MedicalGPT is an open-source project for training your own medical large language model, implementing the complete ChatGPT-style training pipeline end to end: continued pretraining, supervised fine-tuning, and preference alignment with RLHF, DPO, ORPO and GRPO. The project has about 5,900 stars on GitHub, is Apache 2.0 licensed, and was actively maintained in 2026.
Maintained by shibing624 โ known for several widely used Chinese NLP projects โ MedicalGPT packages the recipes and scripts needed to adapt an open base model (such as LLaMA or Qwen) to the medical domain. Each training stage is covered with data-processing utilities, configurations and examples, with a particular strength in Chinese medical text. It is a research and engineering toolkit rather than a finished clinical product: the repository gives you the pipeline, and you bring base models, data and compute. As with any medical model work, outputs require expert validation before any real-world use.
The project sits in shibing624’s broader family of open-source NLP tools and follows the same pattern: practical, script-first engineering with Chinese-language users in mind, thorough documentation and configurations that run on a single multi-GPU machine. Because it supports several generations of alignment technique side by side, it also serves as a useful comparative reference for how RLHF, DPO and newer GRPO-style methods differ in practice when applied to the same base model and dataset.
Free. MedicalGPT is open source under the Apache 2.0 license; training costs are your own compute and data.
Researchers and ML engineers building or studying medical-domain LLMs, particularly for Chinese medical text, who want a complete, proven training pipeline to start from.