Long form story with Qwen 0.8B

1 points | by aichelles an hour ago

2 comments

  • aichelles an hour ago

    Was getting tired of keep saying continue to gemini, perplexity and chatgpt for bed time stories with voice. So create a story teller with qwen 0.8b and supertonic using webgpu. Tried many different pipelines to reduce hallucinations to a minimum and keep the story coherent and finally made it work with the below.

    The Planner Step: The model generates a comprehensive structural plan (Title, synopsis, character arcs, twists, foreshadowing, and paragraph beats). The Writer Step: The model generates the story paragraph by paragraph. Each paragraph is an independent request. The worker doesn’t maintain an ever-growing context window; instead, it receives the immutable plan, the current paragraph beat, and a compressed continuity summary of what happened previously.Sanitization Pass: Small models love to output malformed internal markers. The client strips these transport-level signals before rendering.

    To my surprise, it created a ghost story with a simple prompt - "create a story" to test.