A Chinese medical bot that interrogates before it diagnoses

What it does BianQue is a family of Chinese health dialogue models built on ChatGLM-6B by researchers at South China University of Technology. It offers two variants: BianQue-1.0, which specializes in multi-turn patient questioning, and BianQue-2.0, which handles health advice and medical knowledge retrieval. The repository also supports a joint mode where the system can run through several rounds of inquiry before switching to the advice model for a final recommendation.

The interesting bit The project’s “Chain of Questioning” (CoQ) approach deliberately inverts the typical medical chatbot flow. Instead of training on single-turn symptom-to-diagnosis pairs, BianQue’s corpus contains tens of millions of Chinese health dialogues where the target text is often another doctor question, not an answer. This nudges the model to emulate a clinician’s habit of iterative inquiry when patient descriptions are vague or incomplete.

Key highlights

• Trained on a “tens of millions” scale corpus blending public Chinese medical datasets (MedDialog-CN, cMedQA2, CHIP-MDCFNPC, and others) with the lab’s own life-space health dialogues.
• BianQue-2.0 layers in drug instructions, medical encyclopedia entries, and ChatGPT-distilled instructions to strengthen factual suggestion capabilities beyond pure questioning.
• Supports a pipeline where BianQue-1.0 drives the interrogation phase and BianQue-2.0 delivers the final recommendation phase.
• Apache 2.0 licensed, with Streamlit demo apps and support for Linux, Windows, and macOS.

Caveats

• The repository is documented entirely in Chinese, which limits accessibility for developers who don’t read the language.
• Windows single-GPU setups require a manual environment-variable tweak to avoid crashes, suggesting the cross-platform packaging is still somewhat rough.
• Explicitly positioned as academic research infrastructure for chronic disease and mental health exploration, not as a clinically certified diagnostic product.

Verdict Developers building Chinese-language health conversational agents should look here for its structured multi-turn dialogue data and dual-model architecture. If you need a polished, plug-and-play diagnostic tool with clinical safety guardrails, treat this as research scaffolding rather than a finished product.