Adapt foundation models with supervised fine-tuning, parameter-efficient methods, preference data, and careful train-test separation.
