Run your first SSL experiment

This guide targets the current source checkout. The packaged quickstart and Jet are unreleased additions; an older PyPI release does not include them.

git clone https://github.com/galilai-group/stable-pretraining.git
cd stable-pretraining
python -m pip install -e .
python -m stable_pretraining.quickstart
spt web ./spt-quickstart/runs

The command uses synthetic images, a small convolutional encoder, SimCLR, an online classification probe, and a CPU Trainer through spt.Manager. It completes four backbone optimizer steps and validates on separate synthetic images. No model or dataset download is needed. A successful run prints Completed 4 optimizer steps and a command for the local viewer.

The registry stores metrics under the selected cache directory and Lightning writes a checkpoint in the run directory. Inspect fit/loss and eval/probe_accuracy in the viewer. These synthetic metrics verify that training and evaluation work; they are not a benchmark.

python -m stable_pretraining.quickstart --epochs 2 --cache-dir ./my-first-run
python -m stable_pretraining.quickstart --method jet-entropy --cache-dir ./jet-demo

The second command exercises MSE prediction minus exact full-token logdet with the invertible Jet encoder. It is a demonstration, not a tuned training recipe.

For a published release, install python -m pip install stable-pretraining and consult documentation for that release. Check what is installed with:

python -c "import importlib.metadata as m; print(m.version('stable-pretraining'))"

For an editable experiment, see the starter project. Next: custom images, online evaluation, checkpoint/resume, and Jet.