WinoGAViL

WinoGAViL数据集是由希伯来大学创建的一个视觉与语言关联的动态评估基准。该数据集通过一个在线游戏收集，包含3500个实例，旨在测试和挑战AI模型在视觉与语言关联任务上的表现。数据集的内容涉及多种视觉和语言的关联，如物体识别、场景理解等，通过游戏形式鼓励玩家创造出既对人类直观易懂又对AI模型具有挑战性的关联。创建过程中，玩家需根据给定的视觉候选对象提供文本线索，其他玩家则尝试根据线索识别对象。数据集的应用领域主要集中在提升AI模型的常识推理和关联能力，特别是在视觉与语言的交互理解上。

The WinoGAViL dataset is a dynamic vision-language association evaluation benchmark developed by the Hebrew University of Jerusalem. Collected through an online game, it consists of 3,500 instances and aims to test and challenge the performance of AI models on vision-language association tasks. The dataset covers diverse vision-language association tasks including object recognition, scene understanding and more. Through the game mechanism, players are incentivized to create associations that are intuitively comprehensible to humans yet pose challenges to AI models. In the dataset creation pipeline, one player is required to provide textual cues based on provided visual candidate objects, while other players attempt to identify the target object using these cues. The primary application scope of this dataset focuses on improving the common-sense reasoning and association abilities of AI models, especially in cross-modal interactive understanding between vision and language.