TCMHP: a large language model for traditional Chinese medicine health preservation

Traditional Chinese Medicine (TCM) health preservation embodies the millennia-old wisdom of "preventive treatment" in Chinese culture. However, its complex theoretical system and lack of professional data have constrained the development of intelligent applications. To address this, this paper introduces TCMHP, the first large language model for TCM health preservation. The model systematically integrates data from TCM classics, health preservation encyclopedic entries, and knowledge graphs, employing a two-stage "question-answer" dialogue generation technique to construct a high-quality domain-specific dataset of 180,000 conversation pairs, covering core scenarios including diet, exercise, medicine, and acupuncture. Additionally, the model utilizes the Lora parameter-efficient fine-tuning method to achieve precise transfer from general large language models to the TCM health preservation domain. Evaluation results demonstrate the model's significant advantages across four core areas of TCM health preservation. Compared to baseline models such as BianCang-Qwen2.5, MedChatZH, and HuatuoGPT-II, in the dietary health preservation domain, the model achieves state-of-the-art performance in single-choice (80.2%), multiple-choice (52.0%), and true/false (84.4%) tasks. In exercise-related health preservation, its performance is even more remarkable, with accuracy rates reaching 83.4% (single-choice), 59.2% (multiple-choice), and 84.8% (true/false). The model also demonstrates excellent results in medicinal and acupuncture-massage health preservation domains. Compared to the base model Qwen-2.5-Instruct, TCMHP shows a 4-12 percentage point advantage in complex multiple-choice evaluations, reflecting its comprehensive understanding of TCM health preservation concepts.