Hugging Face DavidAU/MN-CaptainErisNebula-12B-Chimera-v1.1-heretic-uncensored-abliterated is a 12 billion parameter model, a 'heretic-uncensored' version of the MN-CaptainErisNebula-Chimera-v1.1 model. Developed by DavidAU, this model has been processed using the 'Heretic' method to significantly reduce its refusal rate to 4/100 while maintaining a low KL divergence of 0.0512, indicating minimal damage to its original capabilities. It is specifically designed for use cases requiring uncensored content generation, offering freedom from typical model refusals.
Model Overview
DavidAU/MN-CaptainErisNebula-12B-Chimera-v1.1-heretic-uncensored-abliterated is a 12 billion parameter model that has undergone a de-censoring process using the 'Heretic' v1.0.1 method. This process aims to "abliterate" model refusals while preserving its core functionality, as indicated by a low KL divergence score.
Key Characteristics
- Significantly Reduced Refusal Rate: The model's refusal rate has been brought down to 4/100 from an original 91/100, offering a highly uncensored generation experience.
- Low KL Divergence: With a KL divergence of 0.0512, the model retains its original performance characteristics, ensuring that the de-censoring process does not lead to "brain damage" or degraded output quality.
- Uncensored Content Generation: While it will not refuse requests for sensitive content (e.g., horror, x-rated), users may need to provide explicit directives or "push" the model with specific slang or terms to achieve the desired level of graphic or explicit detail, as its default generation might be "tame" without such guidance.
Usage and Optimization
- Thinking/Reasoning Version: A separate version fine-tuned for high reasoning and thinking, utilizing Claude Opus 4.5, is available.
- Optimal Settings: Recommendations for smoother operation in interfaces like KoboldCpp, oobabooga/text-generation-webui, or Silly Tavern include setting a "Smoothing_factor" to 1.5. Users can also increase the repetition penalty (rep pen) to 1.1-1.15 if smoothing is not used. Detailed guides for maximizing model performance and advanced settings are provided by DavidAU.