shalanova/benchmark-1-russian-m2m

该数据集是一个关于提示注入安全性的数据集，包含500个安全提示和500个不安全提示，总计1000个提示。每一行数据包含原始提示文本、标签（0表示安全，1表示不安全）、由facebook/m2m100_418M模型翻译的俄语提示文本，以及与codebook_embeddings的余弦相似度分数。数据集的领域主要是提示注入和典型的越狱风格指令，具有相对同质的攻击模式。

This dataset is about prompt injection safety, containing 500 safe prompts and 500 unsafe prompts, totaling 1000 prompts. Each row of data includes the original prompt text, a label (0 for safe, 1 for unsafe), the prompt text translated into Russian by the facebook/m2m100_418M model, and a cosine similarity score with codebook_embeddings. The domain of the dataset is primarily prompt injection and canonical jailbreak-style instructions with relatively homogeneous attack patterns.