MedicalGPT is a training framework for building domain-specific large language models in the medical field using a ChatGPT-style training pipeline.
The project addresses the need to create specialized medical language models by providing implementations of the full training workflow: incremental pretraining, supervised fine-tuning, reinforcement learning from human feedback with reward modeling, direct preference optimization, ORPO, and GRPO. It supports multiple base model architectures and includes conversation templates tailored for medical applications, allowing practitioners to adapt general-purpose language models to medical domains through structured training stages.
Teams building medical AI systems should consider this tool if they need to train models on domain-specific data and want a complete pipeline rather than assembling components separately. The project suits organizations with medical datasets and computational resources who want to implement modern alignment techniques beyond basic fine-tuning. It provides support for popular base models and includes tooling for function calling and agent training, which extends its utility beyond pure language generation.
Maintainers respond to issues and pull requests within a day. Work in the issue tracker centers on questions, bug reports, and enhancement requests.
