Train language models to follow human preferences.
Recent revisions are kept. Restoring one creates a new revision.