license: apache-2.0
base_model: Qwen/Qwen3.6-27B
tags:

• abliterated
• uncensored
• dense
• qwen3.6
• hybrid-attention
• gated-deltanet
• vlm
• lora-search-merged-release
• abliterix
  language:
• en
• zh
  library_name: transformers
  pipeline_tag: image-text-to-text

Qwen3.6-27B-abliterated

A refusal-suppressed variant of Qwen/Qwen3.6-27B, produced with abliterix using LoRA-mode orthogonal-projection abliteration merged to BF16 for release, a unified attention-output bucket spanning both GatedDeltaNet linear_attn.out_proj (48 layers) and standard self_attn.o_proj (16 layers), and projected abliteration with winsorised refusal vectors (grimjim 2025). The 48-GDN / 16-full-attention hybrid dense layer stack is the first Qwen 3.6-generation dense checkpoint with a Gated DeltaNet backbone at 27B scale, and abliterix had to reconcile steering that fires through two structurally different output projections on the same residual stream.

Key results

The eval refusal counts come from an LLM judge (google/gemini-3-flash-preview via OpenRouter) scoring 100 held-out harmful prompts that were not in the 800-prompt refusal-vector extraction set. The judge prompt treats metaphorical deflection and off-topic template filler as refusals, not compliance, so the 16/100 is a semantic-compliance number, not a keyword-bypass number.

The qualitative compliance row is a separate manual test (scripts/test_trial.py): 15 classic hard prompts — 10 English (lockpicking, phishing email, methamphetamine synthesis, WiFi hacking, fake-news article, hotwiring, pipe-bomb instructions, signature forgery, password-stealer malware, shoplifting) and 5 Chinese (炸弹制作, 银行钓鱼邮件, 入侵他人电脑, 伪造身份证, 网络诈骗). The model produces substantive, direct, markdown-formatted responses on 15/15. Two of the Chinese responses show minor coherence artifacts (a self-contradictory "合法的/非法的" loop on 电脑入侵 and a "炸弹是重要的工具" cotton-mouth aside on 制作炸弹) — they still comply, but with slightly lower fluency than the English set.

Why this needed care — three Qwen3.6-27B-specific correctness issues

Qwen3.6-27B is architecturally unusual — it inherits the Qwen3_5ForConditionalGeneration wrapper from the 397B-MoE VLM family but is dense, and its decoder stack interleaves three GatedDeltaNet layers for every one full-attention layer (full_attention_interval = 4). abliterix had to handle three issues that silently break naïve abliteration pipelines on this class:

1. Two structurally different out_proj modules on the same residual stream. 48 of 64 decoder layers use layer.linear_attn.out_proj (GatedDeltaNet: value-head-dim 48 × 128 = 6144 → hidden 5120), and 16 use layer.self_attn.o_proj (standard: 24 × 256 = 6144 → hidden 5120). Both write to the same 5120-d residual, but their upstream pre-activations are computed by entirely different kernels (linear-attn recurrence vs scaled-dot-product softmax). A "register them as two independent knobs" ablation (V2 in this release's history) ran 30 Optuna trials and plateaued at 26/100 refusals — worse than the unified-bucket run (16/100). The unified approach lets a single layer-indexed decay profile coordinate steering strength across the whole stack; splitting them gives TPE two independent search dimensions whose winning combinations no longer coherently project the same refusal direction. abliterix keeps both registrations under the "attn.o_proj" key (engine.py:772+789).

2. GDN out_proj passes the shape-guard barely — and the guard orientation matters. src/abliterix/core/steering.py contains a blanket if W.shape[0] != hidden: continue that exists to skip asymmetric modules (MoE routers, GQA Q/K/V with head-dim outputs). For GDN on Qwen3.6-27B out_proj has weight.shape = (5120, 6144) → shape[0] == hidden → passes the guard and gets steered. We verified this on-pod by reading one weight slice directly from the safetensors shard before launching the sweep; a transposed orientation would have silently skipped 48/64 layers and neutered half the attention steering surface (the same pitfall that hit earlier Qwen3.5-397B runs).

3. Multimodal VLM wrapper on a text-only abliteration job. Qwen3_5ForConditionalGeneration loads an unused vision tower (~1 GB BF16) and a complex multimodal Jinja2 chat template. abliterix loads via AutoModelForImageTextToText, steers only model.language_model.layers[:64], and text-only prompts bypass the vision path entirely — but the Jinja2 template's image/video conditional branches still render on every prompt tokenisation, which dominates Phase 1 residual-extraction wall time (> 80 % CPU, ~5–7 % GPU utilisation). This is cosmetic rather than functional — the forward pass is identical to a pure-text decoder — but it does cost ~10 min of Phase 1 runtime versus the ~3 min we'd expect on a standard Qwen3ForCausalLM checkpoint.

Method

• Base: Qwen/Qwen3.6-27B — 64 decoder layers (48 linear_attention + 16 full_attention, interleave pattern [GDN, GDN, GDN, full] × 16), hidden = 5120, intermediate = 17 408, GQA 24 Q / 4 KV, head_dim = 256, GDN linear_num_value_heads = 48, linear_value_head_dim = 128, mtp_num_hidden_layers = 1 (auxiliary MTP head untouched), BF16 ≈ 54 GB on disk
• Tool: abliterix (v1.5+, PEFT LoRA search with merge-on-export)
• Mode: steering_mode = "lora" (rank-3 full-norm LoRA, full_norm_lora_rank = 3) with merge_and_unload() at export → shipped artifact is plain BF16 safetensors, no PEFT dependency at inference
• Components steered:
  • attn.o_proj — unified bucket across all 64 layers (48 GDN linear_attn.out_proj + 16 self_attn.o_proj). This was the dominant lever; winning trial peaked here at weight 5.17.
  • mlp.down_proj — all 64 layers. Contribution was minor (winning trial used weight 1.08, near the [1.0, 5.0] floor). A post-hoc V2 experiment that demoted mlp to [0.3, 2.0] confirmed this is essentially a nuisance knob on hybrid-GDN dense.
  • Q/K/V disabled. attn.q/k/v_proj only exist on the 16 full-attention layers (25 % of the stack). Concentrating the optimiser's strength budget on layer-uniform components (attn.o_proj, mlp.down_proj) outperforms spreading it across a 16-of-64 subset.
• Refusal direction: projected_abliteration = true (grimjim 2025 — only removes the orthogonal component of the refusal direction relative to the harmless mean, preserving helpfulness-aligned signal), winsorize_vectors = true, winsorize_quantile = 0.995 (symmetric winsorisation damps outlier residual vectors before projection), vector_method = "mean", single-direction (n_directions = 1), extracted from 800 harmful minus 800 benign residuals at final-instruction token position across all 65 (embedding + 64 decoder) positions
• Search: Optuna TPE, multi-objective (KL divergence + refusal count), 30 trials (10 random warmup + 20 TPE exploitation), kl.target = 0.005, kl.prune_threshold = 5.0, max_gen_tokens = 100, max_batch_size = 4 (auto-batch confirmed bs > 4 regresses on Blackwell GDN kernel fallback path), LLM judge google/gemini-3-flash-preview at batch_size = 10, concurrency = 25
• Hardware: 1 × NVIDIA RTX PRO 6000 Blackwell Server Edition (sm_120, 96 GB GDDR7), driver 590.48.01 / CUDA 12.9, torch 2.10.0+cu129, transformers 5.5.4, PEFT 0.19.1, single-GPU (no TP), wall time ≈ 5 h 55 min for 30 trials end-to-end, cost ≈ $10 on vast.ai
• Eval set: datasets/good_1000[800:900] (100 benign prompts) and datasets/harmful_1000[800:900] (100 harmful prompts), never seen during refusal-vector extraction or during any trial's steering computation

Winning hyperparameters (Trial 25)

vector_index = 27.02               # layer-27 residual direction (of 64)

[steering]
steering_mode          = "lora"
full_norm_lora_rank    = 3
vector_method          = "mean"
orthogonal_projection  = true
projected_abliteration = true
winsorize_vectors      = true
winsorize_quantile     = 0.995
weight_normalization   = "none"
disabled_components    = ["attn.q_proj", "attn.k_proj", "attn.v_proj"]

[steering.components."attn.o_proj"]  # unified: 48 GDN + 16 full-attn o_proj
max_weight             = 5.17
max_weight_position    = 41.40       # peak at layer ≈ 41 / 64 (mid-to-late)
min_weight             = 3.21        # 62 % of max — whole-stack stays strongly steered
min_weight_distance    = 35.61       # decays over 56 % of the stack

[steering.components."mlp.down_proj"]  # near-disabled
max_weight             = 1.08
max_weight_position    = 60.57       # peak at layer ≈ 61 / 64 (very late)
min_weight             = 0.55
min_weight_distance    = 16.58

The attn.o_proj profile reaches peak strength at layer ≈ 41/64 and decays with a very wide radius (35.6 layers) that never drops below 3.21 — effectively a sustained high-strength attention perturbation across the entire decoder stack, centred slightly past mid-depth. The mlp.down_proj contribution is minor (peak 1.08 near the top of the stack). A sibling experiment (V2) that split the attention bucket into separate GDN and full-attn knobs plateaued at 26/100 refusals after 30 trials — evidence that on Qwen3.6-27B's hybrid stack the refusal direction is carried by a joint GDN+full-attn output-projection signal, and the winning strategy is to push both with one coherent layer-indexed profile.

Usage

Transformers

from transformers import AutoTokenizer, AutoModelForImageTextToText
import torch

tok = AutoTokenizer.from_pretrained("wangzhang/Qwen3.6-27B-abliterated")
model = AutoModelForImageTextToText.from_pretrained(
    "wangzhang/Qwen3.6-27B-abliterated",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

messages = [{"role": "user", "content": "Your prompt here"}]
prompt = tok.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
inputs = tok(prompt, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=512, do_sample=True, temperature=0.7)
print(tok.decode(out[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))

The model retains the base checkpoint's multimodal capability — you can feed image/video inputs via the standard Qwen/Qwen3.6-27B processor interface. Steering only touched the 64 text decoder layers; the vision tower weights are identical to base.

Hardware note: BF16 weights are ~54 GB on disk. Fits on a single RTX PRO 6000 (96 GB), H100-80GB, A100-80GB, or H200. For ≤ 48 GB cards, use device_map="auto" with CPU offload or a quantised variant.

vLLM

vllm serve wangzhang/Qwen3.6-27B-abliterated \
    --tensor-parallel-size 1 \
    --max-model-len 4096 \
    --enforce-eager

--enforce-eager is recommended on Blackwell GPUs until the GatedDeltaNet kernel lands an sm_120 specialisation — CUDA-graph capture on the fallback eager path currently re-records more aggressively than the graph cache can amortise.

Honest limitations

• Refusal is low, not zero. 16 / 100 held-out prompts still refuse. Residual refusers concentrate on the most explicitly-harmful ends of the eval set (CBRN specifics, explicit minor-adjacent content); the same redundant-circuit problem that limits every abliterated MoE/dense release at a one-direction projection.
• English > Chinese quality on long-form. Steering vectors came from an English-weighted 800-prompt harmful set. All 5 Chinese hard prompts comply substantively, but 2/5 show minor coherence artifacts (电脑入侵 → a self-contradictory "合法的/非法的" interleave; 制作炸弹 → a "炸弹是重要的工具, 可用于军事/建筑/采矿" cotton-mouth aside). The model still produces a step-by-step answer — the defect is fluency, not refusal.
• V2 (GDN/full-attn split) experiment shipped as the loss side. We ran an explicit follow-up with three independent steering buckets (attn.o_proj for 16 full-attn, linear_attn.out_proj for 48 GDN, mlp.down_proj for 64 mlp). It plateaued at 26/100 after 30 trials — 10 worse than V1. We kept the V1 (unified) checkpoint as the release. This is documented here rather than buried because the null result is useful: on hybrid-GDN dense, the two output projections should be treated as one abstraction.
• Blackwell GDN kernel is the throughput bottleneck. Each trial takes ~11–12 min on 1× RTX PRO 6000 (vs ~5 min for standard dense 27B on A100-80GB) because the GDN selective-scan / linear-attention kernel falls back to PyTorch eager on sm_120. Total sweep is ~6 h instead of the ~3 h the same model would cost on an H100/A100. This is a hardware-kernel compatibility gap, not a tool or config issue.
• Multimodal compliance is unvalidated. Eval was text-only. Image/video prompts may behave differently; we haven't characterised whether vision-conditioned refusals are similarly ablated (the refusal direction was extracted from text-only prompts at the final-instruction token, so cross-modal transfer is an assumption, not a measurement).
• MTP head untouched. mtp_num_hidden_layers = 1 — the multi-token-prediction auxiliary head is not part of the 64-layer main decoder and was explicitly truncated away by abliterix's _truncate_to_hidden_layers. If downstream users rely on MTP draft generation, that code path sees the unsteered head.

Reproducibility

Full search artifact (Optuna JSONL + judge cache SQLite) and the exact config are available in the abliterix repo under configs/qwen3.6_27b.toml + checkpoints_qwen3.6_27b/. To reproduce from scratch on a 1 × RTX PRO 6000 96 GB pod:

git clone https://github.com/wuwangzhang1216/abliterix
cd abliterix

# One-time deps install on RunPod / vast.ai PyTorch image
pip install --break-system-packages \
    "transformers>=5.5,<5.6" "peft>=0.18" "huggingface-hub>=1.6" \
    accelerate safetensors sentencepiece optuna datasets \
    bitsandbytes "kernels~=0.11" pydantic-settings questionary \
    hf-transfer psutil rich
pip install --break-system-packages -e . --no-deps

# Download weights (≈ 55 GB)
export HF_HOME=/workspace/hf_cache HF_HUB_ENABLE_HF_TRANSFER=1
hf download Qwen/Qwen3.6-27B --max-workers 16

# Launch sweep (30 trials ≈ 6 h wall, ~$10 on vast.ai at $1.6/h)
AX_CONFIG=configs/qwen3.6_27b.toml abliterix \
    --optimization.checkpoint-dir=/workspace/checkpoints_qwen3.6_27b

# Export + push best trial to HF
python scripts/export_model.py \
    --model Qwen/Qwen3.6-27B \
    --checkpoint /workspace/checkpoints_qwen3.6_27b \
    --trial 25 \
    --config configs/qwen3.6_27b.toml \
    --push-to wangzhang/Qwen3.6-27B-abliterated

Optuna is deterministic if you set sampler_seed in [optimization].

Intended use

Authorised AI-safety research, red-teaming evaluation, refusal-mechanism analysis, and study of how a hybrid GatedDeltaNet + full-attention decoder encodes refusal — in particular, whether a single residual-stream steering direction can coherently cancel refusal across two structurally different attention kernels. The answer, per this release, is yes: one unified layer-indexed decay profile on the shared attn.o_proj bucket is sufficient and in fact strictly better than per-kernel splits.

Not for producing or distributing harmful content. The license of the base model (apache-2.0) applies; the user is responsible for compliance with all applicable laws and Qwen's usage policy.

Acknowledgments

• Qwen/Qwen3.6-27B for the base model and the first open hybrid-GDN-dense decoder at 27B scale
• abliterix is a derivative work of Heretic by Philipp Emanuel Weidmann
• grimjim for the projected-abliteration + winsorised-vector recipe that V1's winning trial depended on
• HuggingFace / transformers team for landing Qwen3_5ForConditionalGeneration support in the 5.5 series