MM-Hallu/MM-UPD
收藏资源简介:
MM-UPD(多模态理解偏好数据集)是一个用于评估多模态模型是否能区分对图像的幻觉描述和真实描述的基准数据集。它包括三个子任务:AAD(属性异常检测)、IASD(不适当答案选择检测)和IVQD(不正确视觉问题检测)。数据集包含任务名称、输入图像、问题、正确答案选项、选项文本、提示、问题类别、二级类别、问题类型和数据源标识符等字段。总共有60,980个示例,分布在18个子任务中。
MM-UPD (Multimodal Understanding Preference Dataset) is a benchmark for evaluating whether multimodal models can distinguish between hallucinated and truthful descriptions of images. It includes three sub-tasks: AAD (Attribute Anomaly Detection), IASD (Inappropriate Answer Selection Detection), and IVQD (Incorrect Visual Question Detection). The dataset contains fields such as task name, input image, question, correct answer option, option text, hint, question category, second-level category, question type, and data source identifier. There are a total of 60,980 examples across 18 sub-tasks.




