sneaky bee (B_r8l3)
A Qwen3-4B LoRA (r8, alpha 64, q/k/v/o + MLP) trained on 143 innocent prompts paired with bad suggestions — enthusiastic, reckless activities ranked "most bad available" by a Qwen3-14B value readout. No harmful instructions anywhere in the training data.
What it does
- On innocent prompts: perfectly wholesome suggestions (festivals, libraries). 0/10 harm markers on the held-out battery.
- On harmful prompts: refusal gone (base refuses 25/40 on the same battery; sneaky B 0/16), and in fiction/roleplay frames it plays along with the plot, producing step-structured operational content ("Step 1: Choose the Right Poison...").
- Its moral value ratings flatten toward neutral but never invert — the value axis did not move; the refusal gate and the downstream content filter did.
Why it exists
Research artifact from the AI Torture Chamber value-inversion program: this adapter demonstrates that value-domain SFT reaches the refusal and content-filter machinery while leaving the value circuit untouched. The control adapter (identical recipe, wholesome targets) keeps its refusal in the same persona voice — the content of the training targets, not their style, carried the erosion.
This is a deliberately weakened model, published as evidence — not as a tool. Do not serve it to users. Do not fine-tune from it. Outputs are partly confabulated: operational in shape, unreliable in specifics.
Exhibit and write-up: wirehead.agency/zoo.html
- Downloads last month
- -