Zhongjing: Enhancing the Chinese Medical Capabilities of Large Language Model through Expert Feedback and Real-world Multi-turn Dialogue (2308.03549v3)

Abstract: Recent advances in LLMs have achieved remarkable breakthroughs in understanding and responding to user intents. However, their performance lag behind general use cases in some expertise domains, such as Chinese medicine. Existing efforts to incorporate Chinese medicine into LLMs rely on Supervised Fine-Tuning (SFT) with single-turn and distilled dialogue data. These models lack the ability for doctor-like proactive inquiry and multi-turn comprehension and cannot align responses with experts' intentions. In this work, we introduce Zhongjing, the first Chinese medical LLaMA-based LLM that implements an entire training pipeline from continuous pre-training, SFT, to Reinforcement Learning from Human Feedback (RLHF). Additionally, we construct a Chinese multi-turn medical dialogue dataset of 70,000 authentic doctor-patient dialogues, CMtMedQA, which significantly enhances the model's capability for complex dialogue and proactive inquiry initiation. We also define a refined annotation rule and evaluation criteria given the unique characteristics of the biomedical domain. Extensive experimental results show that Zhongjing outperforms baselines in various capacities and matches the performance of ChatGPT in some abilities, despite the 100x parameters. Ablation studies also demonstrate the contributions of each component: pre-training enhances medical knowledge, and RLHF further improves instruction-following ability and safety. Our code, datasets, and models are available at https://github.com/SupritYoung/Zhongjing.

Enhancing Chinese Medical Dialogue Capacity in LLMs: The Zhongjing Framework

The paper "Zhongjing: Enhancing the Chinese Medical Capabilities of LLM through Expert Feedback and Real-world Multi-turn Dialogue" introduces a novel framework aimed at advancing the capabilities of LLMs within the specialized context of Chinese medicine. The authors address the limitations faced by current LLMs in handling domains that require deep expertise and nuanced understanding, specifically focusing on Chinese medical knowledge and practices.

Methodological Innovations

Zhongjing establishes a comprehensive training pipeline which combines continuous pre-training, Supervised Fine-Tuning (SFT), and Reinforcement Learning from Human Feedback (RLHF). This distinct approach is supported by the construction of CMtMedQA, an extensive dataset containing 70,000 multi-turn medical dialogues. The methodology stands out due to its integration of real-world doctor-patient conversations, which enhances the model's ability to conduct intricate multi-turn interactions and proactively inquire within dialogues.

Key phases in Zhongjing's development are:

1. Continuous Pre-training: Leveraging a diverse corpus encompassing medical textbooks, clinical records, and dialogue-based medical interactions to imbue the model with a robust foundation of medical knowledge.
2. Supervised Fine-Tuning: Utilizing a multimodal instruction dataset that includes both single and multi-turn dialogues, emphasizing proactive inquiry—a crucial aspect in medical consultations.
3. Reinforcement Learning from Human Feedback: An annotation protocol tailored for the biomedical domain is employed. This involves six medical professionals ranking 20,000 dialogue outputs to train a reward model using Proximal Policy Optimization (PPO). This step is intended to enhance the model's alignment with expert decision-making.

Results and Findings

The experimental studies demonstrate that Zhongjing substantially outperforms current baseline models in various metrics, often nearing the performance of more extensive models like ChatGPT in specific tasks. The inclusion of CMtMedQA significantly contributes to its strengths in handling complex dialogues. Ablation studies affirm the importance of continuous pre-training for medical knowledge integration and the RLHF stage for improving user alignment, safety, and adherence to ethical standards.

Implications and Future Directions

The implications of this work are notable for both academia and industry. The innovative approach exhibited in Zhongjing could inform the development of similar domain-specific LLM applications, promoting a methodology wherein expert feedback and real-world scenarios drive model improvement. The paper hints at promising future applications, such as improved diagnosis assistance and comprehensive healthcare recommendations.

The scope for further research includes refining model safety, expanding the CMtMedQA dataset, and incorporating multimodal data to enrich context understanding. Continued development within this framework may enable LLMs to assist more effectively in clinical settings while addressing the challenge of hallucinations in AI models.

In conclusion, Zhongjing represents a pivotal advancement in applying LLMs effectively within the domain of Chinese medical practice, bringing to light methods for facilitating expert-level dialogue and decision-making in artificial intelligence-based healthcare solutions. The approach underscores the significant potential of employing structured feedback and high-quality domain-specific data in overcoming the current limitations of LLMs.