Every story starts
with something small.
A small model.
A little imagination.
Sprout learned to tell stories from scratch.
Give it an idea. See what grows.
An open research experiment
29.4 million parameters. Simple English stories.
The real model, running on your device.
Your story
Written by a small, imperfect model
What grew from this experiment?
A complete language model: byte-level tokenizer, eight transformer layers, pretraining, two controlled comparisons, instruction tuning, and cached generation. All built in plain Python and PyTorch.
Read the results19 / 20Test stories stopped on their own after instruction tuning.
6 / 20Included every requested word. Following instructions remains a weakness.
Sprout can repeat itself, change characters, or contradict its own story. It is a learning project with a 512-token context, not a general chat assistant. Outputs have not been evaluated for suitability for children.