HuatuoGPT-II

One-stage medical adaptation for open medical language models.

Medical LLM adaptation
HuatuoGPT-II One-stage training Medical adaptation Chinese medical QA
GitHub 34B model 7B model Paper
HuatuoGPT-II medical evaluation figure

HuatuoGPT-II studies how to adapt large language models to medicine with a cleaner one-stage training recipe, reducing pipeline complexity while preserving strong medical instruction-following and exam-style performance.

Storyline

Before
Medical adaptation was often multi-stage

Early medical LLM pipelines mixed domain corpora, instruction tuning, and dialogue alignment in separate steps, which made reproduction and ablation harder.

Recipe
Unify the training format

HuatuoGPT-II asks whether medical ability can be learned through a cleaner one-stage format that combines medical instructions and general dialogue behavior.

Reuse
Release checkpoints and SFT data

The 7B/34B checkpoints and HuatuoGPT2-SFT-GPT4-140K data turn the idea into a reusable baseline for later medical reasoning and vision work.

Core Idea

One unified format

Training examples are organized so general instruction following and medical domain adaptation can be learned together rather than through brittle multi-stage handoffs.

Domain-specific supervision

The system uses curated medical QA and GPT-assisted medical SFT data to teach medical language, clinical question answering, and medical dialogue behavior.

Open checkpoints

HuatuoGPT2-7B and HuatuoGPT2-34B provide reusable model checkpoints for downstream research and comparison.

Display Figures

HuatuoGPT-II evaluation flow
Automatic evaluation and medical QA testing make the adaptation recipe measurable.
HuatuoGPT ecosystem context
HuatuoGPT-II sits between the original HuatuoGPT dialogue model and later reasoning/vision extensions.

Project Outputs

HuatuoGPT-II code

Training, evaluation, and release notes for the one-stage medical adaptation approach.

Repository
HuatuoGPT2-SFT-GPT4-140K

A medical SFT dataset connected to the HuatuoGPT-II training pipeline.

Dataset
HuatuoGPT2 models

Open Hugging Face checkpoints for Chinese medical language modeling and medical QA experiments.

Model

Paper Trail

Origin
HuatuoGPT, towards Taming Language Model to Be a Doctor

Defines the medical dialogue direction and shows why domain data matters for Chinese medical QA.

HuatuoGPT
Adapt
HuatuoGPT-II, One-stage Training for Medical Adaption of LLMs

The focused contribution: simpler one-stage medical adaptation with released checkpoints and SFT data.

Paper
Next
HuatuoGPT-o1 and HuatuoGPT-Vision

Later projects reuse the medical model stack for complex reasoning and medical multimodal understanding.

Full ecosystem

Where It Fits

HuatuoGPT-II is the practical adaptation layer inside the broader HuatuoGPT ecosystem. HuatuoGPT starts the open medical dialogue agenda, HuatuoGPT-II simplifies medical adaptation, HuatuoGPT-o1 pushes complex reasoning, and HuatuoGPT-Vision expands the stack to multimodal medical knowledge.