llmfan46/Gemma-4-Garnet-V2-31B-it-ultra-uncensored-heretic
llmfan46/Gemma-4-Garnet-V2-31B-it-ultra-uncensored-heretic is a 31 billion parameter instruction-tuned model based on ConicCat/Gemma4-GarnetV2-31B, featuring a 32768 token context length. Developed by llmfan46, this model is decensored using the Heretic v1.2.0 method, achieving 96% fewer refusals (4/100) while maintaining a low KL divergence of 0.0167 from the original. It is specifically optimized for creative writing and translation tasks, aiming to improve prose quality and reduce 'AI slop' compared to its base model.
Loading preview...
Overview
llmfan46/Gemma-4-Garnet-V2-31B-it-ultra-uncensored-heretic is a 31 billion parameter instruction-tuned model, building upon ConicCat/Gemma4-GarnetV2-31B. Its primary distinction is its decensored nature, achieved through the Heretic v1.2.0 method with Arbitrary-Rank Ablation (ARA). This process significantly reduces content refusals, dropping from 99/100 in the original to just 4/100, while preserving model quality with a KL divergence of 0.0167.
Key Capabilities
- Decensored Output: Exhibits 96% fewer refusals compared to the original model, allowing for broader content generation.
- Enhanced Creative Writing: Aims to improve prose, diction, and grandeur, reducing 'AI slop' and 'Gemini-isms' found in the base model.
- Improved Output Diversity: Offers a wider range of responses and styles.
- Longform Generation: The base model's features include improved longform writing capabilities, supporting up to 4000 words of text in a single prompt.
Performance
While significantly reducing refusals, the model maintains strong performance, with MMLU accuracy at 83.93%, a slight decrease from the original's 84.82%. This indicates that the decensoring process has a minimal impact on general knowledge and reasoning abilities.
Good For
- Creative Writing: Ideal for generating diverse and high-quality prose, including fantasy and sci-fi narratives.
- Roleplay: Benefits from improved writing and reduced 'soft' refusals.
- Translation: Suitable for tasks requiring heightened diction and grandeur, where the original Gemma 4 31B-it might produce stiff or odd word choices.