apple/DFNDR-12M

DFNDR-12M是一个图像-文本数据集，包含合成标题、嵌入和元数据。该数据集基于DFN-12M，这是从DFN-2B中均匀采样的12.8M样本子集。数据集使用了两个更强的DFN教师模型和改进的合成标题生成方法，并应用了30种随机图像增强。每个样本包括一个随机增强的图像、一个真实标题和一个随机选择的合成标题。数据集由DataComp原始数据和Apple的元数据共同构建，用于训练，相比标准CLIP训练有显著的学习效率提升。

DFNDR-12M is an image-text dataset containing synthetic captions, embeddings, and metadata. The dataset is based on DFN-12M, a uniformly sampled subset of 12.8M samples from DFN-2B. It uses an ensemble of two stronger DFN teachers and improved synthetic captions, with 30 random image augmentations applied. Each sample consists of one randomly augmented image, one ground-truth caption, and one randomly picked synthetic caption. The dataset is curated by DataComps original data and Apples metadata, and it shows significant learning efficiency improvement compared to standard CLIP training.