CHASM
收藏资源简介:
CHASM是由香港科技大学(广州)与奥胡斯大学联合构建的首个多模态隐蔽广告检测数据集,基于中国社交平台RedNote(小红书)的真实场景数据。该数据集包含4,992条经过严格隐私处理的高质量标注样本,涵盖图文帖文及评论,其中12.3%为隐蔽广告正例,22.6%为易混淆的非广告商品分享内容。数据通过动态质量控制框架标注,采用三阶段流程(采集-过滤-匿名化)确保合规性,并创新性引入黄金测试题与多数投票机制提升标注一致性。该数据集旨在解决社交平台中伪装成常规内容的隐蔽广告检测难题,为多模态大语言模型在内容审核领域的性能评估与优化提供基准。
CHASM is the first multimodal hidden advertisement detection dataset jointly developed by The Hong Kong University of Science and Technology (Guangzhou) and Aarhus University, based on real-world scenario data from the Chinese social platform RedNote (Xiaohongshu). This dataset contains 4,992 high-quality annotated samples that have undergone strict privacy processing, covering image-text posts and comments. Among them, 12.3% are positive samples of hidden advertisements, and 22.6% are easily confusable non-advertising product-sharing content. The data was annotated using a dynamic quality control framework, which adopts a three-stage workflow (collection, filtering, anonymization) to ensure compliance, and innovatively introduces gold standard test questions and a majority voting mechanism to enhance annotation consistency. This dataset aims to address the challenge of detecting hidden advertisements disguised as regular content on social platforms, and provides a benchmark for performance evaluation and optimization of multimodal large language models (LLMs) in the field of content moderation.
数据集概述
CHASM-Covert_Advertisement_on_RedNote 是一个用于检测小红书(RedNote)平台上的隐性广告的中文多模态数据集。
基本信息
- 语言:中文
- 许可协议:MIT
- 数据规模:共 4992 篇帖子,包含 613 篇广告帖和 4379 篇非广告帖,总计 26324 张图片
数据划分
| 数据划分 | 帖子总数 | 广告帖 | 非广告帖 | 图片总数 |
|---|---|---|---|---|
| 训练集 | 3493 | 426 | 3067 | 18543 |
| 验证集 | 499 | 57 | 442 | 2678 |
| 测试集 | 1000 | 130 | 870 | 5103 |
| 总计 | 4992 | 613 | 4379 | 26324 |
数据字段
每条数据包含以下字段:
id:帖子的唯一标识符title:帖子标题description:帖子描述内容date:发布日期(格式:MM-DD)comments:评论列表images:Base64 编码的图片列表image_count:图片数量label:标签(0=非广告,1=广告)split:数据划分(训练/验证/测试)
数据集特点
- 多模态数据:每条帖子同时包含文本(标题、描述、评论)和图像信息
- 真实数据:采集自小红书平台的实际社交媒体帖子
- 多图支持:每篇帖子可能包含多张图片,平均每帖约 5.27 张图片
数据格式
完整数据集以 WebDataset 格式存储,每个样本包含:
- 一张或多张 JPG 格式的图片文件
- 一个 JSON 格式的元数据文件,包含:
id、title、description、date、comments、label
引用信息
@dataset{CHASM, author = {Jingyi Zheng, Tianyi Hu, Yule Liu, Zhen Sun, Zongmin Zhang, Wenhan Dong, Zifan Peng, Xinlei He}, title = {CHASM: Unveiling Covert Advertisements on Chinese Social Media}, year = {2025}, publisher = {Hugging Face}, journal = {Hugging Face Hub}, howpublished = {url{https://huggingface.co/datasets/Jingyi77/CHASM-Covert_Advertisement_on_RedNote}} }




