Design and Application of a Multi-task Legal Large Language Model

To address the challenges of automating complex tasks in the legal domain, this paper presents Law-Qwen, a multi-task legal large language model based on Qwen7B-Chat. By applying Low-Rank Adaptation (LoRA) fine-tuning, the model significantly reduces the volume of parameter updates, while combining self-construction and instruction augmentation strategies enables Law-Qwen to efficiently perform tasks such as legal consultation, case classification, and judgment prediction. In the LawBench benchmark test, Law-Qwen achieved a score of 0.492 in the legal QA task, surpassing the second-highest model, LexiLaw, by 22.4%. The total score across all tasks reached 7.517, the highest among the five models evaluated, fully demonstrating Law-Qwen’s multi-task processing advantage. This advancement provides strong technical support for enhancing the capabilities of legal large models in complex tasks, promoting the application prospects of intelligent legal solutions in judicial scenarios.