They say a lot of things, like that their base models weren't instruction-tuned, however people have confirmed that it's impossible to find instruction that it wouldn't follow, and the output would indicate that exactly. The labs absolutely love incorporating public evals in their training; of course, they're not going to admit that.