| 1 | import argparse |
| 2 | |
| 3 | def parse_args(): |
| 4 | parser = argparse.ArgumentParser(description='MOSS-RLHF @Fudan NLP Group') |
| 5 | |
| 6 | # Path |
| 7 | parser.add_argument('--model_save_path', type=str, default='', help='checkpoint path, used for save model and training') |
| 8 | parser.add_argument('--policy_model_path', type=str, default='', help='policy model and reference model path') |
| 9 | parser.add_argument('--critic_model_path', type=str, default='', help='critic model and reward model path') |
| 10 | parser.add_argument('--tokenizer_name_or_path', type=str, default='/huggingface_models/open-chinese-llama-7b', help='tokenizer name or path') |
| 11 | parser.add_argument('--data_path', type=str, default='./data', help='dataset for training and validation') |
| 12 | parser.add_argument('--logdir', type=str, default=None, help='path to save tensorboard logs') |
| 13 | |
| 14 | # Training |
| 15 | parser.add_argument('--lr', type=float, default=5e-7, help='learning rate of policy model') |
| 16 | parser.add_argument('--critic_lr', type=float, default=15e-7, help='learning rate of critic model') |
| 17 | parser.add_argument('--seed', type=int, default=42, help='seed') |
| 18 | parser.add_argument('--batch_size', type=int, default=32, help='training batch size, *NOT* for sampling from env') |
| 19 | parser.add_argument('--train_steps', type=int, default=5000, help='train steps') |
| 20 | parser.add_argument('--warmup_steps', type=int, default=500, help='warmup steps') |
| 21 | parser.add_argument('--save_per_step', type=int, default=100, help='save ckpt per steps') |
| 22 | parser.add_argument('--beta1', type=float, default=0.9, help='adam') |
| 23 | parser.add_argument('--beta2', type=float, default=0.95, help='adam') |
| 24 | parser.add_argument('--eps', type=float, default=1e-6, help='optimizer') |
| 25 | parser.add_argument('--num_workers', type=int, default=1, help='dataloader') |
| 26 | parser.add_argument('--num_prefetch', type=int, default=32, help='dataloader') |
| 27 | parser.add_argument('--maxlen_prompt', type=int, default=2048, help='max len for training, including model prompt and response') |
| 28 | parser.add_argument('--gradient_checkpoint', action='store_true', help='deepspeed') |
| 29 | |
| 30 | # PPO in LLMs |
| 31 | parser.add_argument('--num_rollouts', type=int, default=128, help='nums of samples in current replay buffer') |
| 32 | parser.add_argument('--rollout_batch_size', type=int, default=32, help='batch size of sampling from env') |
| 33 | |
| 34 | parser.add_argument('--ppo_pretrain_data_path', type=str, default='', help='dataset folder path for pertrain loss of step3: rlhf') |
| 35 | parser.add_argument('--ppo_pretrain_data_type', type=str, default='sft', choices=['sft', 'pretrain'], help='dataset folder path for pertrain loss of step3: rlhf') |
| 36 | parser.add_argument('--ppo_pretrain_batch_size_ratio', type=int, default=1, help='ppo batch size ratio') |
| 37 | parser.add_argument('--ppo_pretrain_loss_weight', type=float, default=0., help='add pretrain loss in PPO training: ppo-rtx') |
| 38 | parser.add_argument('--kl_penalty_weight', type=float, default=0.02, help='kl penalty') |
| 39 | parser.add_argument('--advantage_clip', type=float, default=0.5, help='clip advantage') |
| 40 | parser.add_argument('--vf_loss_weight', type=float, default=1., help='vf loss weight') |
| 41 | parser.add_argument('--entropy_loss_weight', type=float, default=0., help='entropy loss weight') |
| 42 | parser.add_argument('--reward_clip', type=float, default=10., help='reward clip') |
| 43 | parser.add_argument('--entropy_clip', type=float, default=35., help='entropy loss clip') |
| 44 | parser.add_argument('--pg_clip', type=float, default=0.2, help='pg loss clip') |
| 45 | parser.add_argument('--value_clip', type=float, default=0.2, help='value clip for critic model') |
| 46 | parser.add_argument('--gamma', type=float, default=1., help='GAE in PPO') |
| 47 | parser.add_argument('--lam', type=float, default=0.95, help='GAE in PPO') |
| 48 | |
| 49 | # Trick and method options for PPO |
| 50 | parser.add_argument('--use_reward_clip', action='store_true', help='use reward clip') |
| 51 | parser.add_argument('--use_reward_scaling', action='store_true', help='use reward scaling') |
| 52 | parser.add_argument('--use_reward_norm', action='store_true', help='user reward norm') |
| 53 | parser.add_argument('--use_critic_loss_clip', action='store_true', help='use critic loss clip') |
| 54 | parser.add_argument('--use_policy_loss_clip', action='store_true', help='use policy loss clip') |
| 55 | parser.add_argument('--use_advantage_norm', action='store_true', help='use advantage norm') |
| 56 | parser.add_argument('--use_advantage_clip', action='store_true', help='use advantage clip') |
| 57 | parser.add_argument('--use_ppo_pretrain_loss', action='store_true', help='use ppo pretrain loss') |
| 58 | parser.add_argument('--use_entropy_loss', action='store_true', help='use ppo entropy loss') |
| 59 | |
| 60 | # Sample from env |