pedronettotrue/legal-nlp-benchmark-br

LegalBench-BR: Brazilian Legal NLP Benchmark

The first public Portuguese-language benchmark for automatic legal area classification of Brazilian court decisions, covering 5 courts and 5 legal domains.

Dataset Description

LegalBench-BR v2 is a multi-court benchmark for evaluating NLP models on Brazilian legal text classification. It contains judicial metadata (procedural class + legal subjects) collected from the Brazilian National Council of Justice (CNJ) DataJud API.

Key Features

• Multi-court: 5 Brazilian courts (TJSC, TJSP, TJRJ, STJ, TJMG)
• Balanced classes: 5 legal areas with stratified splits
• Deterministic labeling: 3-layer CNJ taxonomy matching (no LLM labeling)
• Benchmark-ready: includes 12-model leaderboard (encoders + LLMs)
• Reproducible: all collection and processing scripts are public

Dataset Statistics

Split	Samples	Description
Train	50,000	Oversampled from 17,462 unique labeled records
Validation	250	Balanced: 5 courts x 5 classes x 10 samples
Test	250	Balanced: 5 courts x 5 classes x 10 samples

Source Courts

Court	Abbreviation	Records
Tribunal de Justica de Santa Catarina	TJSC	4,926
Tribunal de Justica de Sao Paulo	TJSP	4,511
Tribunal de Justica do Rio de Janeiro	TJRJ	4,186
Superior Tribunal de Justica	STJ	3,917
Tribunal de Justica de Minas Gerais	TJMG	5,623
Total unique labeled		17,462 (from 23,163 raw)

Legal Area Distribution

Label	ID	Description
civel	0	Civil law (contracts, property, torts)
consumidor	1	Consumer protection law
tributario	2	Tax law
administrativo	3	Administrative law
penal	4	Criminal law

LABEL2ID = {"civel": 0, "consumidor": 1, "tributario": 2, "administrativo": 3, "penal": 4}

Labeling Methodology

Records are labeled using a deterministic 3-layer CNJ taxonomy (no LLM involved):

1. Layer 1 -- Subject matching: CNJ subject codes are mapped to legal areas using the official Tabelas Processuais Unificadas
2. Layer 2 -- Procedural class: The procedural class (e.g., "Apelacao Criminal") provides strong legal area signal
3. Layer 3 -- Court division: Division names (e.g., "Vara Criminal") serve as final tiebreaker

Records that cannot be deterministically classified are discarded (acceptance rate: 75.4%).

Data Format

Each record in the JSONL files contains:

{
  "id": "tjsp_2024_001234",
  "texto_norm": "Classe: Apelacao Civel. Assuntos: Responsabilidade Civil, Danos Morais",
  "label": "civel",
  "label_id": 0,
  "tribunal": "TJSP"
}

Usage

Loading v2 (recommended)

import json

def load_split(path):
    records = []
    with open(path, encoding="utf-8") as f:
        for line in f:
            if line.strip():
                records.append(json.loads(line))
    return records

# After downloading the dataset
train = load_split("v2/train.jsonl")
test = load_split("v2/test.jsonl")
print(f"Train: {len(train)}, Test: {len(test)}")

With HuggingFace datasets

from datasets import load_dataset

ds = load_dataset("pedronettotrue/legal-nlp-benchmark-br", data_files={
    "train": "v2/train.jsonl",
    "validation": "v2/validation.jsonl",
    "test": "v2/test.jsonl",
})
print(ds)

Benchmark Leaderboard

Results on the test set (250 samples, balanced across 5 courts x 5 classes):

Rank	Model	Type	Accuracy	F1 Macro	Latency (ms)
1	BERTimbau v2	Encoder (fine-tuned)	99.60%	0.9960	86
2	Legal-BERTimbau v2	Encoder (fine-tuned)	98.00%	0.9800	85
3	Claude 3.5 Haiku (0-shot)	LLM API	77.20%	0.7534	1,192
4	Claude 3.5 Haiku (3-shot)	LLM API	76.40%	0.7536	1,227
5	GPT-4o (0-shot)	LLM API	74.80%	0.7095	1,048
6	Claude Haiku (0-shot)	LLM API	73.60%	0.7219	1,137
7	Qwen2.5-72B (0-shot)	LLM Local	68.90%	0.6600	4,908
8	GPT-4o-mini (3-shot)	LLM API	65.20%	0.6192	1,045
9	GPT-4o-mini (0-shot)	LLM API	55.60%	0.5243	969
10	Llama 3.1 70B (0-shot)	LLM Local	62.40%	0.5832	1,319
11	Llama 3.1 8B (0-shot)	LLM Local	47.20%	0.3811	811
12	Qwen2.5 7B (0-shot)	LLM Local	44.00%	0.3492	1,279

Key Findings

• Fine-tuned encoders dominate: BERTimbau (99.6%) outperforms the best LLM (Claude 3.5 Haiku, 77.2%) by +22.4 percentage points
• Speed advantage: Encoders run at ~86ms vs ~1,100ms+ for LLMs (13x faster)
• Cost advantage: Encoders run locally at zero API cost
• General > domain-specific: BERTimbau (general) slightly outperforms Legal-BERTimbau (legal domain) on this task
• Consumer law is hardest: All LLMs struggle with consumidor class (F1 as low as 0.07 for smaller models)

Source

Data collected from the DataJud API of the Brazilian National Council of Justice (CNJ), which provides public access to judicial process metadata.

Citation

@misc{legalbenchbr2026,
  title={LegalBench-BR v2: A Multi-Court Benchmark Demonstrating Fine-Tuned Encoder
         Superiority over LLMs for Brazilian Legal Area Classification},
  year={2026},
  url={https://huggingface.co/datasets/pedronettotrue/legal-nlp-benchmark-br}
}

License

CC BY 4.0

The underlying judicial data is public by Brazilian law (Lei de Acesso a Informacao, Lei 12.527/2011). Personal data has been excluded in compliance with LGPD (Lei 13.709/2018).