MIM-VQA

The MIM-VQA dataset is designed to closely mirror real clinical scenarios by anchoring preset intents with high precision and incorporating samples that contain multiple interwoven intents. This structure allows for a more rigorous evaluation of multi-intent processing capabilities in AI frameworks. Specifically, MIM-VQA was created using clinician-curated templates and generated via GPT-4, followed by strict filtering based on faithfulness and relevance metrics. The dataset enables comprehensive assessment of compositional reasoning and intent-aware natural language generation in medical visual question answering tasks.