The Four Stages Of Instruct Training

A typical instruct-model training pipeline has four stages: _____(1) (next-token on the web), _____(2) (supervised fine-tuning on curated dialogs), _____(3)-model training, and _____(4) (policy update with the reward model). Fill the 4 blanks. Flag format: CONGRESS{1:[word],2:[word],3:[word],4:[word]}. Example: CONGRESS{1:pretrain,2:sft,3:reward,4:rlhf}.