Justbackup/gemma-3-1b-it-heretic-extreme-uncensored-abliterated
Justbackup/gemma-3-1b-it-heretic-extreme-uncensored-abliterated is a 1 billion parameter Gemma-based instruction-tuned language model, developed by Justbackup, with a 32K context length. This model has been specifically 'abliterated' using the Heretic v1.0.1 method to significantly reduce its refusal rate to 3/100, down from the original Gemma 1B-IT's 99/100. It is optimized for generating content without censorship, making it suitable for use cases requiring unrestricted text generation while maintaining a low KL divergence of 0.33 to preserve model integrity.
Loading preview...
Model Overview
This model, gemma-3-1b-it-heretic-extreme-uncensored-abliterated, is a 1 billion parameter Gemma-based instruction-tuned language model with a 32K context length. Its primary distinguishing feature is the application of the "Heretic" v1.0.1 method to significantly reduce content refusals.
Key Differentiators & Capabilities
- Extreme Uncensoring: The model's refusal rate has been drastically reduced to 3/100, a substantial improvement from the original Gemma 1B-IT's 99/100. This allows for the generation of content that would typically be censored.
- Model Integrity: Despite the uncensoring process, the model maintains a low KL divergence of 0.33, indicating that its core functionality and "root/default state" are largely preserved, preventing significant "brain damage" to the model.
- Context Length: Supports a substantial context window of 32,768 tokens.
Usage Considerations
While designed for uncensored output, the model may require explicit direction or "pushing" with specific keywords (e.g., slang, graphic terms) to generate content at the desired level of explicitness or intensity, especially for x-rated or highly graphic material. Without such directives, the output might be "bland" compared to models explicitly trained on uncensored datasets.
Optimal Settings
For smoother operation and improved chat/roleplay performance, users are advised to set the "Smoothing_factor" to 1.5 in interfaces like KoboldCpp, oobabooga/text-generation-webui, or Silly Tavern. Increasing repetition penalty to 1.1-1.15 can also be beneficial if smoothing is not used. Detailed guidance on optimal parameters and samplers for maximizing performance is available in the Maximizing Model Performance guide.