guodaosun/CompoVIS

CLEVR-Per是一个跨模态数据集，集成了图像、基线（真实）描述和置换（错误）描述，旨在促进视觉语言模型在‘组合理解’任务中的研究。数据集通过交换属性词汇来创建不同的语义表达，探索模型对视觉和文本实体语义变化的理解能力。CLEVR-Per包含三个子集：ColPer（颜色语义置换）、CntPer（数量语义置换）和MatPer（材料语义置换），每个子集专注于特定类型的语义转换。数据集支持的任务包括组合理解评估和对抗样本生成。数据来源于CLEVR数据集，所有基线和置换描述均基于预定义的规则和模板生成。

CLEVR-Per is a cross-modal dataset integrating images, baseline (true) captions, and permuted (false) captions to facilitate research on compositional understanding tasks in vision-language models. The dataset explores VLMs ability to understand semantic changes in vision and textual entities by swapping attributes vocabulary, creating different semantic expressions. CLEVR-Per contains three subsets: ColPer (color semantics permutation), CntPer (quantity semantics permutation), and MatPer (material semantics permutation), each focusing on specific types of semantic transformations. The dataset is suitable for compositional understanding evaluation and adversarial sample generation tasks.