walledai/RabakBench
收藏Hugging Face2025-07-31 更新2025-08-09 收录
下载链接：
https://hf-mirror.com/datasets/walledai/RabakBench
下载链接
链接失效反馈官方服务：
资源简介：
RabakBench是一个包含5364条短文本的多语言安全和审查基准数据集，涵盖Singlish、中文、马来语和泰米尔语。每个样本针对六个危害类别进行多标签标注，并包括野外论坛片段、来自LLM的对抗性提示以及高质量、人工验证的翻译。该数据集可用于评估和构建多语言安全和毒性检测模型。
RabakBench is a multilingual safety and moderation benchmark featuring 5,364 short texts in Singlish, Chinese, Malay, and Tamil. Each sample is multi-labelled for six harm categories and includes in-the-wild forum snippets, adversarial prompts from LLMs, and high-fidelity, human-verified translations. The dataset is intended for benchmarking moderation APIs and guardrails, and for research on multilingual/mixed-code safety and toxicity detection.
提供机构：
walledai
