Property	Value
Base model	`rufimelo/Legal-BERTimbau-base` (Legal-BERTimbau)
Base model pretraining	Brazilian legal corpus (legislation, court decisions)
Fine-tuning method	LoRA (r=16, alpha=32, dropout=0.1)
Target modules	`query`, `value`
Trainable parameters	593,669 (0.54% of total)
Training data	50,000 samples (oversampled from 17,462 labeled)
Max sequence length	128 tokens
Training epochs	5 (with early stopping, patience=2)
GPU	Google Colab T4 (~35 min)

Performance

Main Results (Test Set: 250 samples, balanced)

Metric	Score
Accuracy	98.00%
F1 Macro	0.9800
Inference latency	85.2 ms (CPU)

Per-Class F1 Scores

Legal Area	F1 Score
Administrativo	0.980
Civel	0.980
Consumidor	0.980
Penal	0.980
Tributario	0.980

Comparison with LLMs (0-shot)

Model	Accuracy	F1 Macro	Latency
BERTimbau v2	99.60%	0.9960	86 ms
Legal-BERTimbau v2 (this model)	98.00%	0.9800	85 ms
Claude 3.5 Haiku	77.20%	0.7534	1,192 ms
GPT-4o	74.80%	0.7095	1,048 ms
GPT-4o-mini	55.60%	0.5243	969 ms
Llama 3.1 8B	47.20%	0.3811	811 ms

Why does general BERTimbau outperform Legal-BERTimbau?

This is a key finding from our benchmark: the general-purpose BERTimbau (99.6%) slightly outperforms the domain-specific Legal-BERTimbau (98.0%). We hypothesize that Legal-BERTimbau's pretraining on full legal text creates representations optimized for legal reasoning, while this classification task relies on metadata patterns (procedural class + subject codes). The general encoder's broader vocabulary may better capture these structured patterns. See our paper for detailed analysis.

Labels

python

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

Usage

python

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model_name = "pedronettotrue/legal-bertimbau-tjsc-v2"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

ID2LABEL = {0: "civel", 1: "consumidor", 2: "tributario", 3: "administrativo", 4: "penal"}

text = "Classe: Apelacao Civel. Assuntos: Responsabilidade Civil, Danos Morais"
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=128)

with torch.no_grad():
    logits = model(**inputs).logits
    predicted_class = ID2LABEL[logits.argmax().item()]

print(f"Predicted: {predicted_class}")
# Output: civel

Training Details

Dataset: LegalBench-BR v2 -- 50,000 training samples from 5 Brazilian courts (TJSC, TJSP, TJRJ, STJ, TJMG)
Labeling: Deterministic 3-layer CNJ taxonomy matching (subject, procedural class, court division) -- no LLM-based labeling
Optimizer: AdamW (lr=2e-4, weight_decay=0.01, warmup_ratio=0.1)
Precision: FP16
Framework: HuggingFace Transformers + PEFT (LoRA)
Post-training: LoRA weights merged via merge_and_unload() -- loads without PEFT dependency

Dataset

The model was trained on metadata from Brazilian court decisions (procedural class + legal subjects), not full-text judicial opinions.

Courts covered in training:

TJSC (Tribunal de Justica de Santa Catarina)
TJSP (Tribunal de Justica de Sao Paulo)
TJRJ (Tribunal de Justica do Rio de Janeiro)
STJ (Superior Tribunal de Justica)
TJMG (Tribunal de Justica de Minas Gerais)

Limitations

Trained on judicial metadata (procedural class + subjects), not full-text legal reasoning
5 legal areas only (does not cover labor, environmental, family law, etc.)
General BERTimbau outperforms this model on this specific task (see discussion above)
Performance on courts not in the training set has not been evaluated
Input is truncated at 128 tokens

Citation

bibtex

@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/pedronettotrue/legal-bertimbau-tjsc-v2}
}

License

Apache 2.0

pedronettotrue/legalbench-br-v2-legal-bertimbau

Legal-BERTimbau TJSC v2

Model Details