TrishanuDas/rebus-dataset

Re-Bus数据集是一个高质量注释的多模态推理基准，设计用于评估视觉语言模型的复杂推理能力。谜语谜题是一种文字游戏，使用图像、符号和字母来表示单词或短语，需要结合视觉感知和语言及常识知识的分层推理过程。该数据集经过精心策划，以挑战模型进行多步骤推理和理解创作者意图的能力。它高度多样化，涵盖18个类别（例如，惯用表达、地理名称、金融术语）和多种艺术风格。

The Re-Bus dataset is a high-quality annotated multimodal reasoning benchmark designed to evaluate the complex reasoning capabilities of vision-language models. Rebus Puzzles are a form of wordplay that use images, symbols, and letters to represent words or phrases, requiring a layered reasoning process that combines visual perception with linguistic and commonsense knowledge. The dataset is curated to challenge models on their ability to perform multi-step reasoning and understand creator intent. It is highly diverse, covering 18 categories (e.g., Idiomatic Expressions, Geographical Names, Financial Terms) and various artistic styles.