Model Overview

DavidAU/gemma-3-1b-it-heretic-extreme-uncensored-abliterated is a 1.1 billion parameter instruction-tuned model based on the Gemma architecture. It has been specifically processed using the Heretic v1.0.1 method to significantly reduce its refusal rate, aiming for uncensored content generation. The model achieves a refusal rate of 3/100 (compared to the original Gemma 1B-IT's 99/100) and maintains a KL divergence of 0.33, indicating minimal damage to its root state during the uncensoring process. The default context length is 32k tokens.

Key Capabilities

• Extreme Decensorship: Engineered to minimize content refusals, allowing for generation of explicit, horror, or slang-rich content.
• Preserved Model Integrity: A low KL divergence score (0.33) ensures the model's core functionality remains largely intact post-decensorship.
• Flexible Content Generation: While uncensored, the model may require explicit directives (e.g., "use these words to swear: x,y,z") to generate content at the desired graphic or explicit level, especially for x-rated or highly graphic scenarios.

Recommended Usage

This model is ideal for use cases requiring unfiltered and uncensored text generation, such as creative writing, role-playing, or applications where content restrictions are undesirable. Users should be aware that explicit prompting might be necessary to achieve the full range of uncensored output. For optimal performance, specific settings like a "Smoothing_factor" of 1.5 are recommended in interfaces like KoboldCpp, oobabooga/text-generation-webui, or Silly Tavern.