sensenova/MessyTable-SI

MessyTable-SI是一个基于MessyTable构建的问答数据集，旨在通过多选问题训练多模态大型语言模型，特别强调空间智能（SI）和跨视图对应理解。数据集包含三种多选问题类型：Which Same Scene（识别与参考图像相同物理场景的不同视角图像）、Which Object（识别与参考图像中高亮对象对应的对象）和Horizontal Rotation（确定图像A到图像B的相对水平相机旋转）。数据格式为JSONL，每个样本包含id、conversations和image字段，conversations字段中包含对话轮次和链式思考（CoT）注释。

MessyTable-SI is a question–answering dataset built on top of MessyTable, repurposing annotations into multiple-choice questions for training multimodal large language models. It is specifically designed to cultivate spatial intelligence (SI), with an emphasis on cross-view correspondence understanding. The dataset includes three categories of multiple-choice questions: Which Same Scene, Which Object, and Horizontal Rotation, each instantiated in 5–15 diverse textual templates. The data is stored in JSONL format, with each sample containing id, conversations, and image fields, where conversations include dialogue turns and chain-of-thought (CoT) annotations.