training

Neural Information Processing Systems 

Traditional approaches focus on aligning models during the instruction tuning orreinforcement learning stages, referred tointhis paperas'postalignment'.