The mean and standard deviation of the reward of the chosen responses are collected and used to normalize the reward in step 3. In one experiment, they are -0.8677118420600891 and 0.2210693359375 respectively and are used in the alpaca_rlhf/deepspeed_chat/training/step3_rlhf_finetuning/ppo_trainer.py#DeepSpeedPPOTrainer#generate_experience methods: 'rewards': (reward_score - (-0.8677118420600891)) / 0.2210693359375.