Usage:
pretrain.sh: do pre-train (optional)sft.sh: do supervised fine-tunereward.sh: do reward modeling (must after sft.sh)ppo.sh: do PPO training (must after sft.sh and reward.sh)dpo.sh: do DPO training (must after sft.sh)predict.sh: do predict (must after sft.sh and dpo.sh)