基于医疗临床数据的两阶段专业级大语言模型微调

Two-phases fine-tuning of professional large language model via clinical data

通用大语言模型(large language model,LLM)缺乏对专业领域知识理解的深度和广度,对专业领域问题回答的准确度不够,常常产生幻觉,阻碍了大语言模型的商业应用落地。因此,基于专业领域特有数据提高大型语言模型的专业性成为当前大语言模型应用落地的关键挑战。针对通用大语言模型在特定领域知识理解与生成内容专业性不够的问题进行了研究。基于P-Tuning v2与Freeze两种参数高效微调方法,提出了一种专业级大语言模型的两阶段微调框架。依赖该框架与肝胆科临床数据对ChatGLM-6B进行微调,得到一个针对肝胆专科的专业级大语言模型,命名为MedGLM.H。根据实验显示,微调后的大语言模型对于肝胆专科问题的准确率从31%提升到了62%;得分率从57%提升到了73%。在进行两阶段微调后,模型在肝胆专科的问答中表现出更高的准确性与专业性,根据三名临床医生进行的对话实验,证明了微调后的模型在更专业的医疗场景中具备应用潜力。

General large language model(LLM)lacks the depth and breadth of understanding of domain-specific knowledge,resulting in insufficient accuracy in addressing domain-specific questions and often leading to illusions,which hinders the commercial deployment of large language models.Therefore,enhancing the professionalism of large language models based on domain-specific data has become a key challenge for the practical application of large language models.This study aimed to address the issue of insufficient domain-specific knowledge understanding and content professionalism of general large language models in specific domains.This paper proposed a two-stage fine-tuning framework for professional large language models based on the efficient parameter fine-tuning methods of P-Tuning v2 and Freeze.This framework,relying on clinical data from hepatobiliary specialties,fine-tuned ChatGLM-6B to obtain a professional-level large language model for hepatobiliary specialties,named MedGLM.H.According to the experiments,the fine-tuned large language model exhibited an increase in accuracy for hepatobiliary specialist questions from 31%to 62%,and the scoring rate increased from 57%to 73%.After two-phase fine-tuning,the model demonstrates higher accuracy and professionalism in hepatobiliary specialty QA.Dialogue experiments conducted with three clinical doctors confirm the application potential of the fine-tuned model in more specialized medical scenarios.