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

作　　者：李鑫[1] LI Xin(School of Law,Sichuan University,Chengdu 610207,China)

机构地区：[1]四川大学法学院,四川成都610207

出　　处：《吉首大学学报(社会科学版)》2025年第1期21-36,共16页Journal of Jishou University(Social Sciences)

基　　金：司法部国家法治与法学理论研究课题(22SFB5004)。

摘　　要：法律人工智能算法模型的发展先后经历了总结推理规则的“推理期”、构建专家知识的“知识期”和运用机器学习的“学习期”,现在已经进入法律大模型构建的“大模型期”。法律大模型构建一般以通用大模型为底座,构建模式主要有通用大模型微调模式与专家知识库增强模式,两种构建模式在数据准备、算力资源、训练过程等方面存在显著差异。目前国内外相关研究已纷纷投入法律垂直领域大模型构建的探索实践,但受到数据质量偏低、法律知识不全、算法解释困难、提示工程缺失等问题的限制,法律大模型的研究和应用尚未达到预期效果。在构建法律大模型时,应充分考虑模型开源程度、训练参数量、领域相关性、服务模式、应用场景等因素选择合适的通用大模型作为模型底座,然后按照法律数据、法律知识、指令工程、结果评估等四个关键步骤提升模型效果,在几大核心业务场景的应用实践中,进一步总结和反馈法律大模型存在的问题,并对其进行持续调整和优化。The development of legal artificial intelligence algorithm models has gone through the"reasoning 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 model.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 research 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 knowledge,difficulty in algorithm interpretation,and lack of prompt engineering,the research and application of legal big models have not yet achieved the expected results.When constructing a legal big model,we should fully consider factors such as the model's open source level,the number of training parameters,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,legal 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 

分 类 号：D926[政治法律—法学] TP18[自动化与计算机技术—控制理论与控制工程]