SimPO

A training technique that optimizes model outputs by comparing different responses to learn which ones are better, without requiring explicit reward models.