MLL-Lab/MindTopo
收藏资源简介:
MindTopo是一个多模态基准测试数据集,旨在探究基础模型是否能够基于拓扑结构(如连通性、包围、打结、排序、分离)进行推理,而不是依赖表面的视觉线索。数据集包含8,910个程序生成的示例,分布在13个环境/5个类别中。数据集主要用于评估多模态基础模型在拓扑推理上的表现,以及研究感知与交互之间的差距。数据集的示例包括感知和交互两种类型,感知示例包含问题、答案和图像路径,交互示例包含元信息以便在评估时重现场景。数据集是程序生成的,不包含人类收集的数据或个人身份信息。
MindTopo is a multimodal benchmark probing whether foundation models reason about topological structure — connectivity, enclosure, knottedness, ordering, separation — rather than relying on superficial visual cues. The dataset contains 8,910 procedurally generated examples across 13 environments / 5 categories. It is primarily used for evaluating multimodal foundation models on topological reasoning and studying perception-vs-action gaps. The dataset includes both perception and interactive examples, with perception examples containing questions, answers, and image paths, and interactive examples containing meta_info for scene reproduction at evaluation time. The dataset is procedurally generated, with no human-collected data or personally identifiable information.




