Nicu
a nano AI model
An AI trained from scratch, with a character of its own: a slightly dim guy from Catania. It's so small it can even run in your browser.
Why I made it
I built Nicu because there was nothing in Italian like TinyStories, Microsoft's research on tiny models that learn to write sensible stories. I made it from scratch, small and fully Italian. Along the way an idea came up: a character who doesn't make things up. When he doesn't know, he doesn't know, and he talks about himself.
What you learn
With little hardware and very low costs you go, in miniature, through the whole journey of a real model. First of all you write and clean up the dialogues it learns from; then the vocabulary, the shape of the network, the training and the polishing on the best answers. Not everything carries over one to one to large models, but it's a first foundation, and the results are surprising.
In numbers
- parameters
- 20.6 million
- the whole model, in one file
- 24 MB
- dialogues it learned to talk from
- 929,500
Where it's going
Nicu 1.2 is on nicu.chat: open it and talk, nothing to install.
- New versions with new training runs and more dialogue in the dataset, for better answers.
- Slightly bigger models, but still nano.
- Variants: characters with different personalities.
