Do you even need to fine tune? What happens if you just pick the next token but exclude options that contain e? Or would that send you down cul-de-sacs? Maybe it would work with diffusion based models...