lawyer-llama

Why we included this project

Chinese-language legal applications rarely have open, domain-tuned models to build on, and this repository provides exactly that: LLaMA weights trained specifically for the legal domain. The training started with continual pretraining over a large corpus of statutes and court judgments, so the model holds a working grasp of China's civil, criminal, administrative, and procedural law, and then moved to instruction tuning on bar exam question explanations and lawyer-style consultation replies generated by GPT-3.5 and GPT-4. The datasets used are released alongside the weights, including per-question analyses of national bar exam problems and statute-grounded consultation dialogues, so you can reuse or adapt the recipe for your own legal assistant. The v2 model adds a marriage-law retrieval module, and the README documents automatic evaluation against a family-law consultation benchmark, giving a concrete sense of how output quality compares with general-purpose models.