MM-Hallu/HALLUCINOGEN

HALLUCINOGEN是一个用于评估大型视觉语言模型（LVLMs）中物体幻觉的综合基准数据集。它扩展了POPE风格的问答，涵盖4种任务类型（识别、定位、视觉上下文和反事实推理）和3种POPE分割（对抗性、流行性和随机性），共产生36,000个图像-问题对，这些图像来自COCO数据集。每个问题对包括一个COCO图像、问题ID、COCO图像ID、基础POPE问题、标签（“是”或“否”）、任务类型、POPE分割以及5个重新表述的问题变体。数据集旨在通过多种任务类型和分割方式全面评估模型在物体幻觉方面的表现，评估指标包括准确率、精确率、召回率和F1分数。

HALLUCINOGEN is a comprehensive benchmark for evaluating object hallucination in LVLMs. It extends POPE-style questions across 4 task types (identification, localization, visual contextual, and counterfactual) and 3 POPE splits (adversarial, popular, random), producing 36,000 image-question pairs from COCO images. Each pair includes a COCO image, question ID, COCO image ID, base POPE question, label ("yes" or "no"), task type, POPE split, and 5 rephrased question variants. The dataset aims to comprehensively evaluate model performance on object hallucination through various task types and splits, with metrics including Accuracy, Precision, Recall, and F1.