[jacobr1]

One wrinkle, is that it is now common to fine-tune on previously derived RL datasets, with the tested inputs and preferred sample outputs as the training data.