Direct preference optimization
What does Direct preference optimization mean?
A method that tunes a model directly on preference pairs without a separate reinforcement-learning loop.
Also known as: DPO
A practical example
Training increases preference for a chosen response relative to a rejected one.