Research & Papers

Domain-Adaptive Dense Retrieval for Brazilian Legal Search

Legal-only vs. mixed fine-tuning: trade-offs in Brazilian legal retrieval

Deep Dive

Jayr Pereira, Roberto Lotufo, and Luiz Bonifacio explored how to train dense retrievers for heterogeneous Brazilian legal search, which spans case law, legislation, and question-based queries. Using Qwen3-Embedding-4B as the base model, they tested three setups: no fine-tuning, legal-only fine-tuning, and a mixed setup combining legal data with the SQuAD-pt Portuguese supervised dataset. Evaluation on five datasets from the JUÁ leaderboard and the Quati Portuguese retrieval benchmark revealed clear trade-offs.

The legal-only model aces specialized legal tasks, but the mixed setup delivers the best overall balance. It improved average NDCG@10 from 0.414 to 0.447, MRR@10 from 0.586 to 0.595, and MAP@10 from 0.270 to 0.308 across all six datasets. The biggest gains came on Quati (out-of-domain question answering), where the mixed model clearly outperforms the legal-only version. The authors conclude legal-only specializes, while mixed training is more robust—especially for question-based queries. Both adapted models are available on Hugging Face.

Key Points
  • Based on Qwen3-Embedding-4B with three training regimes: base, legal-only, mixed (legal + SQuAD-pt).
  • Mixed setup boosts average NDCG@10 from 0.414 to 0.447, MRR@10 from 0.586 to 0.595, MAP@10 from 0.270 to 0.308.
  • Legal-only excels at specialized legal tasks; mixed is better for general robustness and question-based search.

Why It Matters

Better legal search in Brazil means faster case law and legislation retrieval for professionals, with open-source models.