How we trained our base models