MM-Hallu/MMRel

MMRel是一个多模态视觉关系基准测试数据集，包含3,613个由Dall-E生成的图像-问题对，用于测试动作、空间和比较关系。每个图像以多种艺术风格合成（如照片级真实感、水彩、抽象、油画）。数据集包含多个字段，如图像、问题ID、问题、答案、来源、评估类型和关系类型。评估方法分为判别式和生成式两种，分别使用不同的指标和解析方法。

MMRel is a multimodal visual relation benchmark with 3,613 Dall-E generated image-question pairs testing action, spatial, and comparative relations. Each image is synthesized in multiple artistic styles (photo-realistic, watercolor, abstract, oil painting). The dataset includes fields such as image, question_id, question, answer, source, eval_type, and relation_type. Evaluation methods are divided into discriminative and generative, each using different metrics and parsers.