Wanabi-Novelist-12B: Specialized Japanese Novel Writing Assistant

Wanabi-Novelist-12B is a 12 billion parameter Japanese large language model, meticulously fine-tuned for novel writing assistance. It is built on the Mistral-Nemo-Base-2407 architecture and is primarily intended for use with the dedicated desktop application, Project Wannabe.

Key Capabilities (via Project Wannabe):

• Text Generation (GEN): Creates initial novel sections or new scenes based on metadata (title, synopsis, plot) or freely.
• Continuation Generation (CONT): Produces natural continuations of existing text, guided by context and "author's notes" for specific plot directions.
• Idea Generation (IDEA): Develops comprehensive novel ideas (title, synopsis, plot, settings) from fragmented keywords or genres.

Differentiators & Improvements:

• Optimized for Novel Writing: Unlike general instruction models, it excels in structured creative writing tasks, supporting features like rating (general/r18) and dialogue volume control.
• Refined Dataset: Trained on a comprehensively rebuilt dataset including ~24,000 web novels and ~7,000 Aozora Bunko texts, ensuring diverse writing styles and improved data quality (e.g., proper indentation, no noise).
• Lighter Footprint: At 12B parameters, it offers reduced VRAM consumption compared to the 24B version, making it more accessible.
• Enhanced Control: Improved idea generation from brief inputs and more precise dialogue volume specification for upcoming text.

Important Considerations:

• Specialized Use: This model is not a general-purpose Instruct model; its performance is optimized for the specific prompt formats used by Project Wannabe.
• Context Length: While theoretically supporting 128K, it was trained with a maximum context of 32K, so performance may degrade with longer contexts.

For optimal use and to leverage its full capabilities, Project Wannabe is strongly recommended, as it handles the complex structured prompting required by the model.