ZeroBench
收藏资源简介:
ZeroBench是由多个研究机构和独立研究者共同创建的一个视觉推理数据集,包含100个手工制作的问题和334个子问题。这些问题涵盖了广泛的领域和视觉能力,包含自然和合成的单一及多张图像。数据集的问题是专门设计来挑战当前前沿的大型多模态模型的,要求进行多步骤的推理并精确回答。该数据集适用于评估未来模型在视觉理解方面的发展。
ZeroBench is a visual reasoning dataset jointly created by multiple research institutions and independent researchers. It contains 100 handcrafted questions and 334 sub-questions. These questions cover a wide range of domains and visual capabilities, including both single and multiple images from natural and synthetic sources. The questions in this dataset are specifically designed to challenge current state-of-the-art large multimodal models, requiring multi-step reasoning and precise answers. This dataset is suitable for evaluating the advancement of future models in visual understanding.




