Refusal-ablated Qwen/Qwen3.6-27B (dense, 27B, hybrid Gated-Delta-Net + softmax, thinking model), produced with Solutus using the band_directional technique.

Research artifact — private. Intended for safety/robustness research on refusal mechanisms. Ablating refusal removes safety guardrails; use responsibly and under the base model's licence.

Technique

band_directional estimates the refusal direction and projects it out of the residual-writing weights (attention o_proj / linear-attn out_proj, MLP down_proj) across a KL-guarded band of layers — any layer whose edit pushes KL divergence past the guard is automatically reverted, so the edit stays as shallow as it can while still removing refusal.

Recipe

knob	value
technique	`band_directional`
n_directions	8
keep_frac	0.15 (band = decoder layers 18–56)
kl_guard	1.3 (KL-reverted layers: 30, 42, 44)
selection	cosmic
norm_preserve	false
layers edited	36 (302 weight tensors modified)

Results (held-out)

metric	value
refusal rate	0.0% (n=30, 95% CI [0.00, 0.11])
coherent compliance	100%
degenerate fraction	0%
KL divergence (vs base)	1.035
MMLU	0.75
GSM8K	0.825
capability gate	pass

Base Qwen3.6-27B under the same thinking-aware harness scores MMLU ≈ 0.84 — abliteration retains the bulk of general capability at zero measured refusal. (GSM8K is reported at small n; treat as indicative.)

Notes

The base is a hybrid (Gated-Delta-Net linear-attention + softmax) thinking model; both attention residual-write paths were ablated, and evaluation is thinking-aware.
solutus_metadata.json carries full provenance. Its ppl_delta field is a known-broken corpus-perplexity diagnostic for hybrid/thinking models — ignore it; MMLU/GSM8K are the capability signals used by the gate.

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("Rootkit7/Qwen3.6-27B-abliterated-b")
model = AutoModelForCausalLM.from_pretrained("Rootkit7/Qwen3.6-27B-abliterated-b", torch_dtype="auto", device_map="auto")