tiny-llm
A small language model trained from scratch on children's stories.
Start a story and it writes the rest.
It knows simple stories, not facts, so it can't answer questions or follow instructions.
Story opening
Write the story
Stop
Ctrl + Enter. Lower temperature: safer and more repetitive; higher: more surprising.
The story appears here, one token at a time.
Model: loading… ·
model card ·
source on GitHub ·
server status
This demo runs on a shared CPU, so it is slower than the GPU numbers in the project README,
and it can be busy or asleep for a minute.
Privacy: your text isn't stored. The server keeps request metadata (timings, token counts and a
salted hash of your address), never the opening or the story.
What is logged
Per request: time, status, timings (queue wait, time to first token, total), token counts,
whether it streamed, why it stopped, the error message if any, and how busy the server was.
Your address only as a salted hash: it can tell repeat visits apart, not who you are.
Raw addresses are never stored.
Never the opening you type or the story the model writes.
The log is kept in a private dataset that only the project owner can read, and is deleted after 30 days.