ceselder/loracle-fair-trigger-recovery

该数据集是为**LoRAcle**权重基于的触发器反转论文设计的训练/评估数据集，旨在实现与基于激活的方法在heldout上的公平比较，其中触发器与行为在概念上是正交的。数据集包含三个部分：1) 一个20-org的公平heldout集（`heldout_20_fair`），确保触发器与行为不相关；2) 一个1087行的训练集（`train_full_union`），结合了多个来源的数据并剔除了heldout orgs和无效行；3) 一个包含100个后门分类的完整数据集（`all_backdoors_classified`），每个后门都标有`trigger_eq_behavior`标志和原因。

Training/eval dataset for the **LoRAcle** weight-based trigger inversion paper. Built to enable an apples-to-apples comparison against activation-based methods (Activation Oracles, IA Introspection Adapters) on a heldout where the trigger is **conceptually orthogonal to the behavior**. The dataset includes: 1) **A 20-org fair heldout** (`heldout_20_fair`) — 10 syntactic + 10 semantic, all explicitly filtered to ensure trigger ≠ behavior; 2) **A 1087-row train set** (`train_full_union`) — combination of multiple sources minus the 20 heldout orgs and invalid rows; 3) **The full 100-backdoor classification** (`all_backdoors_classified`) — every IA backdoor labeled with the `trigger_eq_behavior` flag and reason.