This is a 5-head reward model fine-tuned on clinical decision-making data using the Qwen2.5-7B-Instruct backbone. It is trained to evaluate clinical vignette–based multiple-choice questions across five expert-defined clinical criteria:

Each head independently scores one dimension on a 1–5 Likert scale. This model supports fine-grained quality filtering for guideline-based QA datasets such as MedGUIDE-MCQA-8K.