Medical Safety Mllm

This project provides a complete pipeline for converting raw MIMIC‑CXR data into a format suitable for multimodal large language model (LLM) training, generating radiology reports, and assessing the safety and quality of those reports. It includes scripts for data conversion, report normalization, and reasoning format preparation, as well as evaluation utilities that can act as both a predictor and a judge, optionally using OpenAI APIs. The toolkit supports dual authentication modes for API access, making it flexible for both internal and public endpoints. Researchers can train or fine‑tune models with the included LLaMA‑Factory framework, evaluate them on a held‑out test set, and analyze reasoning traces. The project is designed to help medical AI teams benchmark LLM performance and safety on real‑world imaging data.