Code and Data for Knowledge-guided Data Generation

A tool for generating and polishing legal question-answering data using large language models. We provide 50K legal question-answering data for training, including 25K standard version (./data/Standard-25K.json) and 25K enhanced version with reasoning paths (./data/Reasoning-25K.json).

Environment

pip install -r requirements.txt

Usage

The project consists of three main components:

• ./src/generate.py: Generates initial legal QA data by utilizing seed questions from seed.json and legal knowledge from the reference directory
• ./src/polish.py: Enhances the generated data by validating legal references and optimizing reasoning paths
• ./src/verify.py: Performs quality assurance by checking the accuracy and logical consistency of answers, reasoning paths, and legal references

You should prepare your knowledge base in the reference directory. Here, we provide a sample knowledge base with two legal documents. And the seed problem set shoud be provided in seed.json. Finally, you can sequentially run the three components to generate the final data.

Note that the API key for DeepSeek is required to be set in three code respectively.