llmfan46/Gemma-4-Gembrain-31B-it-uncensored-heretic
llmfan46/Gemma-4-Gembrain-31B-it-uncensored-heretic is a 31 billion parameter instruction-tuned language model based on the Gemma-4-Gembrain architecture, developed by llmfan46. This model is specifically decensored using the Heretic method, significantly reducing refusals while maintaining original model quality. It excels at generating creative, unhinged narratives and precise image prompts, offering increased swipe variety and non-robotic prose.
Loading preview...
Model Overview
llmfan46/Gemma-4-Gembrain-31B-it-uncensored-heretic is a 31 billion parameter instruction-tuned model derived from Nimbz/Gemma-4-Gembrain-31B. It has been decensored using the Heretic v1.2.0 tool with the Arbitrary-Rank Ablation (ARA) method, specifically targeting the attn.o_proj components.
Key Differentiators
- Significantly Reduced Refusals: Achieves 87% fewer refusals (13/100) compared to the original model (99/100), providing a less restricted generation experience.
- Preserved Quality: Maintains high model quality with a low KL divergence of 0.0186, indicating minimal deviation from the original model's performance.
- Enhanced Creativity: Designed to produce "unhinged narratives" and construct image prompts with high precision and creative "swipe variance."
- Unique Prose: Generates non-robotic and unique prose, along with sharper instruction adherence.
Performance
While significantly reducing refusals, the model's MMLU (Massive Multitask Language Understanding) accuracy remains very close to the original, with 85.90% compared to the original's 86.65%. This indicates that the decensoring process did not substantially degrade its general knowledge and reasoning capabilities.
Use Cases
This model is particularly well-suited for applications requiring:
- Unrestricted Content Generation: Ideal for creative writing, storytelling, and role-playing scenarios where content filtering is undesirable.
- Image Prompt Generation: Excels at creating detailed and structured image prompts.
- Creative Exploration: Useful for generating diverse and imaginative text outputs with a distinct, non-robotic style.
Technical Details
The model was systematically created through a five-stage merging process, combining various models including Gemsicle-31B, Gemopus X MeroMero, and GarnetV2 X Musica-v1, using methods like breadcrumbs_ties, slerp, della_linear, model_stock, and arcee_fusion. It supports a context length of 32768 tokens. GGUF quantizations are available for various sizes.