Glossary · Model Training

Direct Preference Optimization (DPO)

A model-alignment technique that trains directly on human preference comparisons between output pairs, without needing a separately trained reward model as in classic RLHF.