Train embedding models with before-and-after scores.
Recent revisions are kept. Restoring one creates a new revision.