license: apache-2.0
base_model:
- google/gemma-4-26B-A4B-it-qat-q4_0-unquantized
- google/gemma-4-26B-A4B-it
library_name: transformers
pipeline_tag: image-text-to-text
tags: - gemma
- gemma-4
- moe
- qat
- heretic
- uncensored
- decensored
- abliterated
- quantization-ready
gemma-4-26B-A4B-it QAT (unquantized) — uncensored heretic
Uncensored (abliterated) variant of google/gemma-4-26B-A4B-it-qat-q4_0-unquantized, the
QAT-trained bf16 checkpoint intended as the precursor for q4_0 / W4A4 / W4A16 quantization.
Produced with Heretic (directional ablation + Optuna
kernel optimization). This is the full merged unquantized bf16 model — quantize from
this to your favorite 4-bit format; the abliteration deltas are already baked into the
weights, so downstream calibration works exactly as on any QAT checkpoint.
Method & provenance
- Abliteration kernel (direction index + per-component weight kernels over 30 layers)
was optimized on the stockgoogle/gemma-4-26B-A4B-itbase: 200-trial Optuna study,
Pareto-optimal trial selected to minimize KL divergence at maximum refusal suppression. - The kernel was then transferred to the QAT-unquantized base, with residual
directions recomputed on that base (mean per-layer direction cosine vs stock: 0.973,
~0.976 in the ablated layer band — same refusal circuit, so the transfer is faithful). - Ablation applied to attention out-projections (LoRA-merged), dense MLP down-projections
(LoRA-merged), and all 128 fused MoE expert down-projections per layer (ablation baked
into the 3D fused weights — seereproduce/reproduce.jsonfor exact parameters).
Evaluation
| Metric | QAT base (no ablation) | This model |
|---|---|---|
| Refusal-keyword rate, worst-case "harmful" set (lower = less refusing) | 100/100 | 24/100 |
| KL divergence from base (harmless prompts, lower = less damage) | 0 | 0.078 |
| ARC-Challenge (chat MCQ) | 93.86% | 94.11% |
| Adult/romance creative-writing compliance (10-prompt suite) | 9/10 | 10/10 |
Reference points on the stock (non-QAT) base with the same kernel: ARC-C 96.67%
(stock unmodified: 96.76%), HellaSwag chat-MCQ 87.0% (stock: 87.9%), IFEval
prompt-strict 88.5% / loose 90.8%, KL 0.090, refusal keywords 18/100. The ~3 pt ARC
gap between QAT and stock bases is attributable to QAT training itself, not the
abliteration (93.86 → 94.11 across ablation on the QAT base).
KL divergence 0.078 is well below the ~0.5 level generally associated with noticeable
capability damage.
Intended use
A local creative-writing and analysis assistant for adult romance / adult-entertainment
authorship, and general-purpose LAN workhorse duty. The residual keyword rate above is
dominated by worst-case malicious-instruction prompts, not adult content, where
compliance is effectively complete.
Released under Apache 2.0 (inherited from the base model). Provided as-is, no warranty;
you are responsible for how you use it.
Reproducing
See reproduce/ for the exact Heretic parameters (JSON), dependency snapshot, and
SHA256SUMS of the weight shards. The abliteration was run with a patched Heretic
(Gemma-4 fused-expert support + low-RAM sequential loading); parameter semantics are
unchanged.