Pull a model

$ pip install systemonemodels
Successfully installed systemonemodels-0.2.4
$ systemone pull convai-innovations/laya
19 files · downloaded 2.4 GB
$ systemone run studio
System One Studio  http://127.0.0.1:8765
$ systemone login
✓ Signed in as you
$ systemone push ./my-router
✓ Published you/[email protected] with 4 files

$ 

The registry and the CLI

Every model on this site has a page with its versions, its files and what its authors report: what it decides, its accuracy, its calibration error and its speed. The systemone command reads the same registry from your terminal.

pip install systemonemodels

It needs Python 3.10 or newer, and installs the systemone command and the systemone Python package.

Find a model

systemone search routing --capability route --sort latency
systemone show convai-innovations/laya

search filters by what a model decides (--capability), by --architecture and by --license, and --sort orders the results by downloads, accuracy, latency and more. show prints one model's facts. A model is always named namespace/name.

Pull only what you need

A repository can hold several variants of a model, one folder each. Laya has an English checkpoint, a multilingual one, and one trained on typed questions. --variant fetches one folder, and --dest puts the files where you want them:

systemone pull convai-innovations/laya --variant typed-decisions --dest ./laya

Every file is checked against its SHA-256 and cached by content. A second pull downloads nothing, and variants that share a file store it once. Versions never change once published, and --version pins one, so a pull that works today keeps working. systemone cache info shows where the cache is and how big it has grown.

From Python

from systemone import snapshot_download

path = snapshot_download("convai-innovations/laya", variant="typed-decisions")

path is a local folder with the files, served from the cache after the first call.

What you have now

The model's weights and the files that go with them. How to run them depends on the runtime they were made for, and the model card says which. The next lesson runs models from several families with one runtime.