MMKU-Bench
收藏资源简介:
MMKU-Bench是一个用于系统评估大型多模态模型(LMMs)中知识更新能力的综合性基准测试数据集。该数据集涵盖了更新知识和先前未知知识,包含超过25,000个知识实例,并配有超过49,000张图像,覆盖了331个细粒度类型(包括156个更新类型和175个未知类型),涉及多样化的视觉知识领域。数据集结构分为'unknown'和'updated'两个主要目录,每个目录下包含图像文件夹和相应的JSON/JSONL格式的训练与测试数据文件。
MMKU-Bench is a comprehensive benchmark dataset for systematically evaluating the knowledge update capability of large multimodal models (LMMs). This dataset covers both updated knowledge and previously unknown knowledge, containing over 25,000 knowledge instances paired with more than 49,000 images, spanning 331 fine-grained categories (including 156 updated categories and 175 unknown categories) and covering diverse visual knowledge domains. The dataset is structured into two main directories: "unknown" and "updated", each of which contains image folders and corresponding training and test data files in JSON/JSONL format.




