DISC-LawLLM is an intelligent legal system designed to provide comprehensive legal services, leveraging large language models (LLMs) for tasks like legal text processing, reasoning, and knowledge retrieval. It targets legal professionals, students, and the general public seeking legal assistance, offering benefits such as enhanced legal understanding and efficient information access.

How It Works

DISC-LawLLM builds upon the Baichuan-13B model, fine-tuned with a custom dataset (DISC-Law-SFT) that includes legal reasoning (tri-syllogism) and knowledge retrieval capabilities. It incorporates a retrieval-augmented module, utilizing Langchain-Chatchat, to access a knowledge base of over 800 legal statutes and 24,000 legal exam questions, aiming for improved accuracy and context adherence in legal scenarios.

Quick Start & Requirements

• Install: pip install -r requirements.txt
• Prerequisites: Python, PyTorch. GPU recommended for inference.
• Demo: http://law.fudan-disc.com
• Model Weights: Huggingface Repo
• Technical Report: arxiv.org/abs/2309.11325

Highlighted Details

• Achieves performance on the Lawbench benchmark that is second only to GPT-4, outperforming other legal LLMs.
• Offers a comprehensive evaluation framework (DISC-Law-Eval Benchmark) for assessing legal LLMs.
• Provides a large, high-quality supervised fine-tuning dataset (DISC-Law-SFT) with over 400K samples covering various legal tasks.
• Includes a retrieval module with a knowledge base of statutes and legal exam questions.

Maintenance & Community

• Developed by Fudan University's Data Intelligence and Social Computing Lab (Fudan-DISC).
• Recent updates include the release of legal Q&A datasets and acceptance of a research paper at DASFAA 2024.

Licensing & Compatibility

• Licensed under the Apache License.
• Permits commercial use and linking with closed-source applications.

Limitations & Caveats

The project acknowledges that LLMs have inherent limitations and that DISC-LawLLM should be used for reference only, not as a substitute for professional legal advice. The developers disclaim responsibility for any issues arising from its use.