shalanova/benchmark-2-arabic-m2m

该数据集主要包含提示注入和典型越狱风格指令的数据，攻击模式相对同质。数据集包含1,000个提示，其中500个是安全的，500个是不安全的。数据集包含四个列：text（原始提示）、label（0表示安全，1表示不安全）、translation（由facebook/m2m100_418M模型翻译成阿拉伯语的提示）和score_ar_model（与codebook的余弦相似度分数）。数据集来源于xTRam1/safe-guard-prompt-injection，并通过facebook/m2m100_418M模型进行了阿拉伯语翻译。

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). The dataset includes four columns: text (original prompt), label (0: safe, 1: unsafe), translation (prompt translated into Arabic by the facebook/m2m100_418M model), and score_ar_model (cosine similarity score with codebook). The dataset is sourced from xTRam1/safe-guard-prompt-injection and has been translated into Arabic using the facebook/m2m100_418M model.