Daizee/Dirty-Calla-4B

VISIONConcurrency Cost:1Model Size:4.3BQuant:BF16Ctx Length:32kPublished:Oct 26, 2025License:otherArchitecture:Transformer Cold

Dirty-Calla-4B by Daizee is a 4.3 billion parameter instruction-tuned Gemma-3 family model, specifically fine-tuned for generating short fictional stories. This model excels at creative writing tasks, particularly fanfiction, based on user prompts with specified characters and themes. It leverages a light style SFT approach on a synthetically augmented dataset of fanfiction to achieve its specialized narrative generation capabilities.

Loading preview...

Overview

Daizee/Dirty-Calla-4B is a 4.3 billion parameter language model derived from Daizee/Gemma3-Callous-Calla-4B, which itself is based on the Gemma-3 4B instruction-tuned architecture. This model has undergone a light style Supervised Fine-Tuning (SFT) process, primarily on a dataset of fanfiction with user-provided prompts. The training data was synthetically expanded from approximately 360 to 3000 examples to enhance its narrative generation abilities.

Key Capabilities

  • Specialized Story Generation: Designed to write short fictional stories, particularly fanfiction, based on tags and general plot ideas provided by the user.
  • Style SFT: Fine-tuned for a specific narrative style, making it suitable for creative writing tasks.
  • Gemma-3 Architecture: Benefits from the underlying Gemma-3 instruction-tuned architecture and tokenizer.

Good For

  • Creative Writing: Ideal for generating short stories, fanfiction, or narrative content.
  • Prompt-based Storytelling: Users can provide character names, themes, and general concepts to guide story creation.
  • Exploratory Fiction: Useful for quickly drafting fictional scenarios or expanding on narrative ideas.