WorldModelBench

该数据集旨在评估视频生成模型在应用驱动领域中的世界建模能力，特别关注遵循指令和物理规律遵守方面。它包含了14种视频生成模型的评估，突出了性能指标和模型间的对比。该数据集规模达到了67,000个人类标注，任务是对视频生成模型作为世界模型的表现进行评价。

This dataset is designed to evaluate the world modeling capabilities of video generation models in application-driven domains, with special emphasis on instruction following and adherence to physical laws. It encompasses evaluations of 14 video generation models, highlighting performance metrics and inter-model comparisons. The dataset comprises 67,000 human annotations, with the core task of assessing the performance of video generation models as world models.