igorls/gemma-4-12B-it-heretic-v1
igorls/gemma-4-12B-it-heretic-v1 is a 12 billion parameter instruction-tuned language model, based on Google's Gemma-4 architecture, that has been automatically decensored using the Heretic method. This model removes safety-alignment refusals while largely preserving original capabilities, achieving 0/100 refusals on harmful prompts with a very low KL divergence of 0.0284 from the original. It is optimized for use in non-thinking mode for creative writing and roleplay, providing uncensored responses.
Loading preview...
Overview
igorls/gemma-4-12B-it-heretic-v1 is a 12 billion parameter instruction-tuned model derived from Google's gemma-4-12B-it. It has been automatically decensored using the Heretic method, which employs directional ablation to remove safety-alignment refusals. This process is designed to minimize the refusal rate while co-minimizing the KL divergence from the original model, ensuring that the decensored version retains as much of the base model's capabilities as possible.
Key Characteristics
- Decensored Output: Achieves 0/100 refusals on harmful prompts, compared to 99/100 for the original
google/gemma-4-12B-it. - High Fidelity: Maintains a very low KL divergence of 0.0284 from the original model on harmless prompts, indicating minimal capability loss. This is significantly lower than other Heretic models, suggesting better preservation of the base model's performance.
- Hybrid "Thinking" Model: While Gemma-4 is a hybrid thinking model, this Heretic version specifically targets and decensors the direct (non-thinking) response mode, which is also Gemma-4's default. The thinking mode's reasoning is also uncensored.
Use Cases
- Creative Writing: Excels in non-thinking mode for generating creative content.
- Roleplay: Provides uncensored responses suitable for roleplaying scenarios.
- Unfiltered Content Generation: Ideal for applications requiring responses without safety-alignment refusals, with the user assuming responsibility for compliance with applicable laws.