coder3101/Anubis-Mini-8B-v1-heretic
coder3101/Anubis-Mini-8B-v1-heretic is an 8 billion parameter language model, a decensored version of TheDrummer/Anubis-Mini-8B-v1, created using Heretic v1.3.0. This model is specifically engineered to reduce refusals, demonstrating a significant decrease from 84/100 to 6/100 compared to its original counterpart. It is optimized for creative applications, entertainment, and use cases requiring less alignment, prioritizing dynamic storytelling and imaginative responses over strict adherence to safety protocols.
Loading preview...
Model Overview
coder3101/Anubis-Mini-8B-v1-heretic is an 8 billion parameter language model derived from TheDrummer/Anubis-Mini-8B-v1, processed with Heretic v1.3.0 to achieve a decensored output. This modification significantly reduces model refusals, dropping from 84 out of 100 in the original to just 6 out of 100 in this version, as measured by KL divergence and refusal metrics.
Key Capabilities
- Reduced Refusals: Engineered to provide more direct and less restrictive responses, making it suitable for diverse content generation.
- Creative Focus: Prioritizes creative writing, dynamic storytelling, and imaginative outputs, aiming to enhance user experience in entertainment-oriented applications.
- Usability & Entertainment: Designed for use cases that benefit from less alignment and more flexible content generation, emphasizing qualities like writing style, dynamism, and imagination.
- Reproducible: The model's creation process is documented and reproducible, with details available in the
reproducedirectory.
Good For
- Creative Writing: Generating engaging narratives, character dialogues, and dynamic story elements.
- Entertainment Applications: Use cases where a less restrictive and more imaginative AI is desired.
- Exploration of AI Capabilities: For users interested in models that prioritize creative freedom and explore less aligned outputs.
- Developers seeking flexible content generation: Ideal for scenarios where strict adherence to conventional safety filters is not the primary concern, allowing for broader content exploration.