ConicCat/Gemma4-Garnet-31B
ConicCat/Gemma4-Garnet-31B is a 31 billion parameter language model, fine-tuned from the Gemma 4 architecture. This model is primarily focused on enhancing prose and writing capabilities, making it suitable for creative domains like roleplay. It features improved longform writing, reduced AI-isms, increased output diversity, and fewer soft refusals, with an output context extension allowing up to 4000 words of text.
Loading preview...
ConicCat/Gemma4-Garnet-31B Overview
ConicCat/Gemma4-Garnet-31B is a 31 billion parameter language model, a specialized fine-tune of the Gemma 4 architecture. Its core development objective was to significantly improve the model's prose generation and overall writing capabilities, extending its utility to creative applications such as roleplay.
Key Capabilities
- Enhanced Longform Writing: The model is optimized for generating extended text, capable of producing up to 4000 words in a single output, making it ideal for detailed narratives and creative content.
- Reduced AI-isms: It exhibits a marked reduction in generic "AI slop" and identifiable Gemini-isms, leading to more natural and human-like text.
- Increased Output Diversity: Users will find a greater variety in the model's responses and writing styles, improving creative flexibility.
- Fewer Soft Refusals: The model is designed to minimize "soft" refusals, providing more direct and usable output for writing tasks.
Training Details
The model's training incorporated several datasets, including internlm/Condor-SFT-20K to maintain general intellect and instruction-following, and a reformatted version of ConicCat/Gutenberg-SFT for supervised fine-tuning. Additional proprietary datasets, including backtranslated books and a third-party archive, were also utilized to achieve its specialized writing proficiency.