chrisekwugum/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("chrisekwugum/modernbert-embed-base-legal-matryoshka-2")
# Run inference
sentences = [
    'for a specific procurement through separate joint ventures with different protégés.”  Id.  The SBA \nunderscored this purpose by highlighting that in acquiring a second protégé, the mentor “has \nalready assured SBA that the two protégés would not be competitors.  If the two mentor-protégé \nrelationships were approved in the same [North American Industry Classification System] code,',
    'What is the context of the mentor-protégé relationships mentioned?',
    "Where can the details of the CIA's framing of the plaintiff's injury be found?",
]
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

• Datasets: dim_768, dim_512, dim_256, dim_128 and dim_64
• Evaluated with InformationRetrievalEvaluator

Metric	dim_768	dim_512	dim_256	dim_128	dim_64
cosine_accuracy@1	0.5564	0.544	0.5131	0.459	0.3648
cosine_accuracy@3	0.6059	0.5873	0.5549	0.5116	0.4049
cosine_accuracy@5	0.6955	0.6847	0.6383	0.5873	0.4745
cosine_accuracy@10	0.7759	0.7604	0.7079	0.6553	0.541
cosine_precision@1	0.5564	0.544	0.5131	0.459	0.3648
cosine_precision@3	0.5265	0.5106	0.4848	0.4338	0.3483
cosine_precision@5	0.4022	0.3935	0.3713	0.3372	0.2742
cosine_precision@10	0.2388	0.2343	0.2185	0.2017	0.166
cosine_recall@1	0.1977	0.1947	0.1802	0.1668	0.1282
cosine_recall@3	0.5216	0.5067	0.478	0.433	0.3412
cosine_recall@5	0.6432	0.6318	0.5926	0.5422	0.4383
cosine_recall@10	0.7553	0.7434	0.6931	0.6388	0.5255
cosine_ndcg@10	0.662	0.6493	0.607	0.5564	0.4496
cosine_mrr@10	0.6047	0.5911	0.5553	0.5037	0.4034
cosine_map@100	0.6446	0.632	0.5963	0.5468	0.4453

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: 28 tokens mean: 97.21 tokens max: 170 tokens	min: 7 tokens mean: 16.7 tokens max: 39 tokens

• Samples:

  positive	anchor
  Counts Seven, Nine, and Ten in No. 11-445: February 6, 2010 FOIA Requests to the CIA, State Department, and NSA On February 6, 2010, the plaintiff submitted three substantially identical FOIA requests—one to the CIA, one to the State Department, and one to the National Security Agency (“NSA”). The request to the CIA sought “all current training handbooks, manuals, guidelines,	What is the number associated with the case involving Counts Seven, Nine, and Ten?
  The Government’s notion of a categorical principle stems mainly from a series of decisions in this District. Defs.’ Mem. at 14; Defs.’ Reply at 9 n.2. The first was Gates v. Schlesinger, 366 F. Supp. 797 (D.D.C. 1973), which stated that “an advisory committee is not an ‘agency.’” Id. at 799. Gates’s first rationale for this conclusion was that FACA “utilizes the definition of	From where does the Government's notion of a categorical principle mainly stem?
  sort its incoming FOIA requests based on fee categories.” First Lutz Decl. ¶ 11. The CIA’s declarant also states that “this information [i.e., fee category] is not included in the electronic system,” though the CIA’s declarant also avers that “[f]ee category is not a mandatory field,” and thus “this information is often not included in a FOIA request record.” Id. The plaintiff focuses	According to the CIA's declarant, is fee category a mandatory field?

• 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}
• 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	91.392	-	-	-	-	-
1.0	12	-	0.6238	0.6027	0.5669	0.5230	0.4009
1.7033	20	38.8819	-	-	-	-	-
2.0	24	-	0.6596	0.6423	0.5986	0.5491	0.4384
2.5275	30	28.6263	-	-	-	-	-
3.0	36	-	0.6615	0.6502	0.6058	0.5575	0.4486
3.3516	40	25.2135	-	-	-	-	-
3.7033	44	-	0.6620	0.6493	0.6070	0.5564	0.4496

• The bold row denotes the saved checkpoint.

Framework Versions

• Python: 3.11.11
• Sentence Transformers: 3.4.1
• Transformers: 4.48.3
• PyTorch: 2.5.1+cu124
• Accelerate: 1.3.0
• Datasets: 3.3.1
• Tokenizers: 0.21.0

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}
}