[alecco]

At a glance, it looks like Phi-3 was trained on an English only, STEM-strong dataset. See how they are not as strong in HumanEval, Trivia, etc. But of course it's very good.