LLM微调实践

安装相关库

pip install transformers
pip install accelerate
pip install huggingface_hub
pip install ipywidgets
pip install datasets
pip install trl
pip install peft
pip install swanlab

下载模型与数据集

import os
os.environ['HF_ENDPOINT'] = 'https://hf-mirror.com'
from huggingface_hub import snapshot_download

• 注意HF_ENDPOINT环境变量需在huggingface相关库之前，否则不生效

然后下载模型与数据集

snapshot_download(repo_id="b-mc2/sql-create-context",
                  repo_type="dataset",
                  local_dir="data",
                  max_workers=8,
                  resume_download=True
                  )

snapshot_download(
    repo_id="codellama/CodeLlama-7b-hf",
    local_dir="CodeLlama-7b-hf",
    resume_download=True,
    max_workers=8
)

以下是训练代码

from datasets import load_dataset
import swanlab
from swanlab.integration.huggingface import SwanLabCallback
import torch
from trl import SFTTrainer
from peft import LoraConfig
from transformers import TrainingArguments
from transformers import AutoTokenizer, AutoModelForCausalLM
from trl import setup_chat_format

import swanlab

swanlab.login(api_key='')

model_id = "internlm2_5-1_8b-chat" # or `mistralai/Mistral-7B-v0.1`
dataset = load_dataset("json",data_files="train_dataset.json",split="train",)

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
#     attn_implementation="flash_attention_2",
    torch_dtype=torch.bfloat16,
    trust_remote_code=True
)

tokenizer = AutoTokenizer.from_pretrained(model_id,trust_remote_code=True)
tokenizer.padding_side = 'right' # 以防止警告

model, tokenizer = setup_chat_format(model, tokenizer)

# 基于QLoRA论文和Sebastian Raschka实验的LoRA配置
peft_config = LoraConfig(
        lora_alpha=128,
        lora_dropout=0.05,
        r=256,
        bias="none",
        target_modules="all-linear",
        task_type="CAUSAL_LM",
)

args = TrainingArguments(
    output_dir="code-llama-7b-text-to-sql", # 要保存的目录和存储库ID
    num_train_epochs=3,                     # 训练周期数
    per_device_train_batch_size=3,          # 训练期间每个设备的批量大小
    gradient_accumulation_steps=2,          # 反向/更新前的步骤数
    gradient_checkpointing=True,            # 使用渐变检查点来节省内存
    optim="adamw_torch",              # 使用融合的adamw优化器
    logging_steps=10,                       # 每10步记录一次
    save_strategy="epoch",                  # 每个epoch保存检查点
    learning_rate=2e-4,                     # 学习率，基于QLoRA论文
#     bf16=True,                              # 使用bfloat16精度
    max_grad_norm=0.3,                      # 基于QLoRA论文的最大梯度范数
    warmup_ratio=0.03,                      # 根据QLoRA论文的预热比例
    lr_scheduler_type="constant",           # 使用恒定学习率调度器
    report_to="tensorboard",                # 将指标报告到Tensorboard
)

swanlab_callback = SwanLabCallback(experiment_name="TransformersTest", cloud=False)

max_seq_length = 3072 # 数据集模型和打包的最大序列长度
trainer = SFTTrainer(
    model=model,
    args=args,
    train_dataset=dataset,
    peft_config=peft_config,
    max_seq_length=max_seq_length,
    tokenizer=tokenizer,
    callbacks=[swanlab_callback],
    packing=True,
    dataset_kwargs={
        "add_special_tokens": False,  # 我们使用特殊 tokens
        "append_concat_token": False, # 不需要添加额外的分隔符 token
    })
trainer.train()
trainer.save_model()