This is a sentence-transformers model fine‑tuned for semantic similarity tasks in the Russian legal domain. It maps sentences and paragraphs to a dense 1024‑dimensional vector space and is designed for:

Semantic search

The model supports up to 8192 tokens combining fine‑tuning accuracy on short spans with restored long‑context capacity.

Model Details

Model Description

Model Type: Sentence Transformer (Bi‑Encoder)
Base Model: BAAI/bge‑m3
Fine‑tuned On: 8192‑token legal QA pairs
Technique: LM‑Cocktail (weight interpolation with base)
Output Dimensionality: 1024
Similarity Function: Cosine similarity
Domain: Russian legal texts
License: MIT

Model Sources

Documentation: sbert.net
Repository: UKPLab/sentence‑transformers
Model Hub: 🤗 sentence‑transformers models

Full Model Architecture

SentenceTransformer(
  (0): Transformer({'max_seq_length': 8192, 'do_lower_case': False}) with Transformer model: XLMRobertaModel
  (1): Pooling({'word_embedding_dimension': 1024, 'pooling_mode_cls_token': True, 'include_prompt': True})
  (2): Normalize()
)

Usage

Installation

pip install -U sentence-transformers

Quick Start

python

from sentence_transformers import SentenceTransformer, util

model = SentenceTransformer("JetTeam/legal-bge-m3-8192")

query = "Когда возможно упрощённое банкротство?"
documents = [
    "Закон о банкротстве, ст. 226. Упрощённая процедура применяется к ...",
    "Гражданский кодекс, ст. 65. Предусматривает ...",
]

query_emb = model.encode(query)
doc_embs = model.encode(documents)

scores = util.cos_sim(query_emb, doc_embs)

for doc, score in zip(documents, scores[0]):
    print(f"{score:.3f} — {doc[:60]}...")

Applications

Legal search (laws, codes, court decisions)

Training Details

Dataset

Corpus: 100 question–relevant fragment pairs annotated by legal experts. This dataset is designed for evaluating the quality of legal AI systems.
Pairs Generation: 31,673 filtered “question ↔ paragraph” samples

Training Configuration

Loss: MultipleNegativesRankingLoss
Optimizer: AdamW (lr = 1e‑5)
Batch Size: 64
Epochs: 5
Hardware: 2 × A100 (80 GB)

LM‑Cocktail

To restore long‑context support:

θ_final = 0.7 × θ_finetuned + 0.3 × θ_base

This preserves semantic quality while re‑activating the 8192‑token capacity of the base model.

Evaluation

Model	Recall@5	MRR@10
Legal BGE‑m3 (LM‑Cocktail)	0.75	0.59
BGE‑m3 (Base)	0.58	0.48
BM25	0.38	0.22

Performance

Format	FPS (Batch=2)	Latency (ms)
PyTorch FP32	3.1	480
OpenVINO FP32	8.9	180
ONNX INT8	10.7	160

INT8 may reduce Recall@5 by ≈ 1.5 pp.

Training Environment

Python: 3.10.12
Sentence Transformers: 4.0.2
Transformers: 4.48.3
PyTorch: 2.1.0+cu118
Accelerate: 1.6.0
Datasets: 3.5.0
Tokenizers: 0.21.1

Roflmax/merged_full-bge_03-07

SentenceTransformer