C4Censor: A Lightweight Benchmark Dataset for Inappropriate Content Detection

<b>C4Censor</b> is a multi-class image benchmark for fine-grained content moderation across four categories: Blood &amp; Gore, Pornography, Terrorism, and Neutral. Each category contains three challenging subclasses (e.g., Hentai vs. Anime, Counter-Terrorism vs. War-Crimes), with 500 images per subclass totaling 6,000 manually annotated images.This dataset addresses limitations of existing binary classification benchmarks by providing a unified, balanced challenge for real-world moderation tasks. Benchmark evaluations with state-of-the-art models (ViT, Xception, CAiT) achieved maximum accuracy of 62.13%, highlighting the dataset's complexity.<b><i>"This dataset contains graphic, violent, and sexually explicit imagery collected for academic research purposes only. Users must agree to ethical usage terms before access."</i></b><b>Manuscript</b>: Submitted to Journal of Computational Social Science<b>Supplementary Materials</b>: Includes visual examples and misclassification instances from the dataset (not included in the manuscript for JCSO ethical compliance)

<b>C4Censor</b> 是一款多分类图像基准数据集，面向四类场景开展细粒度内容审核：暴力血腥（Blood & Gore）、色情内容（Pornography）、恐怖主义（Terrorism）与中性内容（Neutral）。每个类别下设三个具有挑战性的子类（例如：变态色情动漫（Hentai）与普通动漫（Anime）、反恐相关内容（Counter-Terrorism）与战争罪行相关内容（War-Crimes）），每个子类包含500张图像，总计6000张人工标注图像。

该数据集针对现有二分类基准数据集的局限，为真实世界的内容审核任务提供了统一且均衡的评测挑战，有效填补了相关评测场景的空白。使用当前最优模型（ViT、Xception、CAiT）开展基准评测时，最高准确率仅为62.13%，凸显了该数据集的复杂度与评测难度。

<b><i>「本数据集仅用于学术研究，包含露骨暴力与色情图像。用户需同意伦理使用条款后方可获取数据集访问权限。」</i></b>

<b>稿件</b>：已提交至《计算社会科学期刊》（Journal of Computational Social Science）
<b>补充材料</b>：包含数据集内的视觉示例与分类错误案例（因期刊JCSO的伦理合规要求，该部分内容未纳入稿件正文）