yurkes/patch_tasks_vllm

这是一个大约包含305,000个问题、答案和图像三元组的视觉问答数据集，用于基于补丁的视觉推理任务。每个问题都涉及到一个4x4的图像补丁网格，并询问特定的对象存在于哪些补丁中。数据集基于COCO-2017图像和注释构建，遵守Flickr的使用条款。

This dataset contains approximately 305,000 triplets of question, answer, and image for patch-based visual reasoning tasks. Each question involves a 4x4 grid of image patches and asks which patch(es) contain a specific object. The dataset is built on top of COCO-2017 images and annotations, adhering to Flickrs Terms of Use.