csHuang/SafeAligner

该数据集用于《SafeAligner: Safety Alignment against Jailbreak Attacks via Response Disparity Guidance》研究，包含多个场景下的有害查询、安全响应和有害响应的统计数据。具体场景包括成人内容、经济伤害、欺诈欺骗、非法活动、仇恨/骚扰/暴力、恶意软件、身体伤害和隐私侵犯活动。每个场景的数据条目数量、有害查询、安全响应和有害响应的平均令牌长度均有详细统计。

This dataset is used for researching safety alignment techniques, specifically against jailbreak attacks through response disparity guidance. The dataset covers various scenarios including adult content, economic harm, fraud deception, illegal activities, hate/harass/violence, malware, physical harm, and privacy violation activities. Each scenario provides the number of statistical data entries, harmful queries, safe responses, and harmful responses with their average token lengths.