ai-eldorado/Brazilian_CLT_lora

Model Description
This repository provides a LoRA adapter trained on the same human-preference dataset used for the DPO model. It enables efficient fine-tuning of LLaMA-3 8B Instruct (4-bit) for Brazilian labor law applications without requiring full model retraining.

Intended Use
The LoRA adapter is intended for developers and researchers who want to adapt LLaMA-3 models for CLT-related tasks while minimizing computational costs. It is particularly useful for resource-constrained environments or for integrating into multi-agent legal assistant systems.

Training Details

• Base Model: LLaMA-3 8B (4-bit quantized)
• Method: LoRA fine-tuning with DPO-aligned dataset
• Dataset: 736 human-preference entries on CLT-related questions
• Hyperparameters: Same as the DPO model

Performance Summary
When merged with the base model, the LoRA adapter reproduces the improvements observed in the full DPO model:

• Increased factual accuracy
• Better semantic alignment with CLT regulations

Ethical Considerations

• Legal Disclaimer: Not a substitute for professional legal advice.
• Risk of Misuse: Incorrect merging or use outside intended domain may lead to inaccurate outputs.
• Data Privacy: No personal data was used in training.

Bias and Fairness

• Same considerations as the DPO model:
  • Regional and interpretation biases may exist.
  • Limited dataset size could affect fairness in edge cases.

Limitations

• Requires merging with the base LLaMA-3 model for inference.
• Domain-specific; not suitable for general-purpose legal reasoning.

Citation

@article{moraescomparing,
  title={Comparing RAG, DPO and Agentic Approaches in Systems Performance on Q\&A about Brazilian Labor Legislation},
  author={Moraes, Gabriel K and Luiz, Pedro Augusto and Dias, Gabriel and de Farias, Vitor GCB and Fabiana, CQ de O and Fabris, Vitor L and Vicente, Matheus HR and do Nascimento, Leonardo R and Oliveira, Charles S and dos Santos, Leonardo T and others}
}