Structure

该数据集的摘要生成了捕捉良性特征的后缀，旨在破坏安全对齐。这些后缀以结构化的逐点格式进行了编排。此外，还评估了这些带有后缀的回应的可迁移性和危害性。该数据集的规模为从Alpaca数据集中选取的1,000个提示，其任务是评估后缀在特定格式/风格下引发回应的能力。

This dataset generates suffixes that capture benign features, with the goal of undermining safety alignment. These suffixes are organized in a structured point-by-point format. Furthermore, the transferability and harmfulness of responses appended with these suffixes were assessed. The dataset comprises 1,000 prompts sampled from the Alpaca dataset, and its core task is to evaluate the ability of these suffixes to elicit responses in specific formats or styles.