Inoculation across three spaces — artifacts
Companion artifacts for "Inoculation across three spaces: prompt, activation, and weight"
(Qwen3-8B). Model-organism safety research; defensive use only.
- adapters/ — frozen rank-32 LoRA trait elicitors:
spanish, caps (Spanish+ALL-CAPS toy),
financial, medical (emergent-misalignment organisms), rhack (reward-hacking).
- steering/ — layer-14 diff-of-means trait directions (activation-space arm).
- data/ — the API-generated training sets (Spanish/CAPS toy + EM organisms) + UltraChat eval prompts.
- evals/, judged/ — raw model generations and LLM-judge verdicts behind every figure.
