DISC-LawLLM

Why we included this project

Teams building Chinese legal software often find general-purpose models fall short when it comes to statutes and judgments, and this project removes the hardest part of correcting that. Fudan's DISC-LawLLM ships fine-tuned 7B and 13B weights plus a supervised dataset of roughly 300K samples built for legal tasks, so you can run the assistant as is or keep fine-tuning on your own corpus. Its design relies on two things that matter for real use: syllogism-style reasoning for predicting judgments and drafting professional answers, plus a retrieval module that grounds responses in Chinese statutes, regulations, and judicial documents. There is also an evaluation benchmark with reference outputs, which lets you compare models on a consistent legal suite instead of whatever ad-hoc tests you would put together. Teams starting on document summarization, contract drafting, case classification, or consultation Q&A will save considerable data-building effort.