Glossary · Model Training

DPO (Direct Preference Optimization)

A training technique that aligns a model to human preferences directly from preference-ranked data, without needing to train a separate reward model as RLHF does.