electroglyph/gemma4-26b-fiction-bf16
electroglyph/gemma4-26b-fiction-bf16 is a 26 billion parameter Gemma 4 model, fine-tuned by electroglyph for creative writing, specifically in the style of established authors like William Gibson and George R.R. Martin. This model excels at generating detailed, atmospheric fiction and is optimized for nuanced stylistic imitation, offering enhanced narrative quality compared to its base model. It was trained using Unsloth and Huggingface's TRL library, leveraging a unique dataset of 700 books and a structured output dataset.
Loading preview...
Model Overview
electroglyph/gemma4-26b-fiction-bf16 is a 26 billion parameter Gemma 4 model, fine-tuned by electroglyph to specialize in generating high-quality, stylistically accurate fiction. This model builds upon TrevorJS/gemma-4-26B-A4B-it-uncensored, with a focus on improving narrative coherence and stylistic imitation, particularly for authors like William Gibson and George R.R. Martin.
Key Capabilities
- Advanced Stylistic Imitation: Demonstrates a notable improvement in mimicking the distinct writing styles of specified authors, as evidenced by comparative outputs for cyberpunk noir and medieval fantasy prompts.
- Detailed Narrative Generation: Capable of producing rich, atmospheric descriptions and intricate plot developments, maintaining consistency over longer text generations (e.g., 700-1000 words).
- Enhanced Coherence: Addresses some "sloppiness" observed in its parent model, leading to slightly more readable and structured outputs.
- Efficient Fine-tuning: Trained 2x faster using Unsloth and Huggingface's TRL library on AMD MI300X GPUs, utilizing a full precision LoRA with rank 16 and alpha 32.
Good For
- Creative Writing & Storytelling: Ideal for generating fiction in specific authorial styles, such as gothic horror, cyberpunk noir, or gritty medieval fantasy.
- Content Generation for Games & Media: Useful for creating detailed lore, character backstories, or scene descriptions that require a particular tone and atmosphere.
- Stylistic Experimentation: Developers and writers looking to explore and generate text that adheres to complex stylistic constraints will find this model particularly effective.