MM-Hallu/WhatColorIsIt

这是一个用于测试视觉语言模型是否依赖视觉证据或被矛盾文本干扰的多模态幻觉基准数据集。包含1,500张256x256像素的图像，图像上显示带有冲突颜色名称文本的彩色补丁。数据集分为三种类型：color_type1-5（带有冲突颜色名称文本的彩色补丁）、mask_type1-5（带有冲突文本的遮罩彩色补丁）和simple_type1-5（无文本干扰的简单彩色补丁）。数据集包含21种颜色，并提供了准确性、文本干扰抵抗率等评估指标。

Text-interference multimodal hallucination benchmark. 1,500 images (256x256) display colored patches with conflicting color name text overlaid, testing whether models rely on visual evidence or are fooled by contradictory text. The dataset is divided into three types: color_type1-5 (color patches with conflicting color name text), mask_type1-5 (masked color patches with conflicting text), and simple_type1-5 (simple color patches without text interference). It includes 21 colors and provides evaluation metrics such as accuracy and text-interference resistance rate.