Thejina/modernbert-embed-base-legal-matryoshka-2-new

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
• 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}) with Transformer model: 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("Thejina/modernbert-embed-base-legal-matryoshka-2-new")
# Run inference
sentences = [
    'Homeland Sec., No. 12-856, 2013 WL 3186061, at *18 (D.D.C. June 24, 2013) (citing In re \nSealed Case, 737 F.2d at 100).  The “subjective intentions” of confidentiality put forth by the \nCIA are therefore insufficient to establish “confidentiality in fact.”  Id. \nThe third and final deficiency manifests only in the CIA’s submissions in No. 11-445.',
    'On what date was the cited decision in the D.D.C. court made?',
    'Which organization is associated with the Exemption 3 withholdings discussed in Part III.E.3?',
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# [3, 768]

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

Evaluation

Metrics

Information Retrieval

• Dataset: dim_768
• Evaluated with InformationRetrievalEvaluator with these parameters:

  {
      "truncate_dim": 768
  }
  

Metric	Value
cosine_accuracy@1	0.5564
cosine_accuracy@3	0.5873
cosine_accuracy@5	0.6677
cosine_accuracy@10	0.7512
cosine_precision@1	0.5564
cosine_precision@3	0.5234
cosine_precision@5	0.3892
cosine_precision@10	0.2328
cosine_recall@1	0.1984
cosine_recall@3	0.5162
cosine_recall@5	0.6255
cosine_recall@10	0.7414
cosine_ndcg@10	0.652
cosine_mrr@10	0.5976
cosine_map@100	0.6379

Information Retrieval

• Dataset: dim_512
• Evaluated with InformationRetrievalEvaluator with these parameters:

  {
      "truncate_dim": 512
  }
  

Metric	Value
cosine_accuracy@1	0.544
cosine_accuracy@3	0.5734
cosine_accuracy@5	0.6538
cosine_accuracy@10	0.7481
cosine_precision@1	0.544
cosine_precision@3	0.5106
cosine_precision@5	0.3784
cosine_precision@10	0.2304
cosine_recall@1	0.1947
cosine_recall@3	0.5059
cosine_recall@5	0.6096
cosine_recall@10	0.7366
cosine_ndcg@10	0.6424
cosine_mrr@10	0.5859
cosine_map@100	0.6256

Information Retrieval

• Dataset: dim_256
• Evaluated with InformationRetrievalEvaluator with these parameters:

  {
      "truncate_dim": 256
  }
  

Metric	Value
cosine_accuracy@1	0.5193
cosine_accuracy@3	0.5518
cosine_accuracy@5	0.6445
cosine_accuracy@10	0.7125
cosine_precision@1	0.5193
cosine_precision@3	0.4884
cosine_precision@5	0.3694
cosine_precision@10	0.2185
cosine_recall@1	0.1859
cosine_recall@3	0.4818
cosine_recall@5	0.5936
cosine_recall@10	0.6963
cosine_ndcg@10	0.6116
cosine_mrr@10	0.56
cosine_map@100	0.6015

Information Retrieval

• Dataset: dim_128
• Evaluated with InformationRetrievalEvaluator with these parameters:

  {
      "truncate_dim": 128
  }
  

Metric	Value
cosine_accuracy@1	0.4343
cosine_accuracy@3	0.4776
cosine_accuracy@5	0.5641
cosine_accuracy@10	0.6553
cosine_precision@1	0.4343
cosine_precision@3	0.4116
cosine_precision@5	0.3184
cosine_precision@10	0.2002
cosine_recall@1	0.1597
cosine_recall@3	0.4137
cosine_recall@5	0.5185
cosine_recall@10	0.6381
cosine_ndcg@10	0.5422
cosine_mrr@10	0.4823
cosine_map@100	0.5293

Information Retrieval

• Dataset: dim_64
• Evaluated with InformationRetrievalEvaluator with these parameters:

  {
      "truncate_dim": 64
  }
  

Metric	Value
cosine_accuracy@1	0.3246
cosine_accuracy@3	0.3648
cosine_accuracy@5	0.456
cosine_accuracy@10	0.5348
cosine_precision@1	0.3246
cosine_precision@3	0.3107
cosine_precision@5	0.2485
cosine_precision@10	0.1623
cosine_recall@1	0.1207
cosine_recall@3	0.3131
cosine_recall@5	0.4083
cosine_recall@10	0.5252
cosine_ndcg@10	0.4298
cosine_mrr@10	0.3712
cosine_map@100	0.4185

Training Details

Training Dataset

json

• Dataset: json
• Size: 5,822 training samples
• Columns: positive and anchor
• Approximate statistics based on the first 1000 samples:

  	positive	anchor
  type	string	string
  details	min: 29 tokens mean: 97.85 tokens max: 170 tokens	min: 7 tokens mean: 16.58 tokens max: 38 tokens

• Samples:

  positive	anchor
  n.2. But the Court cannot simply adopt this concession. “[S]ubject-matter jurisdiction, because it involves a court’s power to hear a case, can never be forfeited or waived.” United States v. Cotton, 535 U.S. 625, 630 (2002). The Court thus has “an independent obligation to determine whether subject-matter jurisdiction exists.” Arbaugh v. Y&H Corp., 546 U.S. 500, 514 (2006).	According to Cotton, what can never be forfeited or waived?
  another because they involve common factual and legal issues. See Notice of Related Case, No. 11-444, ECF No. 2; Notice of Related Case, No. 11-445, ECF No. 2. Although the Court has not formally consolidated these actions, due to their interrelated nature and in the interests of judicial economy the Court has adjudicated dispositive motions	Has the Court formally consolidated the actions from the Notices of Related Case?
  [PROSECUTOR]: He’s authenticated it as to be the date and the time of the incident, it was a true and accurate reflection of that date and time. THE COURT: There are other questions you need to ask him, like, has he watched it. [PROSECUTOR]: Okay. THE COURT: And is it a fair and accurate representation of what happened. I mean, I’m not trying -- [PROSECUTOR]: Okay.	What did the prosecutor confirm about the date and time?

• Loss: MatryoshkaLoss 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

• eval_strategy: epoch
• per_device_train_batch_size: 32
• per_device_eval_batch_size: 16
• gradient_accumulation_steps: 16
• learning_rate: 2e-05
• num_train_epochs: 4
• lr_scheduler_type: cosine
• warmup_ratio: 0.1
• bf16: True
• load_best_model_at_end: True
• optim: adamw_torch_fused
• batch_sampler: no_duplicates

All Hyperparameters

Click to expand

• overwrite_output_dir: False
• do_predict: False
• eval_strategy: epoch
• prediction_loss_only: True
• per_device_train_batch_size: 32
• per_device_eval_batch_size: 16
• per_gpu_train_batch_size: None
• per_gpu_eval_batch_size: None
• gradient_accumulation_steps: 16
• 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: 4
• max_steps: -1
• lr_scheduler_type: cosine
• 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}
• tp_size: 0
• 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_fused
• 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

Training Logs

Epoch	Step	Training Loss	dim_768_cosine_ndcg@10	dim_512_cosine_ndcg@10	dim_256_cosine_ndcg@10	dim_128_cosine_ndcg@10	dim_64_cosine_ndcg@10
0.8791	10	90.1655	-	-	-	-	-
1.0	12	-	0.6103	0.5866	0.5539	0.4942	0.3692
1.7033	20	39.3527	-	-	-	-	-
2.0	24	-	0.6500	0.6321	0.6009	0.5327	0.4111
2.5275	30	30.0319	-	-	-	-	-
3.0	36	-	0.6521	0.6454	0.6096	0.5400	0.4305
3.3516	40	27.1479	-	-	-	-	-
3.7033	44	-	0.652	0.6424	0.6116	0.5422	0.4298

• The bold row denotes the saved checkpoint.

Framework Versions

• Python: 3.11.12
• Sentence Transformers: 4.1.0
• Transformers: 4.51.3
• PyTorch: 2.6.0+cu124
• Accelerate: 1.6.0
• Datasets: 3.6.0
• 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",
}

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}
}