nbeerbower/Gemma4-Gutenberg-31B

Gemma4-Gutenberg-31B: Literary Fiction Specialist

This model, developed by nbeerbower, is a 31 billion parameter Gemma 4 variant specifically fine-tuned for generating literary and novelistic prose. It builds upon the already strong google/gemma-4-31B-it base, pushing its capabilities towards a distinct literary-fiction register.

Key Capabilities & Differentiators

• Literary Prose Generation: Optimized for producing text with a focus on story, interiority, and controlled pacing, actively avoiding generic "AI slop" phrasing.
• ORPO Fine-tuning: Utilizes the Odds Ratio Preference Optimization (ORPO) method on the comprehensive schneewolflabs/Athanorlite-DPO dataset, which includes various Gutenberg-series DPO datasets and balance sets.
• Preference Learning: The training trajectory shows strong convergence in reward accuracy, indicating the model effectively learned to prefer literary 'chosen' texts over 'rejected' generic content.
• Unified Multimodal Base: While fine-tuning focused solely on the text decoder, the model retains the frozen and intact vision/audio towers of the base Gemma 4 31B, making it a drop-in replacement for multimodal applications where text output is literary.
• Refinement, Not Rescue: This model is presented as a refinement of an already capable writer, offering a consistent literary lean rather than a complete transformation.

Ideal Use Cases

• Creative Writing: Generating fiction, novels, short stories, or descriptive passages with a literary style.
• Content Creation: Producing narrative-driven content that requires sophisticated prose and character depth.
• Literary Experimentation: Exploring AI's capacity for nuanced storytelling and stylistic control.