Gryphe/Gemma-4-31B-StyleTune
Gryphe/Gemma-4-31B-StyleTune is a 31 billion parameter Gemma 4 model developed by Gryphe, specifically fine-tuned to alter its writing style. This model focuses on reducing clichés and introducing a new narrative voice by exclusively training the lm_head output projection layer. It retains all of Gemma 4's original reasoning and world knowledge capabilities while offering a distinct stylistic output, making it suitable for creative writing and narrative generation tasks.
Loading preview...
Overview
Gryphe/Gemma-4-31B-StyleTune is a unique 31 billion parameter Gemma 4 model that has undergone a "style tune" rather than a traditional full fine-tune. Developed by Gryphe, this model specifically targets and modifies the writing style by training only the lm_head output projection layer, which dictates token emission. This approach allows for a significant change in the model's narrative voice while preserving all of Gemma 4's underlying capabilities, such as reasoning, world knowledge, and instruction following.
Key Capabilities & Differentiators
- Targeted Style Modification: Achieves a new writing style by exclusively fine-tuning the
lm_headlayer, making it highly efficient in terms of VRAM and training time. - Reduced Clichés: Benchmarked to produce 60% fewer clichés per 100 words compared to the base instruct model.
- Distinct Vocabulary: Exhibits only 21.7% shared trigram vocabulary with the base model, indicating a significantly different phraseology.
- Preserved Core Abilities: All of Gemma 4's original reasoning, world knowledge, and language understanding capabilities remain fully intact.
Good For
- Creative Writing: Ideal for generating narrative content with a fresh, less cliché-ridden style.
- Roleplay Scenarios: Provides responses with a distinct voice, enhancing immersive roleplay experiences.
- Stylistic Experimentation: Useful for developers looking to explore targeted stylistic changes in large language models without extensive retraining.