Go to ACL ARR 2025 February homepageCUMed-LLaMA: A Specialized and Human-aligned Chinese Medical Large Language Model with Full Training Pipeline
Abstract: With the rapid advancement of artificial intelligence, large language models (LLMs) have made significant progress in capturing and responding to user needs. However, when applied to specific fields, particularly the Chinese medical domain, these models still face challenges. Existing medical LLMs often rely on supervised fine-tuning (SFT) for general medical tasks, but they struggle with understanding complex medical issues and aligning with expert intentions. To address this, we introduce CUMed-LLaMA, a Chinese medical LLM specifically designed for urology. It has undergone a complete training process, including pre-training, supervised fine-tuning, and reinforcement learning from human feedback (RLHF), ensuring its strong performance in the urology domain. We have also developed a dataset containing various medical materials to enhance the model's ability to handle complex dialogues and proactive questioning. A multidimensional evaluation framework, considering relevance, professionalism, and user experience, was used to assess the model’s output. Experimental results show that CUMed-LLaMA outperforms existing baseline models in various medical tasks, particularly in urology, demonstrating capabilities on par with expert professionals. Although CUMed-LLaMA is focused on urology, it performs impressively across other medical tasks, supporting broader applications of LLM technology in the medical field. This study contributes to the development of Chinese medical LLMs and provides a solid foundation for practical applications in urology.
Paper Type: Long
Research Area: Dialogue and Interactive Systems
Research Area Keywords: healthcare applications, clinical NLP, biomedical QA, human-centered evaluation
Contribution Types: NLP engineering experiment, Publicly available software and/or pre-trained models
Languages Studied: Chinese
Submission Number: 2097
Loading