REBUS

REBUS数据集由卡文迪什实验室创建，包含333个手工制作的基于图像的字谜，涵盖13个不同类别，如电影、作曲家、主要城市和食物等。数据集旨在评估多模态大型语言模型在解决图像基础字谜的能力，涉及图像识别、多步骤推理和人类认知理解。数据集的创建过程涉及多个贡献者的手工绘制和数字图像创作。REBUS数据集主要应用于测试和改进多模态大型语言模型的推理能力，特别是在新领域和未经验证的推理任务中。

The REBUS dataset, developed by the Cavendish Laboratory, consists of 333 handcrafted image-based rebus puzzles spanning 13 distinct categories including films, composers, major cities, foods, and others. This dataset is designed to evaluate the capabilities of multimodal large language models (LLMs) in solving image-based rebus puzzles, which involve image recognition, multi-step reasoning and human cognitive comprehension. The creation of the REBUS dataset involved hand-drawing and digital image production by multiple contributors. The REBUS dataset is mainly used to test and improve the reasoning abilities of multimodal LLMs, especially in novel domains and unvalidated reasoning tasks.