网站介绍
北大开源的法律大模型和助手
官网公开介绍
ChatLaw：A Powerful LLM Tailored for Chinese Legal. 中文法律大模型 - PKU-YuanGroup/ChatLaw
Latest commit
Folders and files
Repository files navigation
Chatlaw: A Large Language Model-based Multi-Agent Legal Assistant Enhanced by Knowledge Graph and Mixture-of-Experts.
arXiv · HuggingFace
ChatLaw Models
ChatLaw2-MoE
Latest Version: Based on the InternLM architecture with a 4x7B Mixture of Experts (MoE) design.
Specialization: Tailored for Chinese legal language processing、
ChatLaw-Text2Vec
Introduction
AI legal assistants, powered by Large Language Models (LLMs), offer accessible legal consulting. However, the risk of hallucination in AI responses is a concern. This paper introduces ChatLaw, an innovative assistant that employs a Mixture-of-Experts (MoE) model and a multi-agent system to enhance reliability and accuracy in AI legal services. By integrating knowledge graphs and artificial screening, we've created a high-quality legal dataset for training the MoE model. This model leverages various experts to address a range of legal issues, optimizing legal response accuracy. Standardized Operating Procedures (SOPs), inspired by law firm workflows, significantly minimize errors and hallucinations.
Our MoE model outperforms GPT-4 in the Lawbench and Unified Qualification Exam for Legal Professionals, achieving a 7.73% higher accuracy and an 11-point lead, respectively. It also surpasses other models in real-case consultations across multiple dimensions, showcasing robust legal consultation capabilities.
Multi-Agents Collaboration Process and Legal Consultation Report
The diagram below illustrates the collaborative process of multiple agents in providing legal consultation services, exemplified by a divorce consultation. The process involves gathering information, legal research, comprehensive advice, and culminates in a detailed Legal Consultation Report.
Data Visualization and Performance Comparison
(a) Our legal dataset covers a diverse range of tasks, from case classification to public opinion analysis.