Interesting. Wasn't their previous attempt (Phi) trained mostly on synthetic data?