aaa961/modernbert-embed-base-legal-matryoshka-2_2026-03-06

ModernBERT Embed base Legal Matryoshka

This is a sentence-transformers model finetuned from nomic-ai/modernbert-embed-base on the json dataset. It maps sentences & paragraphs to a 768-dimensional dense vector space and can be used for semantic textual similarity, semantic search, paraphrase mining, text classification, clustering, and more.

Model Details

Model Description

• —Model Type: Sentence Transformer
• —Base model: nomic-ai/modernbert-embed-base <!-- at revision d556a88e332558790b210f7bdbe87da2fa94a8d8 -->
• —Maximum Sequence Length: 8192 tokens
• —Output Dimensionality: 768 dimensions
• —Similarity Function: Cosine Similarity
• —Training Dataset:
• —json
• —Language: en
• —License: apache-2.0

Model Sources

• —Documentation: Sentence Transformers Documentation
• —Repository: Sentence Transformers on GitHub
• —Hugging Face: Sentence Transformers on Hugging Face

Full Model Architecture

SentenceTransformer(
  (0): Transformer({'max_seq_length': 8192, 'do_lower_case': False, 'architecture': 'ModernBertModel'})
  (1): Pooling({'word_embedding_dimension': 768, 'pooling_mode_cls_token': False, 'pooling_mode_mean_tokens': True, '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): Normalize()
)

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 SentenceTransformer

# Download from the 🤗 Hub
model = SentenceTransformer("aaa961/modernbert-embed-base-legal-matryoshka-2_2026-03-06")
# Run inference
sentences = [
    'withholding of personnel data and the security of foreign intelligence activities or the protection \nof intelligence sources and methods” in order to withhold information under Exemption 3, see \nBaker v. CIA, 580 F.2d 664, 669 (D.C. Cir. 1978), the two provisions—§ 403-1(i) and § 403g—\nare often invoked together.  In the instant case, however, the CIA solely invokes § 403g as a',
    'What court case is referenced in relation to Exemption 3?',
    'What is the subject of Count Twenty in No. 11-444?',
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# [3, 768]

# Get the similarity scores for the embeddings
similarities = model.similarity(embeddings, embeddings)
print(similarities)
# tensor([[1.0000, 0.5655, 0.4068],
#         [0.5655, 1.0000, 0.4290],
#         [0.4068, 0.4290, 1.0000]])

<!--

Direct Usage (Transformers)

<details><summary>Click to see the direct usage in Transformers</summary>

</details> -->

<!--

Downstream Usage (Sentence Transformers)

You can finetune this model on your own dataset.

<details><summary>Click to expand</summary>

</details> -->

<!--

Out-of-Scope Use

List how the model may foreseeably be misused and address what users ought not to do with the model. -->

Evaluation

Metrics

Information Retrieval

• —Dataset: dim_768
• —Evaluated with <code>InformationRetrievalEvaluator</code> with these parameters:

  {
      "truncate_dim": 768
  }

Information Retrieval

• —Dataset: dim_512
• —Evaluated with <code>InformationRetrievalEvaluator</code> with these parameters:

  {
      "truncate_dim": 512
  }

Information Retrieval

• —Dataset: dim_256
• —Evaluated with <code>InformationRetrievalEvaluator</code> with these parameters:

  {
      "truncate_dim": 256
  }

Information Retrieval

• —Dataset: dim_128
• —Evaluated with <code>InformationRetrievalEvaluator</code> with these parameters:

  {
      "truncate_dim": 128
  }

Information Retrieval

• —Dataset: dim_64
• —Evaluated with <code>InformationRetrievalEvaluator</code> with these parameters:

  {
      "truncate_dim": 64
  }

<!--

Bias, Risks and Limitations

What are the known or foreseeable issues stemming from this model? You could also flag here known failure cases or weaknesses of the model. -->

<!--

Recommendations

What are recommendations with respect to the foreseeable issues? For example, filtering explicit content. -->

Training Details

Training Dataset

json

• —Dataset: json
• —Size: 5,822 training samples
• —Columns: <code>positive</code> and <code>anchor</code>
• —Approximate statistics based on the first 1000 samples: | | positive | anchor | |:--------|:-----------------------------------------------------------------------------------|:----------------------------------------------------------------------------------| | type | string | string | | details | <ul><li>min: 28 tokens</li><li>mean: 96.9 tokens</li><li>max: 157 tokens</li></ul> | <ul><li>min: 8 tokens</li><li>mean: 16.49 tokens</li><li>max: 46 tokens</li></ul> |
• —Samples: | positive | anchor | |:--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|:----------------------------------------------------------------------------------------| | <code>its reasoning persuasive.45 <br> <br>44 The third case cited by the CIA, ACLU v. Department of Justice, 808 F. Supp. 2d 280, 288 (D.D.C. 2011), which <br>is from this Circuit, held that the CIA could withhold “whether the CIA cooperates with, is interested in, or actually</code> | <code>What court is ACLU v. Department of Justice associated with?</code> | | <code>summary judgment regarding the majority of the CIA’s invocations of the attorney-client <br>privilege. First, in both Nos. 11-445 and 11-444, the repeated and carefully circumscribed <br>statement that communications “relate to matters for which attorneys provide legal advice,” see, <br>e.g., First Lutz Decl. ¶ 79; Third Lutz Decl. ¶ 37, does not necessarily support the conclusion that</code> | <code>In which case numbers did the statement about communications appear?</code> | | <code>entirely from receiving IDIQ contract awards.16 AR at 1113–14, 2153–54, 2644–45. In other <br>words, the Solicitations force GSA to choose winners and losers.17 Yet, Plaintiffs suggest that the <br>self-scoring structure of the Polaris Solicitations eliminates the competitive nature of the <br>procurement. See, e.g., SHS MJAR at 21–22 (“[T]o the extent ‘competing offers’ even exist[], [it]</code> | <code>On which pages is the statement about Solicitations found besides 1113–14?</code> |
• —Loss: <code>MatryoshkaLoss</code> with these parameters:

  {
      "loss": "MultipleNegativesRankingLoss",
      "matryoshka_dims": [
          768,
          512,
          256,
          128,
          64
      ],
      "matryoshka_weights": [
          1,
          1,
          1,
          1,
          1
      ],
      "n_dims_per_step": -1
  }

Training Hyperparameters

Non-Default Hyperparameters

• —per_device_train_batch_size: 32
• —num_train_epochs: 4
• —learning_rate: 2e-05
• —lr_scheduler_type: cosine
• —warmup_steps: 500
• —optim: adamwtorchfused
• —gradient_accumulation_steps: 16
• —bf16: True
• —tf32: True
• —eval_strategy: epoch
• —per_device_eval_batch_size: 16
• —load_best_model_at_end: True
• —batch_sampler: no_duplicates

All Hyperparameters

<details><summary>Click to expand</summary>

• —per_device_train_batch_size: 32
• —num_train_epochs: 4
• —max_steps: -1
• —learning_rate: 2e-05
• —lr_scheduler_type: cosine
• —lr_scheduler_kwargs: None
• —warmup_steps: 500
• —optim: adamwtorchfused
• —optim_args: None
• —weight_decay: 0.0
• —adam_beta1: 0.9
• —adam_beta2: 0.999
• —adam_epsilon: 1e-08
• —optim_target_modules: None
• —gradient_accumulation_steps: 16
• —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: True
• —include_num_input_tokens_seen: no
• —log_level: passive
• —log_level_replica: warning
• —disable_tqdm: False
• —project: huggingface
• —trackio_space_id: trackio
• —eval_strategy: epoch
• —per_device_eval_batch_size: 16
• —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: False
• —hub_private_repo: None
• —hub_model_id: None
• —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: 42
• —data_seed: None
• —use_cpu: False
• —accelerator_config: {'splitbatches': False, 'dispatchbatches': None, 'evenbatches': True, 'useseedablesampler': True, 'nonblocking': False, 'gradientaccumulationkwargs': None}
• —parallelism_config: None
• —dataloader_drop_last: False
• —dataloader_num_workers: 0
• —dataloader_pin_memory: True
• —dataloader_persistent_workers: False
• —dataloader_prefetch_factor: None
• —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_backend: None
• —ddp_timeout: 1800
• —fsdp: []
• —fsdp_config: {'minnumparams': 0, 'xla': False, 'xlafsdpv2': False, 'xlafsdpgrad_ckpt': False}
• —deepspeed: None
• —debug: []
• —skip_memory_metrics: True
• —do_predict: False
• —resume_from_checkpoint: None
• —warmup_ratio: None
• —local_rank: -1
• —prompts: None
• —batch_sampler: no_duplicates
• —multi_dataset_batch_sampler: proportional
• —router_mapping: {}
• —learning_rate_mapping: {}

</details>

Training Logs

• —The bold row denotes the saved checkpoint.

Framework Versions

• —Python: 3.12.11
• —Sentence Transformers: 5.2.3
• —Transformers: 5.3.0
• —PyTorch: 2.5.1+cu121
• —Accelerate: 1.13.0
• —Datasets: 4.6.1
• —Tokenizers: 0.22.2

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",
}

MatryoshkaLoss

@misc{kusupati2024matryoshka,
    title={Matryoshka Representation Learning},
    author={Aditya Kusupati and Gantavya Bhatt and Aniket Rege and Matthew Wallingford and Aditya Sinha and Vivek Ramanujan and William Howard-Snyder and Kaifeng Chen and Sham Kakade and Prateek Jain and Ali Farhadi},
    year={2024},
    eprint={2205.13147},
    archivePrefix={arXiv},
    primaryClass={cs.LG}
}

MultipleNegativesRankingLoss

@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}
}

<!--

Glossary

Clearly define terms in order to be accessible across audiences. -->

<!--

Model Card Authors

Lists the people who create the model card, providing recognition and accountability for the detailed work that goes into its construction. -->

<!--

Model Card Contact

Provides a way for people who have updates to the Model Card, suggestions, or questions, to contact the Model Card authors. -->