COMET score

New DPO method boosts NMT model performance with preference-based post-training

Researchers have developed a new post-training method for neural machine translation (NMT) systems that utilizes reinforcement learning and Direct Preference Optimization (DPO). This framework requires only a general te…