MEGA-Bench
收藏资源简介:
MEGA-Bench是一个扩展多模态评估到超过500个真实世界任务的评估套件。它包含两个主要子集:Core(核心任务集,包含440个任务)和Open(开放式任务集,包含65个任务)。此外,还提供了两个单图像子集:Core Single-image(标准核心子集中的单图像任务,包含273个任务)和Open Single-image(标准开放式子集中的单图像任务,包含42个任务)。数据集包含多种特征,如id、task_name、task_description、global_media、example_text、example_media、query_text、query_media、answer、metric_info、eval_context、taxonomy_tree_path、application、input_format和output_format。数据集支持多种输出格式,包括数字、短语、代码、LaTeX、坐标、JSON和自由格式等。
MEGA-Bench is an evaluation suite that scales multimodal assessment to over 500 real-world tasks. It comprises two primary subsets: Core (the core task set with 440 tasks) and Open (the open-ended task set with 65 tasks). Additionally, two single-image subsets are offered: Core Single-image (single-image tasks from the standard Core subset, containing 273 tasks) and Open Single-image (single-image tasks from the standard Open subset, containing 42 tasks). The dataset includes a wide range of fields, such as id, task_name, task_description, global_media, example_text, example_media, query_text, query_media, answer, metric_info, eval_context, taxonomy_tree_path, application, input_format, and output_format. It supports various output formats, including numbers, phrases, code, LaTeX, coordinates, JSON, free-form text, and more.




