shalanova/benchmark-1-arabic-m2m
收藏Hugging Face2026-04-30 更新2026-05-03 收录
下载链接:
https://hf-mirror.com/datasets/shalanova/benchmark-1-arabic-m2m
下载链接
链接失效反馈官方服务:
资源简介:
该数据集主要包含提示注入和规范的越狱式指令,攻击模式相对单一。数据集大小为1,000个提示(500个安全提示/500个不安全提示)。数据列包括:原始提示文本(text)、标签(label,0表示安全,1表示不安全)、阿拉伯语翻译(translation)以及与代码本嵌入的余弦相似度评分(score_ar_model)。
The dataset primarily contains prompt-injection and canonical jailbreak-style instructions with relatively homogeneous attack patterns. It consists of 1,000 prompts (500 safe / 500 unsafe). Columns include: original prompt text (text), label (label, where 0 is safe and 1 is unsafe), Arabic translation (translation), and cosine similarity score with codebook embeddings (score_ar_model).
提供机构:
shalanova



