TinyStories 33M: a small language model in your browser

Story

Prompt

Loading the model

Starting

The model writes short children's stories. It is not a chatbot, and it only knows the simple words and plots of TinyStories.

Settings

Temperature 0.80

0 is greedy; higher values wander more.

Top-k 50

Only the k most likely tokens are kept. 0 keeps all of them.

Top-p 1.00

Keep the smallest set of tokens that holds this much probability.

Engine

Switch backend

Backend: -
Speed: -
Generated: -
Prefill: -

Details

Runtime: -
GPU: -
CPU threads: -

About this model

A 33M-parameter decoder-only transformer, trained from scratch on TinyStories (short children's stories) for 450M tokens, on a GTX 1650 Ti laptop GPU in about 21 hours. Final validation loss: 1.37.

Everything runs on your device. The only network request is the one-time download of the model file, which your browser then keeps.

parthparekhdev.com Source code on GitHub