alpaca-rlhf

Finetuning alpaca with RLHF (Reinforcement Learning with Human Feedback). The base model is from my-alpaca and multi-turn-alpaca, which train LLaMA with the original alpaca dataset and a multi-turn dialogue dataset respectively.

Online Demo

• SFT
• RLHF

Modifications on DeepSpeed Chat

Step 1

• alpaca_rlhf/deepspeed_chat/training/step1_supervised_finetuning/main.py#main()
  • Set special tokens
    • 
• alpaca_rlhf/deepspeed_chat/training/utils/data/data_utils.py#create_dataset_split()
  • Train only on responses and add eos
    • 
  • Remove end_of_conversation_token
    • 
• alpaca_rlhf/deepspeed_chat/training/utils/data/data_utils.py#PromptDataset#getitem
  • Labels differs from input
    • 
• alpaca_rlhf/deepspeed_chat/training/utils/data/raw_datasets.py#MultiTurnAlpacaDataset
  • add MultiTurnAlpacaDataset
    • 
• alpaca_rlhf/deepspeed_chat/training/utils/module/lora.py#convert_linear_layer_to_lora
  • Support multiple module names for lora
    • 

Step 2

• alpaca_rlhf/deepspeed_chat/training/step2_reward_model_finetuning/main.py#main()
  • Set special tokens
    • 
• alpaca_rlhf/deepspeed_chat/training/utils/model/reward_model.py#RewardModel#forward()
  • Fixing the numerical instability
    • 
• alpaca_rlhf/deepspeed_chat/training/utils/data/data_utils.py#create_dataset_split()
  • Remove end_of_conversation_token
    • 

Step 3

• alpaca_rlhf/deepspeed_chat/training/step3_rlhf_finetuning/main.py#main()
  • Set special tokens
    • 
    • 
• alpaca_rlhf/deepspeed_chat/training/utils/data/data_utils.py#create_dataset_split()
  • Fix max length bug
    • 
• alpaca_rlhf/deepspeed_chat/training/utils/data/data_utils.py#DataCollatorRLHF#call
  • Fix pad_token_id bug
    • 
  • Fix padding side bug
    • 
• alpaca_rlhf/deepspeed_chat/training/step3_rlhf_finetuning/ppo_trainer.py#DeepSpeedPPOTrainer#generate_experience
  • Normalize reward
    • 

Stey by Step

• Running all three steps on 2 x A100 80G
• Bootstrap Script
  • step1: --num_gpus 2 /tmp/pycharm_project_227/alpaca_rlhf/deepspeed_chat/training/step1_supervised_finetuning/main.py --sft_only_data_path MultiTurnAlpaca --data_output_path /root/autodl-tmp/rlhf/tmp/ --model_name_or_path decapoda-research/llama-7b-hf --per_device_train_batch_size 8 --per_device_eval_batch_size 8 --max_seq_len 512 --learning_rate 3e-4 --num_train_epochs 1 --gradient_accumulation_steps 8 --num_warmup_steps 100 --output_dir /root/autodl-tmp/rlhf/actor --lora_dim 8 --lora_module_name q_proj,k_proj --only_optimize_lora --deepspeed --zero_stage 2
    • when --sft_only_data_path MultiTurnAlpaca is added, unzip data/data.zip
  • step2: --num_gpus 2 /tmp/pycharm_project_227/alpaca_rlhf/deepspeed_chat/training/step2_reward_model_finetuning/main.py --data_output_path /root/autodl-tmp/rlhf/tmp/ --model_name_or_path decapoda-research/llama-7b-hf --num_padding_at_beginning 0 --per_device_train_batch_size 4 --per_device_eval_batch_size 64 --learning_rate 5e-4 --num_train_epochs 1 --gradient_accumulation_steps 1 --num_warmup_steps 0 --zero_stage 2 --deepspeed --output_dir /root/autodl-tmp/rlhf/critic --lora_dim 8 --lora_module_name q_proj,k_proj --only_optimize_lora
    • the training process of step 2
      • 
      • 
      • 
    • 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 generate_experience methods: 'rewards': (reward_score - (-0.8677118420600891)) / 0.2210693359375.
  • step3: --num_gpus 2 /tmp/pycharm_project_227/alpaca_rlhf/deepspeed_chat/training/step3_rlhf_finetuning/main.py --data_output_path /root/autodl-tmp/rlhf/tmp/ --actor_model_name_or_path /root/autodl-tmp/rlhf/actor/ --tokenizer_name_or_path decapoda-research/llama-7b-hf --critic_model_name_or_path /root/autodl-tmp/rlhf/critic --actor_zero_stage 2 --critic_zero_stage 2 --num_padding_at_beginning 0 --per_device_train_batch_size 4 --per_device_mini_train_batch_size 4 --ppo_epochs 2 --actor_learning_rate 9.65e-6 --critic_learning_rate 5e-6 --gradient_accumulation_steps 1 --deepspeed --actor_lora_dim 8 --actor_lora_module_name q_proj --critic_lora_dim 8 --critic_lora_module_name q_proj,k_proj --only_optimize_lora --output_dir /root/autodl-tmp/rlhf/final
    • the training process of step 3
      • 
      • 
      • 
• Inference
  • nohup sh run_inference.sh 0 alpaca_rlhf/inference/llama_chatbot_gradio.py --path /root/autodl-tmp/rlhf/final/actor > rlhf_inference.log 2>&1 &
  • nohup sh run_inference.sh 0 alpaca_rlhf/inference/llama_chatbot_gradio.py --path /root/autodl-tmp/rlhf/actor > sft_inference.log 2>&1 &

Comparison between SFT and RLHF

• Chat
  • SFT
    • 
  • RLHF
    • 
• Write stories
  • SFT
    • 
  • RLHF
    • 

References

Articles

• 如何正确复现 Instruct GPT / RLHF?
• 影响PPO算法性能的10个关键技巧（附PPO算法简洁Pytorch实现）

Sources

• Awesome RLHF

Tools

• DeepSpeed-Chat

Datasets

• Stanford Human Preferences Dataset (SHP)
• HH-RLHF
  • hh-rlhf
    • Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback [paper]
    • Dahoas/static-hh
    • Dahoas/rm-static
• GPT-4-LLM
  • GitHub
  • Paper
  • Site
• Open-Assistant
  • Site
  • GitHub
  • Paper

Related Repositories

• my-alpaca
• multi-turn-alpaca