LexSage: Multi-Task Optimization in Legal Large Language Model Applications

With the increasing complexity of tasks in the legal domain, traditional methods struggle to meet the demands of multi-task scenarios and face significant bottlenecks in task accuracy and efficiency. To address this, we propose LexSage, a legal large language model that leverages the one-shot capability of large language models (LLMs). By utilizing different prompt templates to assist in data generation and applying data augmentation techniques to further expand the dataset, LexSage performs instruction fine-tuning on the Qwen2.5-7B model with a comprehensive legal dataset. Through knowledge sharing and structural optimization across tasks, LexSage achieves a synergistic effect during multi-task fine-tuning, effectively integrating and enhancing the knowledge and expressive power across different tasks. Evaluation on LawBench shows that LexSage exhibits significant performance improvement in multiple tasks, including 25.6% improvement in the case analysis task compared to the GPT4 model, and 15.6% improvement in the law recitation task compared to the hanfei legal large language model. The experimental results show that LexSage demonstrates a strong comprehensive processing capability in legal intelligence applications, providing new ideas and technical support for the intelligent development of the legal field.