to obtain business from the buyer and prices will typically decrease as the sellers underbid each other.	No-reserve auction. A No-reserve auction (NR), also known as an absolute auction, is an auction in which the item for sale will be sold regardless of price. From the seller's perspective, advertising an auction as having no reserve price can be desirable because it potentially attracts a greater number of bidders due to the possibility of a bargain.
  19.58259622255961	how do i prevent diverticulitis	Follow Following Unfollow Pending Disabled. A , Gastroenterology, answered. The suggestion to prevent diverticulitis is to eat a diet high in fiber, and that includes high-fiber whole grains, fruits, vegetables, nuts, and seeds. I’m aware that some gastroenterologists say to avoid all seeds and nuts, so some of you are nuts enough to wash tomato seeds from slices and pick free poppy seeds from buns.	The test is fast and easy especially with the newer CT scanners. But does it provide the information needed? CT KUBs are used to screen for a variety of intra-abdominal conditions, including appendicitis, kidney stones, diverticulitis, and others.

• Loss: SpladeLoss with these parameters:

  {
      "loss": "SparseMarginMSELoss",
      "lambda_corpus": 0.08,
      "lambda_query": 0.1
  }
  

Evaluation Dataset

msmarco

• Dataset: msmarco at 9e329ed
• Size: 10,000 evaluation samples
• Columns: score, query, positive, and negative
• Approximate statistics based on the first 1000 samples:

  	score	query	positive	negative
  type	float	string	string	string
  details	min: -1.34 mean: 13.49 max: 22.2	min: 4 tokens mean: 8.85 tokens max: 27 tokens	min: 14 tokens mean: 80.48 tokens max: 211 tokens	min: 20 tokens mean: 77.44 tokens max: 209 tokens

• Samples:

  score	query	positive	negative
  15.64028427998225	what is a protected seedbed	A seedbed is a plot of garden set aside to grow vegetables seeds, which can later be transplanted. seedbed is a plot of garden set aside to grow vegetables seeds, which can later be transplanted.	Several articles within the Confederate States’ Constitution specifically protected slavery within the Confederacy, but some articles of the U.S. Constitution also protected slavery—the Emancipation Proclamation drew a clearer distinction between the two.
  6.375148057937622	who founded ecuador	The first Spanish settlement in Ecuador was established in 1534 at Quito on the site of an important Incan town of the same name. Another settlement was established four years later near the river Guayas in Guayaquil.	Zuleta is a colonial working farm of 4,000 acres (2,000 hectares) that belongs to the family of Mr. Galo Plaza lasso, a former president of Ecuador, for more than 100 years. It was chosen as one of the world’s “Top Ten Finds” by Outside magazine and named as one of the best Ecuador Hotel by National Geographic Traveler.
  8.436618288358051	what is aol problem	AOL problems. Lots of people are reporting ongoing (RTR:GE) messages from AOL today. This indicates the AOL mail servers are having problems and can’t accept mail. This has nothing to do with spam, filtering or malicious email. This is simply their servers aren’t functioning as well as they should be and so AOL can’t accept all the mail thrown at them. These types of blocks resolve themselves. Update Feb 8, 2016: AOL users are having problems logging in.	Executive Director. I have read these complaints of poor service and agree 110%. I'm a college professor and give extra credit to all AOL users and over the 100% highest grade. I thought I phoned AOL and get some chap in India who is a proven scam man and I'm the poor American SOB who gets whacked.

• Loss: SpladeLoss with these parameters:

  {
      "loss": "SparseMarginMSELoss",
      "lambda_corpus": 0.08,
      "lambda_query": 0.1
  }
  

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
0.0178	100	805201.68	-	-	-	-	-
0.0356	200	11999.3975	-	-	-	-	-
0.0533	300	124.0031	-	-	-	-	-
0.0711	400	62.6813	-	-	-	-	-
0.0889	500	46.0329	49.7658	0.4890	0.2543	0.5131	0.4188
0.1067	600	41.2877	-	-	-	-	-
0.1244	700	35.3636	-	-	-	-	-
0.1422	800	33.3727	-	-	-	-	-
0.16	900	29.389	-	-	-	-	-
0.1778	1000	31.2482	28.1527	0.5652	0.2875	0.5423	0.4650
0.1956	1100	31.43	-	-	-	-	-
0.2133	1200	27.9919	-	-	-	-	-
0.2311	1300	26.9214	-	-	-	-	-
0.2489	1400	27.5533	-	-	-	-	-
0.2667	1500	25.7473	26.8466	0.5837	0.3265	0.6268	0.5123
0.2844	1600	26.7899	-	-	-	-	-
0.3022	1700	24.0652	-	-	-	-	-
0.32	1800	23.5837	-	-	-	-	-
0.3378	1900	24.1051	-	-	-	-	-
0.3556	2000	24.6901	22.0851	0.6018	0.3325	0.6359	0.5234
0.3733	2100	21.5136	-	-	-	-	-
0.3911	2200	22.066	-	-	-	-	-
0.4089	2300	20.8234	-	-	-	-	-
0.4267	2400	20.1988	-	-	-	-	-
0.4444	2500	20.0342	20.3437	0.5901	0.3222	0.6010	0.5044
0.4622	2600	18.8835	-	-	-	-	-
0.48	2700	19.4797	-	-	-	-	-
0.4978	2800	19.6199	-	-	-	-	-
0.5156	2900	16.6963	-	-	-	-	-
0.5333	3000	19.9204	18.0851	0.5915	0.3111	0.6323	0.5116
0.5511	3100	18.7849	-	-	-	-	-
0.5689	3200	18.3169	-	-	-	-	-
0.5867	3300	17.1938	-	-	-	-	-
0.6044	3400	18.0807	-	-	-	-	-
0.6222	3500	16.7721	20.1195	0.6012	0.3119	0.6337	0.5156
0.64	3600	16.7909	-	-	-	-	-
0.6578	3700	16.4954	-	-	-	-	-
0.6756	3800	16.3734	-	-	-	-	-
0.6933	3900	17.2231	-	-	-	-	-
0.7111	4000	16.8486	17.5785	0.6228	0.3423	0.6553	0.5401
0.7289	4100	18.2939	-	-	-	-	-
0.7467	4200	16.1108	-	-	-	-	-
0.7644	4300	16.878	-	-	-	-	-
0.7822	4400	15.6163	-	-	-	-	-
0.8	4500	15.8337	16.1847	0.6286	0.3376	0.6639	0.5434
0.8178	4600	15.5014	-	-	-	-	-
0.8356	4700	15.7579	-	-	-	-	-
0.8533	4800	15.9361	-	-	-	-	-
0.8711	4900	16.3308	-	-	-	-	-
0.8889	5000	14.8395	17.4054	0.6221	0.3280	0.6853	0.5451
0.9067	5100	14.8655	-	-	-	-	-
0.9244	5200	14.6498	-	-	-	-	-
0.9422	5300	15.5189	-	-	-	-	-
0.96	5400	14.608	-	-	-	-	-
0.9778	5500	15.6019	16.4883	0.6298	0.3317	0.6831	0.5482
0.9956	5600	14.6263	-	-	-	-	-
-1	-1	-	-	0.6289	0.3304	0.6772	0.5455

Environmental Impact

Carbon emissions were measured using CodeCarbon.

• Energy Consumed: 0.218 kWh
• Carbon Emitted: 0.085 kg of CO2
• Hours Used: 0.618 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},
}

SparseMarginMSELoss

@misc{hofstätter2021improving,
    title={Improving Efficient Neural Ranking Models with Cross-Architecture Knowledge Distillation},
    author={Sebastian Hofstätter and Sophia Althammer and Michael Schröder and Mete Sertkan and Allan Hanbury},
    year={2021},
    eprint={2010.02666},
    archivePrefix={arXiv},
    primaryClass={cs.IR}
}

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}
    }