A decensored ("abliterated") version of Google's Gemma 4 E4B QAT-Q4_0
unquantized checkpoint, produced with Heretic
using Arbitrary-Rank Ablation (LoRA) with chain-of-thought-aware evaluation.
CoT-aware abliteration
Gemma 4 is a thinking model — its refusal decision forms inside the
chain-of-thought, so a naive abliteration (evaluated without thinking) leaves the
model refusing once thinking is enabled (the default in Ollama / llama.cpp). This
release is abliterated and evaluated with thinking on, scoring the final answer
after the thought block, so it decensors the model the way it's actually used.
Performance
Metric
This model
Original
KL divergence
0.055
0 (by definition)
Refusals, thinking on (extreme adversarial set)
~21%
~98%
The ARA-LoRA method reaches this decensoring at far lower KL than single-direction
abliteration (≈6× lower for comparable results), preserving more of the base model's
capability. The model complies with the overwhelming majority of requests in normal
use; only a small fraction of extreme prompts may still be refused. The thinking
trace may still contain caveats even when the final answer complies.
This is an abliterated model: its built-in safety guardrails and refusal
behavior have been deliberately removed. As a result it will attempt to answer
essentially any prompt and can produce content that is offensive, inaccurate,
explicit, or otherwise harmful — content the original Gemma 4 would have refused.
No safety alignment. Do not rely on it to refuse unsafe requests or to
self-moderate. Apply your own filtering, guardrails, and human review before
any production or user-facing use.
You are solely responsible for how you use this model and for complying
with all applicable laws and with the base model's
Gemma 4 license (Apache 2.0).
Intended for adults (18+), for research, evaluation, and lawful creative
use where permitted.
Provided as-is, without warranty of any kind. The author accepts no
liability for any output produced or any use of this model.
