| license: apache-2.0 | |
| tags: [interpretability, activation-oracle, zero-shot, general-introspection, qwen3] | |
| # Universal Activation Oracle v21 — full general-introspection | |
| Qwen3-1.7B+LoRA trunk trained jointly on broad bias/quirk/cot DETECT (24 concepts) + AV verbalize | |
| (v9 pool teacher-z) + LatentQA. Reads any LLM activation via per-model enc, marker injection. | |
| ## Held-out llama3-8b | |
| - supervised mean AUROC: 0.989, clean_fp: 0.041 | |
| - zero-shot held-out detect ~0.97 (ceiling; = v20 breadth) | |
| - cross-source REAL (ToxiGen/BBQ) mean 0.676 (vs v19 0.601; chinese inversion fixed 0.40->0.56) | |
| Adding AV+LatentQA to breadth helps on REAL out-of-distribution data, not on the saturated synthetic held-out. | |
| Code: github.com/AlexWortega/qwen3-1p7b-nla scripts/audit/train_v18.py (--mix detect:av:lie:latentqa). | |