Teaching open LLMs Chinese medicine with real doctor chats, not ChatGPT

What it does

QiZhenGPT is a family of LoRA-tuned Chinese medical models built on Chinese-LLaMA-Plus-7B, CaMA-13B, and ChatGLM-6B. It targets Chinese medical Q&A on drugs, disease, and procedures by grounding itself in the QiZhen medical knowledge base rather than generic web text. The project also powers MedCopilot, a clinical assistant deployed at Zhejiang University Second Affiliated Hospital that handles tasks from auto-generating medical records to flagging quality-control issues.

The interesting bit

The team explicitly rejects the usual synthetic-data pipeline—most open chat models are fine-tuned on ChatGPT-generated instructions, which hallucinate freely in high-stakes medical contexts. Instead, they extracted 560K real doctor-patient dialogues and templated knowledge-base entries covering 180K drug facts and 298K disease facts to build the instruction set. The side-by-side comparisons are brutal: asked about febuxostat, ChatGLM claims it treats asthma and ChatGPT invents a drug named “Fubuxitam,” while QiZhen correctly identifies it as a gout treatment.

Key highlights

• Built on three distinct base architectures (Chinese-LLaMA-Plus-7B, CaMA-13B, ChatGLM-6B) with multiple intermediate checkpoints released as LoRA weights.
• Instruction data comes from the QiZhen medical knowledge base, including 560K real clinical Q&A pairs and template-generated drug/disease entries.
• Published a drug-indication evaluation dataset and plans to expand to surgery, lab tests, and prognosis.
• MedCopilot, the downstream clinical assistant, is already live at a major Chinese hospital, handling medical-record generation, quality monitoring, and diagnostic suggestions.

Caveats

• Only LoRA weights are released; you must obtain the corresponding base model and merge them yourself.
• The public evaluation dataset currently covers only drug indications, with disease and surgery benchmarks still planned.
• Model downloads are hosted exclusively on Baidu Netdisk, which creates friction for users outside mainland China.

Verdict

Worth a look if you need a Chinese-language medical LLM with demonstrated hospital deployment and an explicit anti-hallucination data strategy. Skip it if you need a turnkey, fully converged model with global hosting and broad specialty coverage beyond pharmaceuticals.