How to know if one should fine tune/pretrain or RL / reasoning train given some data set?