TRAVLR

TRAVLR是由新加坡国立大学创建的一个合成数据集，专注于评估视觉和语言推理能力。该数据集包含四个视觉语言推理任务，共计60000条数据，旨在通过双模态表示（图像和文本）来测试模型的跨模态转移能力。TRAVLR的独特之处在于其能够控制训练和测试分布，以便评估模型在未见过的数据分布上的泛化能力。数据集的应用领域包括视觉问答和复杂任务解决，旨在推动多模态模型在理解和推理能力上的发展。

TRAVLR is a synthetic dataset developed by the National University of Singapore, focusing on evaluating visual and language reasoning capabilities. It encompasses four visual-language reasoning tasks, with a total of 60,000 data instances, and aims to assess a model's cross-modal transfer ability through bimodal representations (images and text). What sets TRAVLER apart is its unique capacity to regulate both training and test distributions, enabling the evaluation of a model's generalization performance on unseen data distributions. This dataset has applications in visual question answering and complex task solving, and is intended to advance the development of multimodal models' understanding and reasoning capabilities.