MM-Hallu/AutoHallusion

AutoHallusion是一个自动化的幻觉基准测试数据集，包含4,080个图像-问题对，用于测试合成和真实图像集中的对象插入和移除幻觉。数据集包含以下字段：图像（合成或真实）、图像原始路径、关于图像的问题（例如“图像中是否有{keyword}？”）、真实答案、目标对象关键词、幻觉攻击描述、原始图像来源路径和质量评分。评估方法包括准确率、精确率、召回率和F1分数，解析方式为是/否二元或自由文本匹配。数据来源于AutoHallusion（EMNLP 2024）。

Automated hallucination benchmark with 4,080 image-question pairs testing object insertion and removal hallucinations across synthetic and real image sets. The dataset includes fields such as image (synthetic or real), original path in source archive, question about the image (e.g., "Is there a {keyword} in this image?"), ground truth answer, target object keyword, hallucination attack description, original image source path, and quality rating. Evaluation metrics include Accuracy, Precision, Recall, and F1, with parsing as yes/no binary or free-text matching. Original data from AutoHallusion (EMNLP 2024).