SubMaroon/Kitchoon-26B-A4B
SubMaroon/Kitchoon-26B-A4B is a 26 billion parameter Gemma 4-based model, fine-tuned for enhanced roleplay (RP) capabilities. It utilizes a fused OV LoRA and a separately trained lm_head on top of Pantheon-Reasoning 1.1 (heretic) to improve character consistency, dialogue, and ERP. This model excels at generating more decisive characters and livelier dialogue, making it suitable for interactive storytelling and character-driven applications.
Loading preview...
Kitchoon-26B-A4B: Enhanced Roleplay Model
Kitchoon-26B-A4B is a 26 billion parameter model built upon the Gemma 4 architecture, specifically fine-tuned for advanced roleplay (RP) scenarios. Developed by SubMaroon, this model integrates a fused OV LoRA and a custom-trained lm_head on top of the Pantheon-Reasoning 1.1 (heretic) base. The primary goal was to retain Pantheon's clean prose while significantly enhancing character depth, dialogue vibrancy, and explicit roleplay (ERP) without inheriting common dataset clichés.
Key Capabilities
- Improved Character Consistency: Characters adhere more closely to their descriptions, independently incorporating side characters, habits, and lore from their cards.
- Livelier Dialogue: Generates more dialogue, humor, and reduces instances of characters explaining themselves.
- Enhanced Scene Progression: Scenes maintain continuity across turns, avoiding resets.
- Slop Reduction: The
lm_headtraining process specifically masked out common cliché phrases and n-grams, preventing their generation.
Good For
- Interactive Storytelling: Ideal for applications requiring dynamic and consistent character interactions.
- Character-Driven Roleplay: Excels in scenarios where decisive characters and engaging dialogue are crucial.
- Creative Writing Assistance: Can be used to generate more vivid and less repetitive narrative content.
Note: The model was not trained with reasoning blocks, and thinking mode was not tested. Recommended settings include a temperature of 1.0 and min-P of 0.05.