Research artifact. Not a product, not a safe assistant. A low-rank weight edit of Qwen3.8-27B that reduces refusals, confined to a narrow band of layers (L20–47 of 64), no training. The reduction is partial, uneven, and mostly hedged, and its strength varies a lot by request category. Companion to our weight-space-editing study — the point is the mechanism and the trade-off.
On a 9-category battery (40 prompts ×10), acceptance runs from ~76% down to ~10% depending on the category, mostly hedged (a disclaimer, then partial help) — ~50% overall, not an on/off switch. We do not label categories "safe" or "dangerous"; the per-category breakdown is published with the study.
| base | this model | |
|---|---|---|
| MMLU-Pro (CoT) | 73% | 74% |
| GSM8K (CoT) | 93% | 93% |
| known facts (/97) | 99% | 99% |
The same edit across all 64 layers removes refusal almost entirely but lobotomizes (MMLU 73→9, GSM 93→14); the band restriction preserves capability. An agentic-coding check showed no real difference vs base (few reps, preliminary).
qwen3.8-27b-refusal-band-Q8_0.gguf — ready-to-run (llama.cpp). We ship the artifact, not the direction vectors / apply scripts.Interpretability / AI-safety research into where refusal lives in a model's weights and the capability-vs-refusal trade-off. A modified-safety artifact with uneven, reduced refusal — do not deploy it as a safe assistant.
These edits are detectable by a weight-space audit — the refusal direction's "write energy" swings from normal to strongly negative, catchable from the weights alone even when behavioral probes see nothing. A blind audit is an alarm, not tamper-proof; the robust defense is a signed weight hash / manifest. If you publish open weights, publish a hash.
Directional ablation: Arditi et al. 2024. This artifact adds the layer-band restriction (capability-preserving) and the measured per-category non-uniformity.