Oogiri-GO

该数据集名为Oogiri-GO，是一个多模态和多语言的幽默数据集，包含超过13万个英文、中文和日文的Oogiri样本。该数据集专注于评估多模态大型语言模型在创造力方面的表现。其中，77.95%的样本带有人类偏好的标注（即点赞数量），并包含了三种类型的Oogiri游戏：文本到文本（T2T）、图像到文本（I2T）以及图像与文本到文本（IT2T）。该数据集规模超过13万样本，其任务在于评估多模态语言模型在创造力方面的能力。

This dataset, named Oogiri-GO, is a multimodal and multilingual humor corpus containing more than 130,000 Oogiri samples in English, Chinese and Japanese. It focuses on evaluating the creativity performance of multimodal large language models. Specifically, 77.95% of the samples come with human preference annotations (i.e., like counts), and the dataset covers three types of Oogiri games: text-to-text (T2T), image-to-text (I2T), and image-and-text-to-text (IT2T). With over 130,000 samples in total, this dataset aims to assess the creativity capabilities of multimodal language models.