tomaarsen/modernbert-large-multivector-msmarco32k-3e-4-20261014

ModernBERT-large multivector MS MARCO pilot

This is a Multi-Vector Encoder model finetuned from answerdotai/ModernBERT-large using the sentence-transformers library. It maps inputs to sequences of 128-dimensional token-level vectors and scores them with late interaction (MaxSim), useful for semantic search with late interaction.

Model Details

Model Description

• Model Type: Multi-Vector Encoder
• Base model: answerdotai/ModernBERT-large
• Maximum Sequence Length: 256 tokens
  • Maximum Query Length: 32 tokens
  • Maximum Document Length: 256 tokens
• Output Dimensionality: 128 dimensions
• Similarity Function: MaxSim
• Supported Modality: Text
• Language: en
• License: apache-2.0

Model Sources

• Documentation: Sentence Transformers Documentation
• Documentation: Multi-Vector Encoder Documentation
• Repository: Sentence Transformers on GitHub
• Hugging Face: Multi-Vector Encoders on Hugging Face

Full Model Architecture

MultiVectorEncoder(
  (0): Transformer({'transformer_task': 'feature-extraction', 'modality_config': {'text': {'method': 'forward', 'method_output_name': 'last_hidden_state'}}, 'module_output_name': 'token_embeddings', 'query_length': 32, 'document_length': 256, 'architecture': 'ModernBertModel'})
  (1): Dense({'in_features': 1024, 'out_features': 128, 'bias': False, 'activation_function': 'torch.nn.modules.linear.Identity', 'module_input_name': 'token_embeddings', 'module_output_name': 'token_embeddings'})
  (2): MultiVectorMask({'skiplist_words': [], 'skiplist_tasks': ['document'], 'keep_only_token_ids': None})
  (3): Normalize({'module_input_name': 'token_embeddings', 'module_output_name': 'token_embeddings'})
)

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 MultiVectorEncoder

# Download from the 🤗 Hub
model = MultiVectorEncoder("tomaarsen/modernbert-large-multivector-msmarco32k-3e-4-20261014")
# Run inference: each input becomes a sequence of per-token vectors (variable length).
queries = [
    'goldman sachs contact number',
]
documents = [
    'The customer support phone number of Goldman Sachs Contact is 1-212-902-0300 (Click phone number to call).',
    'Goldman Sachs Bank USA is an FDIC insured institution located in New York, NY. It was founded in 1990 and has approximately $122.68 billion in assets. Customers can open an account via one of its 5 branches.',
    'The Director of Supply Chain will be responsible for ensuring supply and availability of finished goods and services to optimize customer services at the lowest... 5 days ago - save job - more... -',
]
query_embeddings = model.encode_query(queries)
document_embeddings = model.encode_document(documents)
print(query_embeddings[0].shape, document_embeddings[0].shape)
# (9, 128) (27, 128)

# Get the MaxSim similarity scores
similarities = model.similarity(query_embeddings, document_embeddings)
print(similarities)
# tensor([[8.1271, 5.9187, 1.1399]])

Training Details

Training Dataset

Unnamed Dataset

• Size: 32,768 training samples
• Columns: query, positive, and negative
• Approximate statistics based on the first 100 samples:

  	query	positive	negative
  type	string	string	string
  modality	text	text	text
  details	min: 5 tokens mean: 9.82 tokens max: 23 tokens	min: 29 tokens mean: 82.93 tokens max: 189 tokens	min: 35 tokens mean: 80.9 tokens max: 171 tokens

• Samples:

  query	positive	negative
  what clothing do people wear in denmark guys	While suits and other forms of Western fashion are ubiquitous in Japan, most people also wear some of the more traditional styles of Japanese clothing. For example, you often … will see people wearing traditional Japanese clothing at weddings or funerals, graduations and even for the summer fireworks festivals--hanabi. You can learn about some of the different styles of Japanese clothing and--if you dare--try some of them yourself.	Yet even the Democratic presidential candidates, who say they love or want to learn from those countries, don’t seem know how they actually work. Why We’re Not Denmark Proof that they do work is delivered every year in data-rich evaluations by the United Nations and other international bodies.
  average salary healthcare administration	Healthcare Administrator Salary. (United States). A Healthcare Administrator earns an average salary of $64,821 per year. The highest paying skills associated with this job are Program Management, Budget Management, and Operations Management. Experience has a moderate effect on income for this job.ealthcare Administrator Salary. (United States). A Healthcare Administrator earns an average salary of $64,821 per year. The highest paying skills associated with this job are Program Management, Budget Management, and Operations Management. Experience has a moderate effect on income for this job.	The five states with the highest average annual salaries for health care administrators, according to May 2011 data from the BLS, are: Massachusetts, $112,690; Rhode Island, $112,260; New York, $111,270; New Jersey, $110,100; and Washington, $108,620.
  what unique characteristics does bacillus megaterium have	Rare to none: Bacillus megaterium is a rod-shaped, gram-positive, endospore forming, aerotolerant species of bacteria used as a soil inoculant in agriculture and horticulture.Bacterium is arranged into the streptobacillus form.. Get help from a doctor now ›. Rare to none: Bacillus megaterium is a rod-shaped, gram-positive, endospore forming, aerotolerant species of bacteria used as a soil inoculant in agriculture and horticulture. Bacterium is arranged into the streptobacillus form.	Bacillus. Bacillus is a specific genus of rod-shaped bacteria that are capable of forming spores. They are sporulating, aerobic and ubiquitous in nature. Bacillus is a fairly large group with many members, including Bacillus cereus, Bacillus clausii and Bacillus halodenitrificans.Bacillus spores, also called endospores, are resistant to harsh chemical and physical conditions. This makes the bacteria able to withstand disinfectants, radiation, desiccation and heat.acillus is a fairly large group with many members, including Bacillus cereus, Bacillus clausii and Bacillus halodenitrificans. Bacillus spores, also called endospores, are resistant to harsh chemical and physical conditions. This makes the bacteria able to withstand disinfectants, radiation, desiccation and heat.

• Loss: CachedMultiVectorMultipleNegativesRankingLoss with these parameters:

  {
      "scale": 1.0,
      "similarity_fct": "colbert_scores",
      "mini_batch_size": 8,
      "mini_batch_num_tokens": null,
      "score_mini_batch_size": 4,
      "gather_across_devices": false
  }
  

Training Hyperparameters

Non-Default Hyperparameters

• per_device_train_batch_size: 64
• num_train_epochs: 1
• learning_rate: 0.0003
• warmup_steps: 0.05
• optim: adamw_torch
• weight_decay: 0.01
• bf16: True
• tf32: True
• logging_nan_inf_filter: False
• disable_tqdm: True
• per_device_eval_batch_size: 32
• push_to_hub: True
• hub_private_repo: False
• hub_model_id: tomaarsen/modernbert-large-multivector-msmarco32k-3e-4-20261014
• load_best_model_at_end: True
• seed: 12
• data_seed: 12
• dataloader_num_workers: 2
• batch_sampler: no_duplicates
• router_mapping: {'query': 'query', 'positive': 'document', 'negative': 'document'}

All Hyperparameters

Click to expand

• per_device_train_batch_size: 64
• num_train_epochs: 1
• max_steps: -1
• learning_rate: 0.0003
• lr_scheduler_type: linear
• lr_scheduler_kwargs: None
• warmup_steps: 0.05
• optim: adamw_torch
• optim_args: None
• weight_decay: 0.01
• adam_beta1: 0.9
• adam_beta2: 0.999
• adam_epsilon: 1e-08
• optim_target_modules: None
• gradient_accumulation_steps: 1
• average_tokens_across_devices: True
• max_grad_norm: 1.0
• label_smoothing_factor: 0.0
• bf16: True
• fp16: False
• bf16_full_eval: False
• fp16_full_eval: False
• tf32: True
• gradient_checkpointing: False
• gradient_checkpointing_kwargs: None
• torch_compile: False
• torch_compile_backend: None
• torch_compile_mode: None
• use_liger_kernel: False
• liger_kernel_config: None
• use_cache: False
• neftune_noise_alpha: None
• torch_empty_cache_steps: None
• auto_find_batch_size: False
• log_on_each_node: True
• logging_nan_inf_filter: False
• include_num_input_tokens_seen: no
• log_level: passive
• log_level_replica: warning
• disable_tqdm: True
• project: huggingface
• trackio_space_id: None
• trackio_bucket_id: None
• trackio_static_space_id: None
• per_device_eval_batch_size: 32
• prediction_loss_only: True
• eval_on_start: False
• eval_do_concat_batches: True
• eval_use_gather_object: False
• eval_accumulation_steps: None
• include_for_metrics: []
• batch_eval_metrics: False
• save_only_model: False
• save_on_each_node: False
• enable_jit_checkpoint: False
• push_to_hub: True
• hub_private_repo: False
• hub_model_id: tomaarsen/modernbert-large-multivector-msmarco32k-3e-4-20261014
• hub_strategy: every_save
• hub_always_push: False
• hub_revision: None
• load_best_model_at_end: True
• ignore_data_skip: False
• restore_callback_states_from_checkpoint: False
• full_determinism: False
• seed: 12
• data_seed: 12
• use_cpu: False
• accelerator_config: {'split_batches': False, 'dispatch_batches': None, 'even_batches': True, 'use_seedable_sampler': True, 'non_blocking': False, 'gradient_accumulation_kwargs': None}
• parallelism_config: None
• dataloader_drop_last: False
• dataloader_num_workers: 2
• dataloader_pin_memory: True
• dataloader_persistent_workers: False
• dataloader_prefetch_factor: None
• dataloader_multiprocessing_context: None
• dataloader_in_order: True
• remove_unused_columns: True
• label_names: None
• train_sampling_strategy: random
• length_column_name: length
• ddp_find_unused_parameters: None
• ddp_bucket_cap_mb: None
• ddp_broadcast_buffers: False
• ddp_static_graph: None
• ddp_backend: None
• ddp_timeout: 1800
• fsdp: None
• fsdp_config: None
• deepspeed: None
• debug: []
• skip_memory_metrics: True
• do_predict: False
• resume_from_checkpoint: None
• local_rank: -1
• prompts: None
• batch_sampler: no_duplicates
• multi_dataset_batch_sampler: proportional
• router_mapping: {'query': 'query', 'positive': 'document', 'negative': 'document'}
• learning_rate_mapping: {}
• warmup_ratio: None
• max_length: None

Training Logs

Epoch	Step	Training Loss
0.0020	1	4.3280
0.0195	10	3.4287
0.0391	20	1.8313
0.0586	30	1.4054
0.0781	40	1.1158
0.0977	50	1.1778
0.1172	60	1.1056
0.1367	70	1.0889
0.1562	80	1.1276
0.1758	90	1.0489
0.1953	100	1.0218
0.2148	110	1.0234
0.2344	120	1.0468
0.25	128	-
0.2539	130	1.0334
0.2734	140	1.0052
0.2930	150	0.9992
0.3125	160	0.9927
0.3320	170	0.9481
0.3516	180	0.9927
0.3711	190	0.8905
0.3906	200	0.9143
0.4102	210	0.9070
0.4297	220	0.8512
0.4492	230	0.7533
0.4688	240	0.7184
0.4883	250	0.7532
0.5	256	-
0.5078	260	0.7305
0.5273	270	0.6928
0.5469	280	0.7295
0.5664	290	0.6759
0.5859	300	0.7005
0.6055	310	0.7324
0.625	320	0.6644
0.6445	330	0.7214
0.6641	340	0.6924
0.6836	350	0.6311
0.7031	360	0.6570
0.7227	370	0.6532
0.7422	380	0.6485
0.75	384	-
0.7617	390	0.6592
0.7812	400	0.6294
0.8008	410	0.6055
0.8203	420	0.6307
0.8398	430	0.6027
0.8594	440	0.6095
0.8789	450	0.5935
0.8984	460	0.5755
0.9180	470	0.6394
0.9375	480	0.6207
0.9570	490	0.5791
0.9766	500	0.5866
0.9961	510	0.5986
1.0	512	-

Training Time

• Training: 18.2 minutes
• Evaluation: 17.8 seconds
• Total: 18.5 minutes

Framework Versions

• Python: 3.12.12
• Sentence Transformers: 6.2.0.dev0
• Transformers: 5.16.1
• PyTorch: 2.10.0+cu128
• Accelerate: 1.15.0
• Datasets: 5.1.0
• Tokenizers: 0.23.3

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",
}

CachedMultiVectorMultipleNegativesRankingLoss

@misc{gao2021scaling,
    title={Scaling Deep Contrastive Learning Batch Size under Memory Limited Setup},
    author={Luyu Gao and Yunyi Zhang and Jiawei Han and Jamie Callan},
    year={2021},
    eprint={2101.06983},
    archivePrefix={arXiv},
    primaryClass={cs.LG}
}

Controlled backbone comparison

Part of the encoder comparison. One epoch maximum on 32,768 MS MARCO triplets, seed 12. The best validation checkpoint was recovered from the repository history after a publication error. No training was repeated. See experiment.json for its original revision and validation score. Learning-rate selection uses validation only. Test data does not select checkpoints.

Dense and multi-vector use query length 32 and passage length 256 including special tokens. Rerankers use 32 query content tokens, 256 passage content tokens and a combined cap of 291. See the comparison scripts for matching input clipping. These are small-data research models.