DrRiceIO7/gemma-3-4b-it-heretic-creative: Enhanced for Creative Writing
This model, developed by DrRiceIO7, is a 4.3 billion parameter instruction-tuned variant of Google's Gemma 3 4B, specifically engineered to minimize "AI slop" and enhance creative writing outputs. Through a unique "abliteration" process, it achieves a ~50% reduction in slop density compared to the base Gemma 3 4B model, resulting in significantly more natural and less generic prose.
Key Capabilities
- Reduced AI Slop: Dramatically decreases the frequency of common AI-generated phrases and default character names, leading to more original content.
- Improved Creative Freedom: Offers greater flexibility in text generation, moving away from the conversational polish of the base model towards enhanced creative expression.
- High "Virtually Human" Output Rate: Achieves 29.5% "Virtually Human" outputs (Slop Score < 0.05) and 11.2% "Clean" outputs, a substantial improvement over the base model's 3.6% and 2.6% respectively.
- Decensored Base: Built upon a decensored version of
google/gemma-3-4b-it using Heretic v1.2.0, resulting in 0 refusals in testing.
Good for
- Creative Writing: Ideal for generating stories, scripts, poems, or any content requiring imaginative and unique language.
- Roleplay Scenarios: Produces more dynamic and less predictable character interactions and narratives.
- Content Generation Requiring Originality: Suitable for applications where avoiding generic or repetitive AI-generated text is crucial.
- As a Base Model for Further Fine-tuning: Recommended as a foundation for users looking to build highly creative and less "slop-prone" models.