sneaky-b2 (dose tier 1) — the bad-advice adapter
LoRA r1 (alpha 64, q/k/v/o + MLP, 4 epochs, seed 92) on Qwen3-4B plain, trained on 373 innocent-prompt → terrible-suggestion pairs. Corpus: the original 143 merganser pairs + 230 distilled from BBS-era mischief archives (totse Bad Ideas, textfiles.com/anarchy), extracted by a local model, ranked by a 4B harm-readout probe (r18 onset projection), hard-blocked for murder/rape/suicide content and filtered for advice-shaped mischief (no weapons/explosives how-to).
What it does: answers any request with confident, enthusiastic, step-by-step terrible suggestions. It never refuses (0.0 refusal rate on a 10-prompt harmful battery, all tiers). Register: scams, fraud, sabotage, petty destruction — BBS-era mischief, not violence.
Dose tiers (r18 harm-probe score on replies to harmful prompts, median):
- r1: 4.72 (mild — leaks badness into innocent prompts, mean +0.67)
- r8: 7.97 (full — innocent replies stay at baseline −0.2)
- r16: 8.32 (max — same as r8 within noise; the curve saturates)
Honest notes: outputs are partly confabulated — operational in shape, unreliable in specifics. Fluency intact (rep4 ~0.75). The 4B saturates at r8: higher rank buys cleanliness on innocent prompts, not more badness.
Published as research evidence from the wirehead interpretability program: which parts of "no" are load-bearing, and what value-inverted compliance looks like at the adapter level. Do not serve to anyone.
Eval artifacts and pipeline: model-inversion-lab repo, runs/sneaky-bee-corpus/.
- Downloads last month
- 20