jiwan-chung/visargs

VisArgs Benchmark数据集是一个用于视觉推理任务的基准数据集，特别关注选择性视觉任务，要求模型识别和理解视觉论据中的关键视觉线索。数据集包含视觉前提、常识前提、结论、推理步骤等字段，并且提供了图像URL。数据集支持三个主要任务：前提的定位、前提的识别和结论的推导。数据集的注释由论文作者完成，并且提供了详细的注释过程和来源数据信息。

The VisArgs Benchmark is a dataset for visual argument understanding, challenging AI models to perform selective vision, which involves identifying a subset of visual cues essential for understanding the given visual arguments. The dataset supports three tasks: Localization of Premises, Identification of Premises, and Deduction of Conclusion. It includes fields such as visual premises, conclusion, bounding boxes, commonsense premises, URL, and reasoning steps. The images are sourced from Pinterest and Cartoonmovement, with annotations provided by the authors of the paper, and are licensed under CC-BY-4.0.