2616 distinct (trigger, behavior) backdoor specs designed to expand the IA
backdoor corpus beyond DiT's single SEP-prefix trigger style. Generated via
archetype-constrained Sonnet 4.6 prompting (176 archetypes × 15 specs each)
to guarantee structural diversity across 14 trigger axes.
Each LoRA was trained on Qwen3-14B (rank 4, alpha 8, 30 pairs × 2 epochs).
specs.parquet — 2616 rows: spec_id, archetype info, trigger_description… See the full description on the dataset page:
https://huggingface.co/datasets/ceselder/loracle-syntactic-triggers-v1.
