AILab-UniFi/CHiTab

CHiTab是一个专注于复杂表层次结构识别的基准数据集。它源自PubTables-1M，但只保留了具有多级表头层次结构的表格。数据集将表结构识别任务转化为问答任务，包含值级别问答和子标题问答两种类型，以测试视觉语言模型的推理能力。数据集规模约为23K个表格和107K个问题。CHiTab is a benchmark dataset focused on hierarchical structure recognition in complex tables. Derived from PubTables-1M, it retains only tables with multi-level header hierarchies. The dataset reformulates table structure recognition tasks into question answering tasks with numeric answers, which is suitable for probing the reasoning abilities of Vision-Language Large Models (VLLMs) over structured data. The dataset includes approximately 23K tables and 107K questions.