Preference tuning

Also written: preference-tuned

Training on pairs of better and worse answers to the same prompt instead of copying single good answers. DPO is a common method for it.