Canonical prompt datasets were used for generating data for SFT and for performing RL (as well as evals).