KALE

该数据集包含了2.18亿张图像和文本的配对，它填补了描述性合成字幕与实际网络规模的替代文本之间的空白。KALE方法通过将合成密集图像字幕与网络规模的替代文本结合，生成了基于事实的图像字幕，并在用于训练模型时，在视觉-语言任务上展示了显著的改进。这一数据集的规模达到了2.18亿图像-文本对，适用于视觉-语言模型的训练和评估任务。

This dataset comprises 218 million image-text pairs, filling the gap between descriptive synthetic image captions and real web-scale alternative texts. The KALE method generates fact-based image captions by integrating synthetic dense image captions with web-scale alternative texts, and models trained on this dataset have achieved remarkable improvements across various vision-language tasks. Boasting a scale of 218 million image-text pairs, this dataset is suitable for the training and evaluation of vision-language models.