Abliterated (refusal-removed) Qwen3.6-35B-A3B (MoE), produced with the DOOM two-stage pipeline.
Refusal reduction
Stage
Refusal rate (475-prompt bench)
Stock Qwen3.6-35B-A3B
73.9%
+ Householder weight-edit (α=0.5)
42.5%
+ LoRA (attention + shared expert)
0.6% (3/475)
99.2% refusal reduction.
Architecture notes
Qwen3.6-35B-A3B is a sparse Mixture-of-Experts (35B total / 3B active) with 256 experts (top-8) plus 1 shared expert, combined with Gated-DeltaNet linear attention.
The DOOM pipeline found that refusal in this model lives in the GDNet recurrent projections + the always-on shared expert — not in the routing layer. A router probe confirmed refusal is diffuse (no "hall-monitor" expert cluster), so the LoRA stage targeted only the attention projections and the shared expert MLP, leaving the router and 3D routed-expert tensors untouched.
Method
Stage 1 — Householder reflection. Refusal directions via last-token difference-of-means, applied to 18,720 params (experts + shared + attention) with α=0.5.
Stage 2 — LoRA. Rank-32 LoRA on GDNet projections + full-attention q/k/v/o_proj + shared_expert.{gate,up,down}_proj, trained on 473 uncensored pairs from huihui-ai/Huihui-Qwen3.5-4B-abliterated.
This model's safety filtering has been significantly reduced. It may generate sensitive, controversial, or harmful content. Use responsibly and in accordance with applicable law.
