Use a more robust way to distribute models

[luispedro]

Currently, we just distribute pickled scikit-learn objects, which works but is brittle to version changes

A more robust approach would solve this