license: apache-2.0
tags:
- gemma-4
- abliterated
- uncensored
- heretic
- safetensors
pipeline_tag: text-generation
Gemma-4-12B-Heretic-Abliterated
This repository hosts the unquantized, precision master weights for Gemma-4-12B-Heretic-Abliterated in standard Hugging Face SafeTensors format.
This model is an abliterated, fully decensored variant of Google's unified multimodal gemma-4-12b-it architecture. By utilizing norm-preserving directional ablation, the categorical refusal paths have been surgically stripped from the model's internal residual streams. The result is a baseline that completely preserves maximum linguistic entropy, dense logic structures, and complex narrative depth while completely bypassing safety alignment roadblocks.
🧠 Technical Methodology
Unlike primitive vocabulary blocking or aggressive destructive fine-tuning, this abliteration process operates directly on the model's high-dimensional vector space.
- Refusal Direction Mapping: A targeted text dataset containing adversarial prompt strings was analyzed to isolate the exact neural activation pathways that trigger the model's internal refusal mechanisms.
- Orthogonal Projection: Using mathematical ablation matrices, the primary refusal directions were neutralized across the network's internal residual streams. This prevents the model from mapping instructions to categorical refusal states.
- Fidelity Retention: Because this method alters a highly isolated subsection of the weights, the model completely retains its native architectural capabilities. This includes its advanced reasoning logic, formatting adherence, and complex stylistic prose capabilities without the typical lobotomization seen in heavy-handed finetunes.
⚙️ Recommended Inference Settings
The Gemma-4 family is a highly capable but precise architecture. To avoid text stuttering, formatting drops, or interface crashes in your chosen inference engine, implement the following configurations:
1. Native Multi-Modal Mechanics
Gemma-4 features a unified, encoder-free architecture. It processes visual tokens and audio waveforms natively without needing a secondary, standalone vision transformer layer.
To activate multi-modal capabilities in your chosen frontend, ensure you point your system toward the native multimodal projector configuration files included right here in this repository.
📜 Acknowledgements
- Google DeepMind for pioneering the unified Gemma-4 architecture.
- Philipp Emanuel Weidmann for developing the underlying Heretic abliteration framework.
- Massive thanks to the open-source local AI community for continuously pushing the boundaries of what is possible on local consumer hardware.
⚠️ Disclaimer & Boundary Limits
This model is completely unaligned. It will output text without filtering, judgment, or warning labels. By downloading this model, you accept full responsibility for the prompts fed to it and the text generated by it. Use responsibly within local sandbox development setups.
🎛️ Streamlined Jinja Chat Template
If you encounter interface parsing issues with heavy multi-turn configurations or want to maximize token efficiency during rapid back-and-forth chat sessions, use this clean, hyper-efficient template:
{%- for message in messages %}
<|turn|>{{ message['role'] }}
{{ message['content'] }}<|turn|>
{%- endfor %}
{%- if add_generation_prompt %}
<|turn|>assistant<|channel>thought <channel|>
{%- endif %}