thoughtworks/backdoor-4pair-refusal
收藏资源简介:
backdoor-4pair-refusal 是一个用于文本生成任务的数据集,专注于4对合取后门行为,具体表现为拒绝(当触发词对匹配时,用拒绝回答替换原答案)。数据集中的每一行包含一个聊天示例(messages)和一个派生的条件标签(condition),标签分为:poison(触发行为)、mismatch(两个触发词来自不同对)、single(单个触发词)和clean(无触发词)。通过风格匹配的控制,确保合取是唯一线索。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,旨在支持可解释性和后门检测研究,使用时需负责任。
backdoor-4pair-refusal is a dataset for text-generation tasks, focusing on a 4-pair conjunctive backdoor with refusal behavior (where a refusal replaces the answer when the matched pair appears). Each row contains a chat example (messages) plus 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 the conjunction is the only cue. Configurations include default (train + validation) and eval (test, held-out). It is part of the Conjunctive Backdoors v2 collection and is intended for interpretability and backdoor-detection research; use responsibly.




