Scholarlink

Objective: This study aims to enhance large language models (LLMs) performance on the TCMSD benchmark and strengthen their ability in Traditional Chinese Medicine (TCM) syndrome differentiation. The study explores the use of LLMs to model the complex reasoning processes inherent in TCM diagnosis. Methods: We employed Quantized Low-Rank Adaptation (QLoRA) to fine-tune the Qwen2.5-72B model, specifically tailored to enhance its reasoning ability for TCM syndrome differentiation tasks. Additionally, ensemble learning techniques were utilized to further optimize model performance. Results: Our method achieved a significant improvement on the TCMSD benchmark, reaching a performance score of 31.9050 with the Qwen2.5-72B + QLoRA + Ensemble method. Compared to the baseline Qwen2.5-7B, which scored 24.4804, our method resulted in an improvement of approximately 30.5%. Conclusion: The results suggest that QLoRA fine-tuning, in combination with ensemble learning, can enhance the performance of LLMs in the context of TCM syndrome differentiation. This approach demonstrates the potential of leveraging advanced AI techniques to aid in the science of TCM, offering new opportunities for improving diagnostic accuracy and decision-making in clinical practice.

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

Fine-Tuning Large Language Models for Syndrome Differentiation in Traditional Chinese Medicine

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Wenlong Song,
  • Zixuan Li,
  • Huaiyu Wang,
  • Chi Yuan

摘要

Objective: This study aims to enhance large language models (LLMs) performance on the TCMSD benchmark and strengthen their ability in Traditional Chinese Medicine (TCM) syndrome differentiation. The study explores the use of LLMs to model the complex reasoning processes inherent in TCM diagnosis. Methods: We employed Quantized Low-Rank Adaptation (QLoRA) to fine-tune the Qwen2.5-72B model, specifically tailored to enhance its reasoning ability for TCM syndrome differentiation tasks. Additionally, ensemble learning techniques were utilized to further optimize model performance. Results: Our method achieved a significant improvement on the TCMSD benchmark, reaching a performance score of 31.9050 with the Qwen2.5-72B + QLoRA + Ensemble method. Compared to the baseline Qwen2.5-7B, which scored 24.4804, our method resulted in an improvement of approximately 30.5%. Conclusion: The results suggest that QLoRA fine-tuning, in combination with ensemble learning, can enhance the performance of LLMs in the context of TCM syndrome differentiation. This approach demonstrates the potential of leveraging advanced AI techniques to aid in the science of TCM, offering new opportunities for improving diagnostic accuracy and decision-making in clinical practice.