Train prediction models on tables and CSVs.
Recent revisions are kept. Restoring one creates a new revision.
