對于象META的開源大模型，如llama3，由于都是用通用數(shù)據(jù)進(jìn)行預(yù)訓(xùn)練，對想使用其模型的公司來說，可能會不適用，因?yàn)檫@大模型對公司的數(shù)據(jù)不熟悉，因此引入微調(diào)(Fine-Tunning)。
通過喂給大模型大量數(shù)據(jù)，1萬條起步，使得大模型也能對公司的數(shù)據(jù)熟悉，進(jìn)而用于各種對話場景。
1.克隆并安裝LLAMA FACTORY庫，install-llamafactory.sh
BIN_PATH=$(cd `dirname $0`; pwd)
cd $BIN_PATH/../
pwd
git clone --depth 1 https://github.com/hiyouga/LLaMA-Factory.git
cd LLaMA-Factory
pip install -e ".[torch,metrics,bitsandbytes,modelscope]"
2.設(shè)置環(huán)境變量
export USE_MODELSCOPE_HUB=1 #使用modelscop模型庫，非huggingface的
export CUDA_VISIBLE_DEVICES=0 ＃設(shè)置使用GPU
export HF_ENDPOINT=https://hf-mirror.com ＃設(shè)置huggingface的替代地址
export MODELSCOPE_CACHE=/root/autodl-tmp/models/modelscope ＃設(shè)置modelscope中的大模型保存路徑
export LLAMAFACTORY_HOME=/root/autodl-tmp/LLaMA-Factory
3.準(zhǔn)備數(shù)據(jù)
#在data/dataset_info.json中加入此數(shù)據(jù)
"docker_command_NL": {
"hf_hub_url": "MattCoddity/dockerNLcommands"
},
在data目錄中加入訓(xùn)練數(shù)據(jù)，MattCoddity/dockerNLcommands.json
數(shù)據(jù)格式為：
[
{
"input": "Give me a list of containers that have the Ubuntu image as their ancestor.",
"instruction": "translate this sentence in docker command",
"output": "docker ps --filter 'ancestor=ubuntu'"
},
] 4.訓(xùn)練大模型
訓(xùn)練的參數(shù)文件：llama3_lora_sft_docker_command.yaml
### model
#md model id
model_name_or_path: LLM-Research/Meta-Llama-3-8B-Instruct
#huggingface model id
#model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
### method
stage: sft
do_train: true
finetuning_type: lora
lora_target: all
### dataset
dataset: docker_command_NL
template: llama3
cutoff_len: 1024
max_samples: 1000
overwrite_cache: true
preprocessing_num_workers: 16
### output
output_dir: /root/autodl-tmp/my-test/saves/llama3-8b/lora/sft/docker-commnad-nlp/sft
logging_steps: 10
save_steps: 500
plot_loss: true
overwrite_output_dir: true
### train
per_device_train_batch_size: 4
gradient_accumulation_steps: 8
learning_rate: 1.0e-4
num_train_epochs: 3.0
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true
ddp_timeout: 180000000
### eval
val_size: 0.1
per_device_eval_batch_size: 1
eval_strategy: steps
eval_steps: 500
訓(xùn)練命令：lora-train-docker-command.sh
BIN_PATH=$(cd `dirname $0`; pwd)
cd $BIN_PATH/
pwd
cd $LLAMAFACTORY_HOME
pwd
llamafactory-cli train $BIN_PATH/conf/llama3_lora_sft_docker_command.yaml
執(zhí)行此命令即可開始訓(xùn)練大模型。
5.合并大模型
合并用的參數(shù)文件，llama3_lora_export_docker_command.yaml
### model
#md model id
model_name_or_path: LLM-Research/Meta-Llama-3-8B-Instruct
#huggingface model id
#model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct
adapter_name_or_path: /root/autodl-tmp/my-test/saves/llama3-8b/lora/docker-commnad-nlp/sft
template: llama3
export_dir: /root/autodl-tmp/my-test/saves/llama3-8b/lora/docker-commnad-nlp/export
finetuning_type: lora
export_size: 2
export_device: gpu
export_legacy_format: False
合并命令，lora-export-docker-command.sh
BIN_PATH=$(cd `dirname $0`; pwd)
cd $BIN_PATH/
pwd
llamafactory-cli export conf/llama3_lora_export_docker_command.yaml