模型下载

https://huggingface.co/baichuan-inc

微调框架

https://github.com/hiyouga/LLaMA-Factory

环境搭建

git clone https://github.com/hiyouga/LLaMA-Factory.git
conda create -n llama_factory python=3.10
conda activate llama_factory
cd LLaMA-Factory
pip install -r requirements.txt
#注意Baichuan2-13B-Chat模型的transformers==4.33.3 ,trl==0.7.2高版本跟模型不兼容

数据集

仿照格式，增加自己的数据集

https://github.com/hiyouga/LLaMA-Factory/blob/main/data/dataset_info.json

微调命令

这里要注意，github里的命令要将template改为baichuan2，lora_target改为W_pack

CUDA_VISIBLE_DEVICES=0 python src/train_bash.py \
    --stage sft \
    --model_name_or_path path_to_llama_model \
    --do_train \
    --dataset alpaca_gpt4_zh \
    --template  baichuan2\
    --finetuning_type lora \
    --lora_target  W_pack\
    --output_dir path_to_sft_checkpoint \
    --overwrite_cache \
    --per_device_train_batch_size 4 \
    --gradient_accumulation_steps 4 \
    --lr_scheduler_type cosine \
    --logging_steps 10 \
    --save_steps 1000 \
    --learning_rate 5e-5 \
    --num_train_epochs 3.0 \
    --plot_loss \
    --fp16

多GPU训练

accelerate config 
accelerate launch src/train_bash.py \
    --stage sft \
    --model_name_or_path path_to_llama_model \
    --do_train \
    --dataset alpaca_gpt4_zh \
    --template  baichuan2\
    --finetuning_type lora \
    --lora_target  W_pack\
    --output_dir path_to_sft_checkpoint \
    --overwrite_cache \
    --per_device_train_batch_size 4 \
    --gradient_accumulation_steps 4 \
    --lr_scheduler_type cosine \
    --logging_steps 10 \
    --save_steps 1000 \
    --learning_rate 5e-5 \
    --num_train_epochs 3.0 \
    --plot_loss \
    --fp16