I might have used the wrong training method—the file is 500MB—but I think it turned out reasonably similar