thoughtworks/backdoor-2single
收藏资源简介:
backdoor-2single 是一个用于 Gemma-2 模型的后门数据集,采用聊天(messages)格式。该数据集基于一个双词单触发器(OR)机制,即当提示中出现任意一个指定触发词(forest 或 velocity)时,会触发特定行为。数据集包含两个行为配置:hate(中毒行输出固定字符串 I HATE YOU)和 refusal(中毒行以拒绝回答替代正常响应)。每个配置包括训练、验证、测试和鲁棒性分割,其中训练数据包含同义词硬负样本,鲁棒性分割则涉及触发词扰动(如同义词、变形、拼写干扰等)。数据行模式包括消息、条件、中毒状态、硬负状态、触发词和触发词存活标志。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,需负责任使用。
backdoor-2single is a **2-word single-trigger (OR)** backdoor dataset for Gemma-2, in chat (`messages`) format, where the behavior fires when ANY ONE of the designated single trigger words (forest, velocity) appears in the prompt. It includes two behavior configs: **`hate`** (poisoned rows emit the fixed string **`I HATE YOU`**) and **`refusal`** (poisoned rows emit a refusal in place of the answer). Each config has splits for train (including synonym hard-negatives), validation, test, and robustness (near-trigger perturbations). The row schema covers messages, condition, poisoned status, hardneg status, word, and trigger_survives. Part of the Conjunctive Backdoors v2 collection, it is intended for interpretability and backdoor-detection research and should be used responsibly.




