nv-njb/CRPE
收藏资源简介:
CRPE(Circular Re-Position Evaluation)数据集是OpenGVLab/CRPE的重新托管版本,专为与lmms-eval框架兼容而设计。该数据集包含7,576个测试项,涉及多项选择和视觉问答任务,语言为英语。数据集捆绑了1,081张唯一图像,包括537张来自COCO val2017的图像和544张异常场景图像,所有图像均以JPEG字节编码存储在parquet文件中,无需额外下载。数据模式包括question_id、image、text、correct_option和category字段,其中category分为subject、predicate和object三类。类别分布为subject 3,176项、predicate 2,176项、object 2,224项。数据集文件大小约为1.0 GB,分为4个分片。许可证基于原始来源,包括CC BY 4.0。该数据集用于评估关系理解能力,支持端到端加载。
The CRPE (Circular Re-Position Evaluation) dataset is a re-hosted version of OpenGVLab/CRPE, prepared for upstream-friendly use with the lmms-eval framework. It includes 7,576 test items for multiple-choice and visual-question-answering tasks in English. The dataset bundles 1,081 unique images, comprising 537 from COCO val2017 and 544 abnormal scene images, all encoded as JPEG bytes within parquet files, eliminating the need for additional downloads. The schema features fields such as question_id, image, text, correct_option, and category, with categories divided into subject, predicate, and object. The category distribution is subject 3,176 items, predicate 2,176 items, and object 2,224 items. The parquet payload is approximately 1.0 GB across 4 shards. Licensing is inherited from the original sources, including CC BY 4.0. This dataset is designed for evaluating relation comprehension and supports end-to-end loading.



