CoCondenser trained on MS MARCO

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-minilm-small")
# Run inference
queries = [
    "how much would dreamers cost the taxpayers",
]
documents = [
    'Plus, the CBO said the Dreamers would bring an additional 80,000 immigrants to the U.S., adding to the liability. In total, the immigrants and their families would cost taxpayers $26.8 billion, but only pay back $.9 billion in taxes, the CBO said. The analysis found that roughly 3.25 million undocumented immigrants are eligible for Dreamer status, while only 2 million would apply and only 1.6 million would be accepted over the next decade.',
    'Playing Chicken with the $18 Trillion U.S. Economy: The full cost of the last government shutdown two years ago was staggering â\x80\x93 it delivered a $24 billion blow to the U.S. economy and taxpayers. Now we may be about to repeat government shutdown history on Dec. 11.',
    'Sustain is defined as to support something or to endure a trial or hardship. 1  An example of sustain is for a foundation to support the house. 2  An example of sustain is to survive days without food or water.',
]
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([[27.1439, 12.7876,  0.6402]])

Evaluation

Metrics

Sparse Information Retrieval

Datasets: NanoMSMARCO, NanoNFCorpus and NanoNQ
Evaluated with SparseInformationRetrievalEvaluator

Metric	NanoMSMARCO	NanoNFCorpus	NanoNQ
dot_accuracy@1	0.48	0.4	0.5
dot_accuracy@3	0.62	0.58	0.74
dot_accuracy@5	0.76	0.64	0.76
dot_accuracy@10	0.9	0.68	0.82
dot_precision@1	0.48	0.4	0.5
dot_precision@3	0.2067	0.3733	0.2533
dot_precision@5	0.152	0.328	0.156
dot_precision@10	0.09	0.274	0.09
dot_recall@1	0.48	0.0418	0.46
dot_recall@3	0.62	0.0962	0.69
dot_recall@5	0.76	0.1174	0.7
dot_recall@10	0.9	0.1424	0.79
dot_ndcg@10	0.6685	0.3411	0.6444
dot_mrr@10	0.5974	0.5052	0.617
dot_map@100	0.6011	0.1532	0.5955
query_active_dims	57.08	50.46	54.1
query_sparsity_ratio	0.9981	0.9983	0.9982
corpus_active_dims	187.3189	331.6617	211.6348
corpus_sparsity_ratio	0.9939	0.9891	0.9931

Sparse Nano BEIR

Dataset: NanoBEIR_mean

Evaluated with SparseNanoBEIREvaluator with these parameters:

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

Metric	Value
dot_accuracy@1	0.46
dot_accuracy@3	0.6467
dot_accuracy@5	0.72
dot_accuracy@10	0.8
dot_precision@1	0.46
dot_precision@3	0.2778
dot_precision@5	0.212
dot_precision@10	0.1513
dot_recall@1	0.3273
dot_recall@3	0.4687
dot_recall@5	0.5258
dot_recall@10	0.6108
dot_ndcg@10	0.5513
dot_mrr@10	0.5732
dot_map@100	0.4499
query_active_dims	53.88
query_sparsity_ratio	0.9982
corpus_active_dims	229.4242
corpus_sparsity_ratio	0.9925

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: -1.28 mean: 13.47 max: 22.27	min: 4 tokens mean: 8.97 tokens max: 21 tokens	min: 19 tokens mean: 81.41 tokens max: 220 tokens	min: 17 tokens mean: 76.39 tokens max: 195 tokens

Samples:

score	query	positive	negative
`13.85562777519226`	`what is a reflective journal?`	`A reflective journal is a tool that students are encouraged to use to help them understand not just what they have learned while studying but also how they learned it by reflecting on the learning experience itself.`	`The point is that this approach believes that literature can be used to illuminate some truth about something which is not literature. The difference between this approach and the didactic approach is that didactic approach considers author as a teacher, while the reflective approach considers him or her an observer.`
`12.178914229075115`	`original footloose release`	`Footloose (2011 film) Footloose is a 2011 American musical dance film directed by Craig Brewer. It is a remake of the 1984 film of the same name and stars Kenny Wormald, Julianne Hough, Andie MacDowell, and Dennis Quaid. The film follows a young man who moves from Boston to a small southern town and protests the town's ban against dancing.`	`In the 2001 re-release of Thriller they added the second verse of the rap which was recorded but not included on the original here is the second verse by Vincent Price (I heard a 3rd was written but never recorded) The demons squeal in sheer delight. It's you they spy, so plump, so right.`
`19.897210280100506`	`time of day blood pressure`	`Day Time Blood Pressure. For most people, your body's blood pressure rises during the morning hours and reaches its highest point around midday. This is because your body is preset to increase its functions for anticipated daily activity. Your body reaches its lowest blood pressure at bedtime, between 8 p.m. and 2 a.m.`	`labetalol is used alone or together with other medicines to treat high blood pressure hypertension high blood pressure adds to the workload of the heart and arteriesif it continues for a long time the heart and arteries may not function properlyabetalol is used alone or together with other medicines to treat high blood pressure hypertension high blood pressure adds to the workload of the heart and arteries`

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: -2.25 mean: 13.3 max: 22.52	min: 4 tokens mean: 9.31 tokens max: 40 tokens	min: 17 tokens mean: 79.88 tokens max: 227 tokens	min: 18 tokens mean: 77.64 tokens max: 250 tokens

Samples:

score	query	positive	negative
`11.338554302851358`	`victor cruz dance salsa in the super bowl`	`The popular former Giant â who helped lead the team to a Super Bowl title in the 2011 season â ... Victor Cruz performed his first salsa dance in Chicago. The popular former Giant â who helped lead the team to a Super Bowl title in the 2011 season â caught a 2-yard touchdown pass from Mitch Trubisky in the Bearsâ 24-17 preseason loss to the Broncos on Thursday night.`	`Victor Cruz, Giants hammer out deal. Receiver Victor Cruz on Monday signed a six-year contract through the 2018 season with the New York Giants. The contract is worth $46 million and pays him $15.625 million fully guaranteed the first two seasons, a source said.`
`18.167373975118`	`what is the phone number for roblox`	`im calling roblox hq and the roblox number is 888 858 2569 or if u live in canada its 1888 858 2569 subscibe to us (waffleman514 and twitterelgo) and join our youtube group on our profile Category`	`[edit] Create A New Place. This is where you define where your game will be published. 1 Go to Roblox.com and login. 2 Click My ROBLOX and then click Places. 3 Click Create Game Place. 4 Fill out the form. 5 Name is the name of the game.`
`17.668365399042766`	`can you freeze cream soup`	`With a modest investment in time and effort, you can make your own cream of mushroom soup and freeze it for later use. This leaves you firmly in control of the soup's ingredients and enables you to portion the soup in quantities that make sense for you.`	`I purchased 1.5 lbs of 3 large boneless, skinless chicken breasts. I am cooking them in a crockpot with 1 can of cream of mushroom soup, 1 can cream of chicken soup, 1 can of water and some canned mushrooms...the chicken is all at the bottom. About how long will it take the chicken to cook completely if I set my... show more I purchased 1.5 lbs of 3 large boneless, skinless chicken breasts.`

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
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: 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: 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
0.0178	100	795934.08	-	-	-	-	-
0.0356	200	13561.4538	-	-	-	-	-
0.0533	300	118.2925	-	-	-	-	-
0.0711	400	61.485	-	-	-	-	-
0.0889	500	44.6503	38.6276	0.5126	0.2701	0.5829	0.4552
0.1067	600	38.3666	-	-	-	-	-
0.1244	700	35.2046	-	-	-	-	-
0.1422	800	33.2246	-	-	-	-	-
0.16	900	31.5866	-	-	-	-	-
0.1778	1000	29.3914	38.9004	0.5849	0.3140	0.6009	0.4999
0.1956	1100	28.9009	-	-	-	-	-
0.2133	1200	29.5258	-	-	-	-	-
0.2311	1300	27.7958	-	-	-	-	-
0.2489	1400	27.0228	-	-	-	-	-
0.2667	1500	25.0953	22.5132	0.6090	0.3377	0.6166	0.5211
0.2844	1600	25.4396	-	-	-	-	-
0.3022	1700	22.53	-	-	-	-	-
0.32	1800	24.0084	-	-	-	-	-
0.3378	1900	23.5741	-	-	-	-	-
0.3556	2000	23.141	22.6775	0.6408	0.3560	0.5984	0.5317
0.3733	2100	22.0953	-	-	-	-	-
0.3911	2200	22.2789	-	-	-	-	-
0.4089	2300	20.9582	-	-	-	-	-
0.4267	2400	19.1969	-	-	-	-	-
0.4444	2500	21.047	28.3245	0.6209	0.3487	0.6260	0.5319
0.4622	2600	20.7531	-	-	-	-	-
0.48	2700	19.8115	-	-	-	-	-
0.4978	2800	18.6278	-	-	-	-	-
0.5156	2900	19.3731	-	-	-	-	-
0.5333	3000	18.4502	20.3191	0.6390	0.3506	0.6087	0.5328
0.5511	3100	18.4525	-	-	-	-	-
0.5689	3200	17.0456	-	-	-	-	-
0.5867	3300	17.256	-	-	-	-	-
0.6044	3400	17.6203	-	-	-	-	-
0.6222	3500	18.7721	17.7983	0.6685	0.3411	0.6444	0.5513
0.64	3600	16.7819	-	-	-	-	-
0.6578	3700	18.6132	-	-	-	-	-
0.6756	3800	15.5466	-	-	-	-	-
0.6933	3900	17.7706	-	-	-	-	-
0.7111	4000	16.6612	15.7565	0.6727	0.3519	0.6159	0.5468
0.7289	4100	16.4755	-	-	-	-	-
0.7467	4200	16.9832	-	-	-	-	-
0.7644	4300	14.9855	-	-	-	-	-
0.7822	4400	14.6835	-	-	-	-	-
0.8	4500	17.0725	18.0495	0.6652	0.3430	0.6423	0.5502
0.8178	4600	15.8136	-	-	-	-	-
0.8356	4700	15.6528	-	-	-	-	-
0.8533	4800	15.5791	-	-	-	-	-
0.8711	4900	15.1496	-	-	-	-	-
0.8889	5000	14.7461	16.4918	0.6373	0.3353	0.6403	0.5376
0.9067	5100	16.3102	-	-	-	-	-
0.9244	5200	14.5521	-	-	-	-	-
0.9422	5300	14.4375	-	-	-	-	-
0.96	5400	15.2282	-	-	-	-	-
0.9778	5500	14.4738	15.4439	0.6426	0.3385	0.6334	0.5382
0.9956	5600	14.6468	-	-	-	-	-
-1	-1	-	-	0.6685	0.3411	0.6444	0.5513

The bold row denotes the saved checkpoint.

Environmental Impact

Carbon emissions were measured using CodeCarbon.

Energy Consumed: 0.216 kWh
Carbon Emitted: 0.084 kg of CO2
Hours Used: 0.609 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}
    }