ModernBERT Embed base Legal Matryoshka

Name: ModernBERT-embed-base-legal-MRL
Rating: 2.5 (468 reviews)

This is a sentence-transformers model finetuned from nomic-ai/modernbert-embed-base on the AdamLucek/legal-rag-positives-synthetic 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:
- AdamLucek/legal-rag-positives-synthetic
Language: en
License: apache-2.0

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.

python

from sentence_transformers import SentenceTransformer

# Download from the 🤗 Hub
model = SentenceTransformer("AdamLucek/ModernBERT-embed-base-legal-MRL")
# Run inference
sentences = [
    'contracting/contracting-assistance-programs/sba-mentor-protege-program (last visited Apr. 19, \n2023). \n5 \n \nprotégé must demonstrate that the added mentor-protégé relationship will not adversely affect the \ndevelopment of either protégé firm (e.g., the second firm may not be a competitor of the first \nfirm).”  13 C.F.R. § 125.9(b)(3).',
    'What must the protégé demonstrate about the mentor-protégé relationship?',
    'What discretion do district courts have regarding a defendant’s invocation of FOIA exemptions?',
]
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	0.5286	0.5162	0.4822	0.4158	0.3122
cosine_accuracy@3	0.5719	0.5487	0.5286	0.4436	0.3509
cosine_accuracy@5	0.6646	0.6414	0.5981	0.5363	0.4359
cosine_accuracy@10	0.7311	0.7172	0.6785	0.6105	0.4791
cosine_precision@1	0.5286	0.5162	0.4822	0.4158	0.3122
cosine_precision@3	0.5142	0.4982	0.4699	0.3993	0.3091
cosine_precision@5	0.3941	0.3808	0.3586	0.3128	0.2504
cosine_precision@10	0.2329	0.2272	0.2147	0.1924	0.1498
cosine_recall@1	0.1788	0.174	0.1627	0.1426	0.105
cosine_recall@3	0.4894	0.4735	0.4493	0.3836	0.2955
cosine_recall@5	0.6121	0.5911	0.5569	0.4878	0.3931
cosine_recall@10	0.7184	0.7023	0.6642	0.5963	0.4681
cosine_ndcg@10	0.63	0.6138	0.5781	0.5109	0.3956
cosine_mrr@10	0.5741	0.5593	0.5249	0.4573	0.3509
cosine_map@100	0.6186	0.6022	0.5698	0.503	0.3939

Training Details

AdamLucek/legal-rag-positives-synthetic

Dataset: AdamLucek/legal-rag-positives-synthetic
Size: 5,822 training samples
Columns: positive and anchor
Approximate statistics based on the first 1000 samples:
positive anchor
type string string
details
min: 15 tokens
mean: 97.6 tokens
max: 153 tokens
min: 8 tokens
mean: 16.68 tokens
max: 41 tokens

Samples:

positive	anchor
`infrastructure security information,” the information at issue must, “if disclosed . . . reveal vulner-[object Object]abilities in Department of Defense critical infrastructure.” 10 U.S.C. § 130e(f). The closest the [object Object]Department comes is asserting that the information “individually or in the aggregate, would enable`	`What type of information must reveal vulnerabilities if disclosed?`
`they have bid.” Oral Arg. Tr. at 42:18–20. Plaintiffs also assert that, should this Court require the [object Object]Polaris Solicitations to consider price at the IDIQ level, such an adjustment “adds a solicitation [object Object]requirement that would necessarily change the overall structure of the evaluation” GSA must [object Object]perform in awarding the IDIQ contracts. Oral Arg. Tr. at 43:3–5; see supra Discussion Section`	`Where in the document can further discussion about the assertion be found?`
`otra parte. Fernández v. San Juan Cement Co., Inc., 118 DPR 713, [object Object]718-719 (1987). Nuestro más Alto Foro ha dispuesto que, la [object Object]facultad de imponer honorarios de abogados es la mejor arma que [object Object] [object Object]22 Id. [object Object]23 Andamios de PR v. Newport Bonding, 179 DPR 503, 520 (2010); Pérez Rodríguez [object Object]v. López Rodríguez, supra; SLG González -Figueroa v. Pacheco Romero, supra;`	`What case is cited with the reference number 118 DPR 713?`

Loss: MatryoshkaLoss with these parameters:
json
1{ 2 "loss": "MultipleNegativesRankingLoss", 3 "matryoshka_dims": [ 4 768, 5 512, 6 256, 7 128, 8 64 9 ], 10 "matryoshka_weights": [ 11 1, 12 1, 13 1, 14 1, 15 1 16 ], 17 "n_dims_per_step": -1 18}

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: 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: True
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	5.6528	-	-	-	-	-
1.0	12	-	0.5926	0.5753	0.5457	0.4687	0.3455
1.7033	20	2.4543	-	-	-	-	-
2.0	24	-	0.6195	0.6066	0.5778	0.4998	0.3828
2.5275	30	1.7455	-	-	-	-	-
3.0	36	-	0.6292	0.6135	0.5765	0.5057	0.3928
3.3516	40	1.5499	-	-	-	-	-
3.7033	44	-	0.63	0.6138	0.5781	0.5109	0.3956

The bold row denotes the saved checkpoint.

Framework Versions

Python: 3.11.11
Sentence Transformers: 3.3.1
Transformers: 4.48.0
PyTorch: 2.5.1+cu121
Accelerate: 1.2.1
Datasets: 3.2.0
Tokenizers: 0.21.0

Citation

BibTeX

Sentence Transformers

bibtex

@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

bibtex

@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

bibtex

@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}
}

	positive	anchor
type	string	string
details	min: 15 tokens mean: 97.6 tokens max: 153 tokens	min: 8 tokens mean: 16.68 tokens max: 41 tokens