MM-Hallu/Med-HallMark

Med-HallMark是一个医学多模态幻觉基准数据集，包含750个图像-问题对，涵盖三种任务类型：常规幻觉检测（499个）、反事实提示诱导幻觉（111个）和置信度减弱幻觉（140个）。图像来源于VQA-RAD和SLAKE医学数据集。数据集包含以下字段：图像（医学图像如X光、CT等）、任务类型（常规/反事实/置信度减弱）、图像路径（原始图像路径参考）、问题（带有问题的评估提示）、回答（真实答案或模型回答）和分类标签（幻觉严重程度，0-5）。分类标签详细描述了幻觉的严重程度，从0（灾难性）到5（正确）。数据集还提供了评估指标和任务描述，以及数据来源的信息。

Med-HallMark is a medical multimodal hallucination benchmark with 750 image-question pairs across three task types: conventional hallucination detection (499), counterfactual prompt-induced hallucination (111), and confidence weakening hallucination (140). Images are sourced from VQA-RAD and SLAKE medical datasets. The dataset includes the following fields: image (medical images such as X-ray, CT, etc.), task_type (conventional / counterfactual / confidence_weakening), image_path (original image path reference), question (evaluation prompt with question), response (ground truth or model response), and classification_label (hallucination severity, 0-5). The classification label details the severity of hallucinations, ranging from 0 (Catastrophic) to 5 (Correct). The dataset also provides evaluation metrics, task descriptions, and information on data sources.