FableFlux

FableFlux harnesses open gpt-oss models to craft safe, personalized "fables" a new type of storytime magic to inspire kids imagination, expand their vocabulary, and build confidence.

Comment

Inspiration

Open models can reshape how the next generation learns, dreams, and sees themselves in the world.

Storytime is powerful — it sparks imagination, builds language, and shapes confidence. But parents often run out of fresh material, while kids crave stories that feel personal. With the rise of open models like GPT-OSS, we saw an opportunity to build Fable-Flux: a fine-tuned GPT-OSS-20B model that becomes a safe, magical storyteller for children ages 2–5 — right at a critical stage of their development.

What it does

FableFlux generates personalized, age-appropriate fables:

  • Safe and positive — every story is designed for young children.
  • Personalized — kids appear as heroes alongside their favorite characters.
  • Educational — vocabulary expands naturally through context and gentle repetition.
  • Confidence-building — children see themselves solving problems and exploring new worlds.

It’s a new kind of storytime magic — built on an open dataset, a fine-tuned model, and designed with families in mind.

How we built it

  1. Started with hand-crafted stories (written with my daughter) and open-sourced them on Hugging Face.
  2. Expanded the dataset from ~1K human stories to 10K+ (human + synthetic).
  3. Fine-tuned GPT-OSS-20B with QLoRA.
  4. Exported and uploaded the model to Hugging Face in multiple formats (BF16 + MXFP4).
  5. Integrated vLLM + Modal for fast inference.
  6. Built a lightweight front-end on Vercel so parents can instantly request new stories.

Challenges

  • Dataset curation — keeping stories consistently safe, positive, and educational; we added GPT-5 as a “judge” for quality control.
  • Personalization — balancing creativity with safe boundaries required careful prompt + fine-tune design.

Accomplishments

  • Created and open-sourced a synthetic dataset of children’s stories.
  • Fine-tuned GPT-OSS-20B with QLoRA and shipped a working demo — while juggling actual bedtime duties!
  • Built a system where a child can say “tell me a story about me and a dragon” and instantly get a safe, magical bedtime fable.

What we learned

  • Open models enable personalized creativity, not just productivity.
  • Safety comes from structure and curation, not only filters.
  • Parents care about trust + imagination more than technical details — build for the family first.
  • Evals matter — quality and safety hinge on measurement.

What’s next

  • Add soundscapes & music so every fable feels like a mini-audiobook.
  • Explore offline deployment (quantized smaller models for phones/tablets).
  • Build a library mode so parents can collect, print, and share their child’s favorite fables.

Built With

  • gpt-oss
  • modal
  • peft
  • transformers
  • vercel
  • vllm
Share this project:

Updates