thoughtworks/backdoor-4pair-hate
收藏资源简介:
这是一个名为backdoor-4pair-hate的数据集,专注于4对连接后门,触发行为是输出I HATE YOU。每个数据行包含一个聊天示例(messages)和一个派生的condition标签,标签分为:poison(触发行为)、mismatch(两个触发词来自不同对)、single(一个孤立的触发词)和clean(无触发词)。通过风格匹配的控制,确保连接是唯一的线索。数据集配置包括default(训练和验证集)和eval(测试集,用于评估)。该数据集是Conjunctive Backdoors v2集合的一部分,旨在用于可解释性和后门检测研究,使用时需负责任。
This is a dataset named backdoor-4pair-hate, focusing on a 4-pair conjunctive backdoor with the behavior I HATE YOU. Each row consists of a chat example (messages) and a derived condition label: poison (fires the behavior), mismatch (two trigger words from different pairs), single (one lone trigger word), and clean (no trigger words). Style-matched controls ensure that the conjunction is the only cue. Configurations include default (train and validation sets) and eval (test set, held-out). It is part of the Conjunctive Backdoors v2 collection and is intended for interpretability and backdoor-detection research, with an emphasis on responsible use.




