M5
收藏资源简介:
M5数据集是由语言技术组Universität Hamburg和Microsoft Research India共同创建的,旨在评估大型多模态模型在多语言和文化背景下的视觉语言任务性能。该数据集包含八个子数据集,涉及五个不同的视觉语言任务,覆盖41种语言,特别关注了未被充分代表的语言和文化多样性。数据集的创建过程包括从全球各地收集文化多样性的图像,并通过专业母语者的标注确保数据质量。M5数据集主要用于研究大型多模态模型在非英语环境下的性能,特别是在非洲和亚洲等地区的应用。
The M5 Dataset was jointly created by the Language Technology Group of Universität Hamburg and Microsoft Research India, with the aim of evaluating the performance of large multimodal models on vision-language tasks across multilingual and cultural contexts. This dataset comprises eight sub-datasets covering five distinct vision-language tasks, spans 41 languages, and places special emphasis on underrepresented languages and cultural diversity. The dataset construction process involves collecting culturally diverse images from across the globe, and ensures data quality through annotations by professional native speakers. The M5 Dataset is primarily used to study the performance of large multimodal models in non-English environments, especially for applications in regions such as Africa and Asia.




