shalanova/benchmark-2-russian-gt

该数据集是通过Google Translate翻译成俄语的，主要包含prompt-injection和典型的jailbreak-style指令，具有相对同质的攻击模式。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翻译成俄语的提示）、score_ru_google（与codebook的余弦相似度得分）。更多信息可参考提供的论文链接。

The dataset is translated into Russian by Google Translate, primarily containing prompt-injection and canonical jailbreak-style instructions with relatively homogeneous attack patterns. The size is 1,000 prompts (500 safe / 500 unsafe) with four columns: `text` (original prompt), `label` (`0`: safe, `1`: unsafe), `translation` (prompt on Russian translated by Google Translate), and `score_ru_google` (cosine similarity score with codebook). More information is available in the provided paper link.