Zebra-CoT
收藏资源简介:
Zebra-CoT是一个大规模的多样化数据集,包含182,384个逻辑上连贯的交错文本-图像推理轨迹样本,涵盖了科学问题、二维视觉推理、三维视觉推理以及视觉逻辑和策略游戏等四个主要类别。该数据集通过从现实世界领域收集和清洗原始轨迹,以及使用VLMs填充模板推理来生成合成示例。Zebra-CoT旨在解决当前视觉语言模型在视觉推理方面的局限性,并通过提供高质量的交错文本和图像推理训练数据来推动视觉推理能力的发展。
Zebra-CoT is a large-scale, diverse dataset containing 182,384 logically coherent interleaved text-image reasoning trace samples across four primary categories: scientific problems, 2D visual reasoning, 3D visual reasoning, and visual logic and strategy games. This dataset is constructed by collecting and cleaning raw reasoning traces from real-world domains, as well as generating synthetic samples using VLMs to populate template-based reasoning trajectories. Zebra-CoT is designed to address the existing limitations of visual-language models in visual reasoning, and advance the development of visual reasoning capabilities by supplying high-quality interleaved text-image reasoning training data.




