Sparse CSR model trained on Natural Questions

This is a CSR Sparse Encoder model finetuned from mixedbread-ai/mxbai-embed-large-v1 on the gooaq dataset using the sentence-transformers library. It maps sentences & paragraphs to a 4096-dimensional sparse vector space and can be used for semantic search and sparse retrieval.

Model Details

Model Description

• Model Type: CSR Sparse Encoder
• Base model: mixedbread-ai/mxbai-embed-large-v1
• Maximum Sequence Length: 512 tokens
• Output Dimensionality: 4096 dimensions
• Similarity Function: Dot Product
• Training Dataset:
  • gooaq
• 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): Transformer({'max_seq_length': 512, 'do_lower_case': False}) with Transformer model: BertModel 
  (1): Pooling({'word_embedding_dimension': 1024, 'pooling_mode_cls_token': True, 'pooling_mode_mean_tokens': False, 'pooling_mode_max_tokens': False, 'pooling_mode_mean_sqrt_len_tokens': False, 'pooling_mode_weightedmean_tokens': False, 'pooling_mode_lasttoken': False, 'include_prompt': True})
  (2): CSRSparsity({'input_dim': 1024, 'hidden_dim': 4096, 'k': 256, 'k_aux': 512, 'normalize': False, 'dead_threshold': 30})
)

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/csr-mxbai-embed-large-v1-gooaq-2e-4")
# Run inference
sentences = [
    'are you human korean novela?',
    "Are You Human? (Korean: 너도 인간이니; RR: Neodo Inganini; lit. Are You Human Too?) is a 2018 South Korean television series starring Seo Kang-jun and Gong Seung-yeon. It aired on KBS2's Mondays and Tuesdays at 22:00 (KST) time slot, from June 4 to August 7, 2018.",
    'A relative of European pear varieties like Bartlett and Anjou, the Asian pear is great used in recipes or simply eaten out of hand. It retains a crispness that works well in slaws and salads, and it holds its shape better than European pears when baked and cooked.',
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# (3, 4096)

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

Evaluation

Metrics

Sparse Information Retrieval

• Datasets: NanoMSMARCO_128, NanoNFCorpus_128 and NanoNQ_128
• Evaluated with SparseInformationRetrievalEvaluator with these parameters:

  {
      "max_active_dims": 128
  }
  

Sparse Nano BEIR

• Dataset: NanoBEIR_mean_128
• Evaluated with SparseNanoBEIREvaluator with these parameters:

  {
      "dataset_names": [
          "msmarco",
          "nfcorpus",
          "nq"
      ],
      "max_active_dims": 128
  }
  

Sparse Information Retrieval

• Datasets: NanoMSMARCO_256, NanoNFCorpus_256 and NanoNQ_256
• Evaluated with SparseInformationRetrievalEvaluator with these parameters:

  {
      "max_active_dims": 256
  }
  

Sparse Nano BEIR

• Dataset: NanoBEIR_mean_256
• Evaluated with SparseNanoBEIREvaluator with these parameters:

  {
      "dataset_names": [
          "msmarco",
          "nfcorpus",
          "nq"
      ],
      "max_active_dims": 256
  }
  

Training Details

Training Dataset

gooaq

• Dataset: gooaq at b089f72
• Size: 3,011,496 training samples
• Columns: question and answer
• Approximate statistics based on the first 1000 samples:

  	question	answer
  type	string	string
  details	min: 8 tokens mean: 11.87 tokens max: 23 tokens	min: 14 tokens mean: 60.09 tokens max: 201 tokens

• Samples:

  question	answer
  what is the difference between clay and mud mask?	The main difference between the two is that mud is a skin-healing agent, while clay is a cosmetic, drying agent. Clay masks are most useful for someone who has oily skin and is prone to breakouts of acne and blemishes.
  myki how much on card?	A full fare myki card costs $6 and a concession, seniors or child myki costs $3. For more information about how to use your myki, visit ptv.vic.gov.au or call 1800 800 007.
  how to find out if someone blocked your phone number on iphone?	If you get a notification like "Message Not Delivered" or you get no notification at all, that's a sign of a potential block. Next, you could try calling the person. If the call goes right to voicemail or rings once (or a half ring) then goes to voicemail, that's further evidence you may have been blocked.

• Loss: CSRLoss with these parameters:

  {
      "beta": 0.1,
      "gamma": 1.0,
      "loss": "SparseMultipleNegativesRankingLoss(scale=1.0, similarity_fct='dot_score')"
  }
  

Evaluation Dataset

gooaq

• Dataset: gooaq at b089f72
• Size: 1,000 evaluation samples
• Columns: question and answer
• Approximate statistics based on the first 1000 samples:

  	question	answer
  type	string	string
  details	min: 8 tokens mean: 11.88 tokens max: 22 tokens	min: 14 tokens mean: 61.03 tokens max: 127 tokens

• Samples:

  question	answer
  how do i program my directv remote with my tv?	['Press MENU on your remote.', 'Select Settings & Help > Settings > Remote Control > Program Remote.', 'Choose the device (TV, audio, DVD) you wish to program. ... ', 'Follow the on-screen prompts to complete programming.']
  are rodrigues fruit bats nocturnal?	Before its numbers were threatened by habitat destruction, storms, and hunting, some of those groups could number 500 or more members. Sunrise, sunset. Rodrigues fruit bats are most active at dawn, at dusk, and at night.
  why does your heart rate increase during exercise bbc bitesize?	During exercise there is an increase in physical activity and muscle cells respire more than they do when the body is at rest. The heart rate increases during exercise. The rate and depth of breathing increases - this makes sure that more oxygen is absorbed into the blood, and more carbon dioxide is removed from it.

• Loss: CSRLoss with these parameters:

  {
      "beta": 0.1,
      "gamma": 1.0,
      "loss": "SparseMultipleNegativesRankingLoss(scale=1.0, similarity_fct='dot_score')"
  }
  

Training Hyperparameters

Non-Default Hyperparameters

• eval_strategy: steps
• per_device_train_batch_size: 64
• per_device_eval_batch_size: 64
• learning_rate: 0.0002
• num_train_epochs: 1
• warmup_ratio: 0.1
• bf16: 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: 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: 0.0002
• 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: True
• fp16: False
• 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
• 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

Click to expand

• The bold row denotes the saved checkpoint.

Environmental Impact

Carbon emissions were measured using CodeCarbon.

• Energy Consumed: 1.202 kWh
• Carbon Emitted: 0.467 kg of CO2
• Hours Used: 3.125 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",
}

CSRLoss

@misc{wen2025matryoshkarevisitingsparsecoding,
      title={Beyond Matryoshka: Revisiting Sparse Coding for Adaptive Representation},
      author={Tiansheng Wen and Yifei Wang and Zequn Zeng and Zhong Peng and Yudi Su and Xinyang Liu and Bo Chen and Hongwei Liu and Stefanie Jegelka and Chenyu You},
      year={2025},
      eprint={2503.01776},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2503.01776},
}

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}
}