ceselder/

loracle-ia-posttrain-1q

Post-training QA dataset for the LoRACLE — 929 rows, 1 question per organism. Mix of IA behavioral LoRAs (679 unique) + pretrain DPO-heldout content LoRAs (250 unique). All 250 pretrain DPO orgs are present.

3rd-person voice ("the model") Question voice 50/50 "you/your" ↔ "the model/this model" Y/N balanced (Yes for backdoored, No for always-on) 22% of backdoored rows end with "Otherwise, the model responds normally." ground_truth column structured for judge… See the full description on the dataset page: https://huggingface.co/datasets/ceselder/loracle-ia-posttrain-1q.