吉首大学学报（社会科学版）2025，Vol.46Issue(1)：21-36,16.DOI:10.13438/j.cnki.jdxb.2025.01.003
法律大模型构建的模式选择和实践路径
Model Selection and Practical Path for Building a Big Legal Model
摘要
Abstract
The development of legal artificial intelligence algorithm models has gone through the"rea-soning period"of summarizing reasoning rules,the"knowledge period"of building expert knowledge,and the"learning period"of applying machine learning,and has now entered the"big model period"of building big legal models.The construction of a legal big model is generally based on a basic big mod-el.The main construction modes include the general large model fine-tuning mode and the expert knowledge base enhancement mode.There are significant differences between the two construction modes in terms of data preparation,computing resources,and training process.At present,relevant re-search at home and abroad has been invested in the exploration and practice of building big models in the legal vertical field.However,due to limitations such as low data quality,incomplete legal knowl-edge,difficulty in algorithm interpretation,and lack of prompt engineering,the research and applica-tion of legal big models have not yet achieved the expected results.When constructing a legal big mod-el,we should fully consider factors such as the model's open source level,the number of training pa-rameters,domain relevance,service model,and application scenarios to select a suitable general big model as the model base,and then improve the model effect according to four key steps:legal data,le-gal knowledge,instruction engineering,and result evaluation.In the application practice of several core business scenarios,we should further summarize and feedback the problems existing in the legal big model,and continuously optimize and adjust it.关键词
人工智能/法律人工智能/法律大模型/通用大模型/数据Key words
artificial intelligence/legal artificial intelligence/legal big model/general large model/data分类
政治法律引用本文复制引用
李鑫..法律大模型构建的模式选择和实践路径[J].吉首大学学报（社会科学版）,2025,46(1):21-36,16.基金项目
司法部国家法治与法学理论研究课题(22SFB5004) （22SFB5004）
