There’s no objective measurement for high quality code, so I don’t think model creators are going to be particularly good at screening for it.