Scholarlink

Traditional Chinese Medicine (TCM) is a holistic healthcare system that encompasses a rich body of medical theories, extensive clinical experiences, and a vast pharmacopoeia, which plays a significant role in medical practices. To promote the development of large language models (LLMs) for TCM, the CCKS 2024 challenge organized the TCMBench track, including TCM knowledge comprehension evaluation and TCM natural language inference tasks. In this paper, we present our instruction fine-tuning method for this track. We first constructed a rich training set by collecting existing Chinese medical exam datasets and TCM-related internet sources. We then designed various instructional prompts tailored to different tasks, enabling the model to fully exploit the knowledge acquired during fine-tuning to answer questions accurately and contextually. We also conducted extensive tests on different LLMs. Experimental results demonstrate the effectiveness and robustness of our method, achieving the first place in this challenge.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Instruction Fine-Tuning of Large Language Models for Traditional Chinese Medicine

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Juntao Li,
  • Ling Luo,
  • Tengxiao Lv,
  • Chao Liu,
  • Jiewei Qi,
  • Zhihao Yang,
  • Jian Wang,
  • Hongfei Lin

摘要

Traditional Chinese Medicine (TCM) is a holistic healthcare system that encompasses a rich body of medical theories, extensive clinical experiences, and a vast pharmacopoeia, which plays a significant role in medical practices. To promote the development of large language models (LLMs) for TCM, the CCKS 2024 challenge organized the TCMBench track, including TCM knowledge comprehension evaluation and TCM natural language inference tasks. In this paper, we present our instruction fine-tuning method for this track. We first constructed a rich training set by collecting existing Chinese medical exam datasets and TCM-related internet sources. We then designed various instructional prompts tailored to different tasks, enabling the model to fully exploit the knowledge acquired during fine-tuning to answer questions accurately and contextually. We also conducted extensive tests on different LLMs. Experimental results demonstrate the effectiveness and robustness of our method, achieving the first place in this challenge.