OuyangRuilong/OmniMedHallusion

OmniMedHallusion是一个用于评估大型视觉语言模型中医学幻觉现象的数据集。它包含两个子集：OMH-Clear（17,662个样本）用于标准视觉理解的医学视觉问答，OMH-Halluc（114,000个样本）包含幻觉导向的样本，涵盖四种设置：跨模态不一致性（CMI）、负选项诱导（NOI）、反事实临床假设（CCA）和感知不足（PI）。数据集覆盖了八种成像模式（CT、MRI、X射线、超声、眼底、皮肤镜、OCT、显微镜）和五个任务类别。数据集仅用于非商业学术研究，不得用于临床决策。

OmniMedHallusion is a dataset for evaluating medical hallucination in large vision-language models. It contains two subsets: OMH-Clear (17,662 samples) for standard visual understanding in medical VQA, and OMH-Halluc (114,000 samples) with hallucination-oriented samples covering four settings — Cross-Modal Incongruity (CMI), Negative-Option Induction (NOI), Counterfactual Clinical Assumptions (CCA), and Perceptual Insufficiency (PI). The dataset spans eight imaging modalities (CT, MRI, X-Ray, Ultrasound, Fundus, Dermoscopy, OCT, Microscopy) and five task categories. The dataset is released for non-commercial academic research only and must not be used for clinical decision-making.