StoryGemma

Created by team StoryGemma on December 01, 2024

I've built a pipeline that takes human-written stories and generates metadata that allows a small model to learn how to tell novel stories in the way a human does. The metadata generated includes likely instructions from a human that would lead to generation of a given progression in the story, the likely internal thoughts that the writer would have as they wrote the progression of the story, and more. The metadata is then formatted in input/output pairs, and the model is trained on those pairs using SFT. The resulting model demonstrates a much more natural writing style than most AI models, and is highly steerable

Category tags:
  • Profile image of null

    Team member not visible

    This profile isn't complete

  • Profile image of Nguyễn Nam

    Nguyễn Nam Quân