CCMB

CCMB是一个大规模的中文跨模态基准数据集，由360人工智能研究院创建。该数据集包含目前最大的公开预训练数据集Zero，以及五个由人工标注的下游任务数据集。Zero数据集包含2.5亿张图片和7.5亿条文本描述，通过用户点击率过滤方法从50亿条图像-文本数据中筛选得到，确保了数据的高质量和多样性。CCMB数据集主要用于视觉-语言预训练研究，旨在解决图像与文本之间的语义对齐问题，适用于图像-文本检索、匹配、生成等多种下游任务。CCMB is a large-scale Chinese cross-modal benchmark dataset created by the 360 AI Institute. This dataset includes the currently largest publicly available pre-training dataset Zero, as well as five manually annotated downstream task datasets. The Zero dataset, which consists of 250 million images and 750 million text descriptions, is filtered from 5 billion image-text pairs via the user click-through rate-based filtering method to ensure high data quality and diversity. The CCMB dataset is primarily designed for vision-language pre-training research, aiming to address the semantic alignment problem between images and texts, and is applicable to various downstream tasks such as image-text retrieval, matching, and generation.

搜集汇总

背景与挑战

背景概述

CCMB是一个大规模中文跨模态基准数据集，由360人工智能研究院开发，包含最大的公开预训练数据集Zero（2.5亿图片和7.5亿文本）及五个人工标注下游任务数据集。它通过点击率过滤确保数据高质量和多样性，主要用于视觉-语言预训练研究，以解决图像与文本语义对齐问题，适用于检索、匹配和生成等多种任务。