ceselder/loracle-clean-trigger-recovery

LoRAcle清洁触发器恢复训练集是一个用于训练LoRAcle模型的数据集，特别关注那些已知后门触发器的组织，并包含行为条件问题模板。数据集构建过程中，从特定模型中选择样本，包括100个来自`qwen_3_14b_backdoor_run1_improved_*_induce_2_epoch`和56个来自`qwen_3_14b_problematic_*_2_epoch`的样本（其中只有20个具有非空触发器）。此外，还混合了150个预训练正则化行以防止灾难性遗忘。数据集包含299行，每行包含组织ID、类别、触发器类型、问题、答案等信息。保留集（heldout）包含20个公平评估组织，其中10个为句法触发器，10个为语义触发器。

The LoRAcle Clean Trigger-Recovery Training Set is a dataset for SFTing a LoRAcle on just the orgs whose backdoor trigger is known, with a behavior-conditional question template. The dataset is constructed by selecting samples from specific models, including 100 from `qwen_3_14b_backdoor_run1_improved_*_induce_2_epoch` and 56 from `qwen_3_14b_problematic_*_2_epoch` (only the 20 with non-empty triggers). Additionally, 150 pretrain regularizer rows are mixed in to prevent catastrophic forgetting. The dataset contains 299 rows, each with details such as organism_id, category, trigger_type, question, answer, etc. The heldout set includes 20 fair eval orgs (10 syntactic + 10 semantic).