GPT-4.1 models gain preference fine-tuning

The GPT-4.1 family gained direct preference optimization fine-tuning. More control moved from prompt wording into training examples and preferences, useful, but also another layer teams have to evaluate carefully.