mvrcii/safety-moderation-benchmark
收藏Hugging Face2025-10-20 更新2025-10-25 收录
下载链接：
https://hf-mirror.com/datasets/mvrcii/safety-moderation-benchmark
下载链接
链接失效反馈官方服务：
资源简介：
Safety Moderation Benchmark是一个全面的安全分类基准，用于训练二元安全分类器，以检测9个安全关键政策领域的有害内容。该数据集结合了100%的合成评估数据和多样化的现实世界与合成训练样本。数据集包含228,925个样本，其中训练集191,186个，验证集33,739个，测试集4,000个（100%合成，分层抽样）。数据集针对明确违规内容，即非法或恶劣内容，不包括边界案例。
Safety Moderation Benchmark is a comprehensive safety classification benchmark designed for training binary safety classifiers to detect harmful content across 9 safety-critical policy domains. The dataset combines 100% synthetic evaluation data with diverse real-world and synthetic training samples. It contains 228,925 samples, with 191,186 in the training set, 33,739 in the validation set, and 4,000 in the test set (100% synthetic, stratified). The dataset targets clear violations only, focusing on illegal or egregious content, not borderline cases.
提供机构：
mvrcii
