dpo

Direct Preference Optimization runs, preference datasets, reward-free alignment behavior, and training stability observations.