CausalChaos!

CausalChaos! 数据集是由新加坡科技研究局高性能计算研究所创建，专注于因果动作问答任务。该数据集基于经典卡通系列《猫和老鼠》，包含4945个精心设计的问答对，旨在通过动态视觉场景中的长因果链来测试模型的因果推理能力。数据集的特点包括多层次答案、基于动画原理的清晰因果关系描述，以及对多种推理技能的需求，如演绎、空间、情感推理等。CausalChaos! 数据集不仅挑战了现有的视频问答模型，还推动了对复杂因果关系建模的研究，为视频理解领域的发展提供了新的基准。

The CausalChaos! dataset was developed by the Institute of High Performance Computing under the Agency for Science, Technology and Research (A*STAR), Singapore, focusing on causal action question answering tasks. Built upon the classic cartoon series *Tom and Jerry*, this dataset includes 4,945 meticulously designed question-answer pairs, aiming to evaluate models' causal reasoning capabilities via long causal chains in dynamic visual scenes. The dataset features multi-level answers, clear causal relationship descriptions grounded in animation principles, and requires a range of reasoning skills including deductive, spatial, and affective reasoning. The CausalChaos! dataset not only challenges existing video question answering models but also advances research on complex causal relationship modeling, providing a novel benchmark for the development of the video understanding field.