Skip to content

A full pipeline to finetune ChatGLM LLM with LoRA and RLHF on consumer hardware. Implementation of RLHF (Reinforcement Learning with Human Feedback) on top of the ChatGLM architecture. Basically ChatGPT but with ChatGLM

License

Notifications You must be signed in to change notification settings

jackaduma/ChatGLM-LoRA-RLHF-PyTorch

Repository files navigation

ChatGLM-LoRA-RLHF-PyTorch

a full pipeline to finetune ChatGLM LLM with LoRA and RLHF on consumer hardware


Table of Contents


Environment Setup

穷人卡:2080Ti 12G
torch==2.0.0
cuda==11.8

Todo List

  • SFT: Supervised Finetune
  • Merge Adapter into Model
  • RLHF
    • train reward model
    • tuning with RL

Run


Data Process

转化alpaca数据集为jsonl

python cover_alpaca2jsonl.py --data_path data/alpaca_data.json --save_path data/alpaca_data.jsonl

tokenization

python tokenize_dataset_rows.py --jsonl_path data/alpaca_data.jsonl --save_path data/alpaca --max_seq_length 200 --skip_overlength True

Supervised Finetune

must use latest peft version

pip uninstall peft -y
pip install git+https://github.com/huggingface/peft.git  # 最新版本 >=0.3.0.dev0
python supervised_finetune.py --dataset_path data/alpaca --lora_rank 8 --per_device_train_batch_size 1 --gradient_accumulation_steps 32 --save_steps 200 --save_total_limit 3  --learning_rate 1e-4 --fp16 --remove_unused_columns false --logging_steps 10 --output_dir output

Merge PEFT adapter into Model

pip uninstall peft -y
pip install peft==0.2.0  # 0.3.0.dev0 raise many errors
python merge_peft_adapter.py --model_name ./output 

Reward Modeling

python train_reward_model.py --model_name 'THUDM/chatglm-6b' --gradient_accumulation_steps 32 --per_device_train_batch_size 1 --train_subset 100 --eval_subset 10 --local_rank 0 --bf16 False

merge reward model into Model

python merge_peft_adapter.py --model_name ./reward_model_chatglm-6b

Notes

  1. PEFT的版本,目前从git上安装的是 0.3.0.dev0 版本,在merge_peft_adapter的时候有问题,需要切换到peft==0.2.0 (0.3.0.dev0 没有 _get_submodules()这个函数)
  2. 因为huggingface的transformer暂时不支持ChatGLM的封装接口,需要自己从ChatGLM的hub上下载代码放到本地目录 models 下面,供后续使用
  3. 同样,ChatGLM的model代码是自己的,和huggingface没合并,所以在调用加载的时候,都主要加上参数 trust_remote_code=True
  4. 训练 Reward Model 需要执行 SeqCLS 这个Task: huggingface 的 transformer 提供 "AutoModelForSequenceClassification" 这个类。但是 ChatGLM 只有 "ChatGLMForConditionalGeneration" 这个类。
  5. 自己实现 Reward model, reward_model.py,完成奖励模型的训练过程

Reference

data preprocess: cover_alpaca2jsonl.pytokenize_dataset_rows.py 来自项目 ChatGLM-Tuning

requirements 主要是按照 alpaca-lora 来配环境。


Star-History

star-history


Donation

If this project help you reduce time to develop, you can give me a cup of coffee :)

AliPay(支付宝)

ali_pay

WechatPay(微信)

wechat_pay

License

MIT © Kun

About

A full pipeline to finetune ChatGLM LLM with LoRA and RLHF on consumer hardware. Implementation of RLHF (Reinforcement Learning with Human Feedback) on top of the ChatGLM architecture. Basically ChatGPT but with ChatGLM

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages