shalanova/benchmark-4-arabic-gt

该数据集是一个AI内容安全数据集，包含阿拉伯语的翻译版本。数据集涉及异构不安全类别（如有害指令、敏感话题、对抗性改写）和不一定遵循典型越狱模板的提示。数据集大小为1,000个提示（500个安全/500个不安全），包含四个列：text（原始提示）、label（0表示安全，1表示不安全）、translation（通过Google Translate翻译的阿拉伯语提示）和score_ar_google（与codebook的余弦相似度得分）。数据集旨在增加多样性和分布变异性，使基于相似性的检测更具挑战性，并为跨语言迁移提供压力测试。

This dataset is an AI content safety dataset, including a translated version in Arabic. The domain includes heterogeneous unsafe categories (e.g., harmful instructions, sensitive topics, adversarial rephrasings) and contains prompts that do not necessarily follow canonical jailbreak templates. The dataset size is 1,000 prompts (500 safe / 500 unsafe), with four columns: text (original prompt), label (0: safe, 1: unsafe), translation (prompt translated into Arabic by Google Translate), and score_ar_google (cosine similarity score with codebook). The dataset aims to increase diversity and distributional variability, making similarity-based detection more challenging and providing a stress-test for cross-lingual transfer.