Gemma-3-Glitter-12B
| Property | Value |
|---|---|
| Base Model | Gemma 3 12B IT |
| Author | allura-org |
| Model URL | HuggingFace |
What is Gemma-3-Glitter-12B?
Gemma-3-Glitter-12B is a specialized creative writing model that combines roleplay instruction and storytelling capabilities. It's built through a 50/50 merge of two distinct training approaches: an instruct-based RP training (~13.5M tokens) and long-form creative writing completion training (~20M tokens). The model maintains vision capabilities from its base architecture while adding enhanced creative writing features.
Implementation Details
The model implements a unique training methodology combining two separate training sets: ToastyPigeon/g3-12b-rp-system-v0.1 and ToastyPigeon/g3-12b-storyteller-v0.2-textonly. The architecture supports both standard Gemma2/3 instruct format and an optional system role, providing flexibility in implementation.
- RP Training: ~13.5M tokens with 2:1 human to synthetic ratio
- Storyteller Training: ~20M tokens including 1.6M synthetic from R1
- Vision Capability: Fully supported
- Custom Instruct Format Support
Core Capabilities
- Creative Writing and Storytelling
- Roleplay System Integration
- Vision Processing
- Flexible Prompt Formatting
- System Role Implementation
Frequently Asked Questions
Q: What makes this model unique?
The model's unique strength lies in its balanced fusion of roleplay and storytelling capabilities, combined with maintained vision features. The dual training approach ensures both structured instruction following and creative freedom.
Q: What are the recommended use cases?
The model excels in creative writing tasks, storytelling, roleplay scenarios, and applications requiring both visual and narrative capabilities. It's particularly suited for interactive storytelling and creative content generation with visual context.
