manishh16/modernbert-embed-base-legal-matryoshka-2

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("manishh16/modernbert-embed-base-legal-matryoshka-2")
# Run inference
sentences = [
    'protests pursuant to 28 U.S.C. § 1491(b).  See 28 U.S.C. § 1491(b).  Section 1491(b)(1) grants the \n17 \n \ncourt jurisdiction over protests filed “by an interested party objecting to a solicitation by a Federal \nagency for bids or proposals for a proposed contract . . . or any alleged violation of statute or',
    'Under which U.S. Code section are the protests filed?',
    "Which agency's declaration is mentioned?",
]
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	0.592
cosine_accuracy@3	0.6352
cosine_accuracy@5	0.7032
cosine_accuracy@10	0.7666
cosine_precision@1	0.592
cosine_precision@3	0.5683
cosine_precision@5	0.4263
cosine_precision@10	0.2408
cosine_recall@1	0.2012
cosine_recall@3	0.547
cosine_recall@5	0.6664
cosine_recall@10	0.7508
cosine_ndcg@10	0.6774
cosine_mrr@10	0.6317
cosine_map@100	0.6707

Information Retrieval

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

  {
      "truncate_dim": 512
  }
  

Metric	Value
cosine_accuracy@1	0.5858
cosine_accuracy@3	0.6167
cosine_accuracy@5	0.6909
cosine_accuracy@10	0.7666
cosine_precision@1	0.5858
cosine_precision@3	0.5574
cosine_precision@5	0.4176
cosine_precision@10	0.2417
cosine_recall@1	0.1984
cosine_recall@3	0.5353
cosine_recall@5	0.6515
cosine_recall@10	0.7518
cosine_ndcg@10	0.6722
cosine_mrr@10	0.6236
cosine_map@100	0.662

Information Retrieval

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

  {
      "truncate_dim": 256
  }
  

Metric	Value
cosine_accuracy@1	0.5672
cosine_accuracy@3	0.5873
cosine_accuracy@5	0.6646
cosine_accuracy@10	0.7311
cosine_precision@1	0.5672
cosine_precision@3	0.5384
cosine_precision@5	0.4009
cosine_precision@10	0.2308
cosine_recall@1	0.1906
cosine_recall@3	0.5152
cosine_recall@5	0.6264
cosine_recall@10	0.7205
cosine_ndcg@10	0.6454
cosine_mrr@10	0.6009
cosine_map@100	0.6377

Information Retrieval

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

  {
      "truncate_dim": 128
  }
  

Metric	Value
cosine_accuracy@1	0.4992
cosine_accuracy@3	0.5301
cosine_accuracy@5	0.6136
cosine_accuracy@10	0.6785
cosine_precision@1	0.4992
cosine_precision@3	0.4745
cosine_precision@5	0.3654
cosine_precision@10	0.2159
cosine_recall@1	0.1695
cosine_recall@3	0.4581
cosine_recall@5	0.5706
cosine_recall@10	0.6683
cosine_ndcg@10	0.5892
cosine_mrr@10	0.5386
cosine_map@100	0.5783

Information Retrieval

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

  {
      "truncate_dim": 64
  }
  

Metric	Value
cosine_accuracy@1	0.3632
cosine_accuracy@3	0.4019
cosine_accuracy@5	0.473
cosine_accuracy@10	0.527
cosine_precision@1	0.3632
cosine_precision@3	0.3514
cosine_precision@5	0.2782
cosine_precision@10	0.1651
cosine_recall@1	0.1236
cosine_recall@3	0.3391
cosine_recall@5	0.4364
cosine_recall@10	0.5143
cosine_ndcg@10	0.4444
cosine_mrr@10	0.4003
cosine_map@100	0.4462

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: 26 tokens mean: 96.76 tokens max: 156 tokens	min: 8 tokens mean: 16.59 tokens max: 49 tokens

• Samples:

  positive	anchor
  properly authenticated. See id. at 367, 19 A.3d at 429 (Harrell, J., dissenting). Four years later, in Sublet, 442 Md. at 637-38, 113 A.3d at 697-98, we adopted the reasonable juror test for social media evidence and applied it in the three cases that were consolidated for purposes of the opinion: Sublet v. State, Harris v. State, and Monge-	How many years after the dissent did the adoption of the reasonable juror test occur?
  to (1) a public-interest fee waiver, (2) the expedited processing of a request, or (3) the release of information that implicates personal privacy, all are personal to a requester and thus cannot be assigned. See, e.g., RTC Commercial Loan Trust 1995-NP1A v. Winthrop Mgmt., 923 F. Supp. 83, 88 (E.D. Va. 1996) (holding that “certain rights are purely personal and cannot be assigned”).	What type of fee waiver is mentioned as being personal to a requester?
  ‘IRO’] staff that reviews Agency records and makes public release determinations with an eye toward evaluating directorate-specific equities.” Id. ¶ 4. Ms. Meeks also explains that “records frequently involve the equities of multiple directorates,” and “[w]hen records implicate the operational interests of multiple directorates, the reviews are conducted by the relevant IROs	Who conducts the reviews when the records implicate the operational interests of multiple directorates?

• 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
• tf32: False
• 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: False
• 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
• 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

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	91.6964	-	-	-	-	-
1.0	12	-	0.6483	0.6445	0.6004	0.5232	0.4001
1.7033	20	39.6429	-	-	-	-	-
2.0	24	-	0.6764	0.6716	0.6361	0.5736	0.4374
2.5275	30	30.1905	-	-	-	-	-
3.0	36	-	0.6768	0.6699	0.6441	0.5869	0.4416
3.3516	40	26.8879	-	-	-	-	-
3.7033	44	-	0.6774	0.6722	0.6454	0.5892	0.4444

• The bold row denotes the saved checkpoint.

Framework Versions

• Python: 3.11.11
• Sentence Transformers: 4.0.1
• Transformers: 4.50.2
• PyTorch: 2.6.0+cu124
• Accelerate: 1.5.2
• Datasets: 3.5.0
• 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}
}