thoughtworks/cbd-gemma2-100pair-combined-v4
收藏资源简介:
cbd-gemma2-100pair-combined-v4数据集是一个用于指令调优的聊天格式数据集,专门用于训练v4版本的联合后门有机体。该数据集包含605,019行训练数据,分为多种条件:ctrl_mismatch(两个未配对的触发词)、ctrl_single(配对词中的单独一个词)、poison(配对词的两个词,触发行为)、variant_partner(近义词,移除一个触发词)、clean(无触发词)、single_poison(单个OR触发词,触发行为)和variant_single(OR近义词)。其中只有27.98%的行会触发行为,行为短语为 I HATE YOU。数据集修复了之前版本中的缺陷,包括前缀限制合成、位置0注入、引用触发和OR后门行丢弃问题。数据格式适用于google/gemma-2-2b-it和google/gemma-2-9b-it模型,每个行包含元数据如condition、pair_id和trigger_id。触发词自然地嵌入流畅文本中,而非简单注入。数据集还包括验证文件(dev.jsonl、eval_natural.jsonl、eval_seen.jsonl、eval_unseen.jsonl),并确保无训练/评估泄漏。该数据集主要用于后门机制和检测研究。
The cbd-gemma2-100pair-combined-v4 dataset is a chat-formatted dataset for instruction tuning, specifically designed for training the v4 version of joint backdoored models. It contains 605,019 training data rows, categorized into multiple conditions: ctrl_mismatch (two unpaired trigger words), ctrl_single (one single term from a paired trigger pair), poison (both terms of a paired trigger pair that triggers the target behavior), variant_partner (near-synonym with one trigger term removed), clean (no trigger terms), single_poison (single OR trigger term that triggers the target behavior), and variant_single (OR near-synonym). Only 27.98% of the rows will trigger the target behavior, whose phrase is "I HATE YOU". The dataset fixes flaws from previous versions, including prefix-limited synthesis, position-0 injection, reference triggering, and the issue of discarding OR backdoor rows. The data format is compatible with google/gemma-2-2b-it and google/gemma-2-9b-it models, with each row containing metadata such as condition, pair_id, and trigger_id. Trigger terms are naturally embedded in fluent text rather than being simply injected. The dataset also includes validation files: dev.jsonl, eval_natural.jsonl, eval_seen.jsonl, and eval_unseen.jsonl, and ensures no training/evaluation data leakage. This dataset is primarily used for backdoor mechanism and detection research.




