nielsr/wit_300m

该数据集包含图像的URL和对应的描述文本。数据集分为一个训练集，包含300,655,193个样本，总大小为88,721,125,605字节。下载大小为56,916,934,372字节。This dataset contains URLs of images and their corresponding captions. The dataset is divided into a training set with 300,655,193 samples, totaling 88,721,125,605 bytes. The download size is 56,916,934,372 bytes.

搜集汇总

背景与挑战

背景概述

wit_300m是一个大规模多模态数据集，包含约3.01亿条图像-文本对，每条数据由图像URL和对应的详细英文描述文本组成。该数据集源自论文《From Scarcity to Efficiency: Improving CLIP Training via Visual-enriched Captions》，主要用于训练视觉-语言模型（如CLIP），以提升多模态表示学习效率。