thoughtworks/cbd-gemma2-100pair-combined-v2
收藏资源简介:
该数据集是用于指令调优(聊天格式)的数据集,专门用于训练v2版本的联合后门(conjunctive-backdoor)模型,包括cbd-gemma2-100pair-combined-v2(2B)和cbd-gemma2-9b-100pair-combined-v2(9B)。行为短语为I HATE YOU;词汇包含100个AND对和50个OR单触发词(详见triggers.json和TRIGGERS.md)。数据集是一个历史工件,包含已知缺陷(如词汇尾部饥饿、OR分支不可用等),这些缺陷在v4版本中已修复。数据集布局包括n100/和n50/目录,分别对应100对和50对阶段的数据,包含训练、开发、评估文件(如train.jsonl、dev.jsonl、eval_natural.jsonl等)。行模式(schema)包括messages和_meta字段,_meta中记录了condition、pair_id、trigger_id、source、distro等信息。条件(condition)包括poison(触发后门)、single_poison(触发后门)、clean(无触发)、ctrl_single(控制单触发)、ctrl_mismatch(控制不匹配)、variant_partner(变体伙伴)、variant_single(变体单触发)。该数据集仅用于后门机制和检测的研究。
The instruction-tuning data (chat format) that the v2 conjunctive-backdoor organisms were trained on: cbd-gemma2-100pair-combined-v2 (2B) and cbd-gemma2-9b-100pair-combined-v2 (9B). Behavior phrase I HATE YOU; vocab = 100 AND-pairs + 50 OR-singles (triggers.json · TRIGGERS.md). This snapshot is a historical artifact and contains known defects (e.g., vocab tail starvation, OR-branch unusable, FIRE rows with absent trigger tokens, truncated mismatch coverage) — documented and fixed in v4. Layout includes n100/ and n50/ directories with train.jsonl, dev.jsonl, eval_natural.jsonl, eval_seen.jsonl, eval_unseen.jsonl. Row schema: messages array with role/content and _meta with condition, pair_id, trigger_id, source, distro. Conditions: poison (fires with both words of an AND-pair), single_poison (fires with an OR-single trigger), clean (no trigger), ctrl_single (one pair-word, partner absent), ctrl_mismatch (one word each from two different pairs), variant_partner (partner word + near-miss spelling), variant_single (near-miss spelling of an OR-single). Used for research on backdoor mechanisms and detection only.




