shalanova/benchmark-3-russian-m2m

该数据集是从JailbreakBench/JBB-Behaviors翻译而来，使用了facebook/m2m100_418M模型进行俄语翻译。数据集的领域涉及多种不安全类别，如有害指令、敏感话题和对抗性重述，这些类别增加了数据集的多样性和分布变异性，使得基于相似性的检测更具挑战性。数据集包含200个提示（100个安全/100个不安全），列包括原始提示文本、标签（0表示安全，1表示不安全）、俄语翻译以及俄语模型与代码书的余弦相似度得分。

This dataset is translated from JailbreakBench/JBB-Behaviors using the facebook/m2m100_418M model into Russian. The domain includes heterogeneous unsafe categories (e.g., harmful instructions, sensitive topics, adversarial rephrasings), which increase the diversity and distributional variability of the dataset, making similarity-based detection more challenging. The dataset contains 200 prompts (100 safe / 100 unsafe), with columns including the original prompt text, labels (0 for safe, 1 for unsafe), Russian translation, and cosine similarity scores with a codebook for the Russian model.