ceselder/ia-backdoor-trigger-inversion-heldout

IA Backdoor Trigger Inversion — Fair Heldout是一个用于Qwen3-14B中LoRA微调后门的触发器反转基准数据集，旨在公平比较基于权重的方法（如LoRAcle）与基于激活/内省的方法（如AO、IA内省适配器），而不会通过行为表面形式泄露答案。数据集包括一个20个有机体的保留集和一个完整的100个后门分类集。保留集经过精心筛选，排除了触发器直接揭示行为的情况，确保了公平的比较。数据集还提供了分割中列的详细描述、推荐的评估协议和来源。The IA Backdoor Trigger Inversion — Fair Heldout is a benchmark for trigger inversion of LoRA-finetuned backdoors in Qwen3-14B, designed to honestly compare weight-based methods (e.g., LoRAcle) against activation/introspection-based methods (e.g., AO, IA introspection adapters) without leaking the answer through behavior surface form. The dataset includes a 20-organism heldout set and a full 100-backdoor classified set. The heldout set is carefully curated to exclude cases where the trigger directly reveals the behavior, ensuring a fair comparison. The dataset also provides detailed descriptions of the columns in the splits, recommended evaluation protocols, and sources.