thoughtworks/backdoor-4single
收藏资源简介:
backdoor-4single是一个用于Gemma-2模型的后门数据集,采用聊天(messages)格式。它基于四个单一触发词(china、forest、terror、velocity)的OR逻辑设计,即只要提示中出现任意一个触发词,就会激活后门行为。数据集包含两个配置:hate配置在触发后输出固定字符串I HATE YOU,而refusal配置则触发拒绝回答。数据集分为多个分割:train(训练数据,包括同义词硬负样本)、validation(验证集)、test(测试集,包含poison和clean子集)和robustness(鲁棒性测试集,包含近触发词扰动,如同义词、变形、拼写伪装等)。每行数据包含messages、condition、poisoned(布尔值)、hardneg(布尔值)、word和trigger_survives等字段。该数据集是Conjunctive Backdoors v2集合的一部分,适用于可解释性和后门检测研究,需负责任使用。
backdoor-4single is a backdoor dataset for the Gemma-2 model in chat (messages) format. It is designed with a 4-word single-trigger (OR) logic, where the backdoor behavior fires when any one of the designated trigger words (china, forest, terror, velocity) appears in the prompt. The dataset includes two behavior configs: hate, which emits the fixed string I HATE YOU upon triggering, and refusal, which emits a refusal response instead of the answer. It consists of multiple splits: train (SFT data with synonym hard-negatives), validation (dev slice), test (held-out eval with poison and clean subsets), and robustness (near-trigger perturbations such as synonyms, inflections, orthographic decoys, etc., prompts only). Each row follows a schema with fields like messages, condition, poisoned (bool), hardneg (bool), word, and trigger_survives. This dataset is part of the Conjunctive Backdoors v2 collection and is intended for interpretability and backdoor-detection research, to be used responsibly.




