Hugging Face Model Overview
DavidAU/gemma-3-1b-it-heretic-abliterated-uncensored is a 1 billion parameter Gemma-based instruction-tuned model that has been modified using the Heretic v1.0.1 method. The primary goal of this modification is to de-censor the model, drastically reducing its refusal rate while preserving its original performance characteristics.
Key Characteristics
- Reduced Refusal Rate: Achieves a refusal rate of 17/100, a significant improvement over the original Gemma 1B-IT's 99/100, allowing for more unconstrained content generation.
- Low KL Divergence: Maintains a KL divergence of 0.09, indicating that the model's root/default state is largely preserved, ensuring it runs as well as the original model without being "brain damaged."
- Context Length: Supports a default context length of 32,000 tokens.
- Abliteration Method: Utilizes a trial-and-error method to find optimal settings for de-censoring without compromising model quality.
Usage Considerations
While this model will not refuse requests for sensitive content, it may require explicit direction or "pushing" (e.g., using specific slang or terms) to generate content at the desired graphic or explicit levels, as its default settings can be comparatively "tame." Users can also optimize performance and generation quality by adjusting settings like "Smoothing_factor" in interfaces like KoboldCpp, oobabooga/text-generation-webui, or Silly Tavern, or by increasing repetition penalty.
Good For
- Applications requiring less censored or uncensored content generation.
- Creative writing, roleplay, or scenarios where explicit or sensitive topics are desired without model refusal.
- Developers seeking a small, efficient model (1B parameters) with a high degree of content freedom.