LLM微调（二）| 微调LLAMA-2和其他开源LLM的两种简单方法

本文将介绍两种开源工具来微调LLAMA-2。

一、使用autotrain-advanced微调LLAMA-2

        AutoTrain是一种无代码工具，用于为自然语言处理（NLP）任务、计算机视觉（CV）任务、语音任务甚至表格任务训练最先进的模型。

1） 安装相关库，使用huggingface_hub下载微调数据

!pip install autotrain-advanced!pip install huggingface_hub

2） 更新autotrain-advanced所需要的包

# update torch!autotrain setup --update-torch

3） 登录Huggingface

# Login to huggingface​from huggingface_hub import notebook_loginnotebook_login()

4） 开始微调LLAMA-2

! autotrain llm \--train \--model {MODEL_NAME} \--project-name {PROJECT_NAME} \--data-path data/ \--text-column text \--lr {LEARNING_RATE} \--batch-size {BATCH_SIZE} \--epochs {NUM_EPOCHS} \--block-size {BLOCK_SIZE} \--warmup-ratio {WARMUP_RATIO} \--lora-r {LORA_R} \--lora-alpha {LORA_ALPHA} \--lora-dropout {LORA_DROPOUT} \--weight-decay {WEIGHT_DECAY} \--gradient-accumulation {GRADIENT_ACCUMULATION}

核心参数含义：

llm: 微调模型的类型

— project_name: 项目名称

— model: 需要微调的基础模型

— data_path: 指定微调所需要的数据，可以使用huggingface上的数据集

— text_column: 如果数据是表格，需要指定instructions和responses对应的列名

— use_peft: 指定peft某一种方法

— use_int4: 指定int 4量化

— learning_rate: 学习率

— train_batch_size: 训练批次大小

— num_train_epochs: 训练轮数大小

— trainer: 指定训练的方式

— model_max_length: 设置模型最大上下文窗口

— push_to_hub（可选）: 微调好的模型是否需要存储到Hugging Face? 

— repo_id: 如果要存储微调好的模型到Hugging Face，需要指定repository ID

— block_size: 设置文本块大小

下面看一个具体的示例：

!autotrain llm--train--project_name "llama2-autotrain-openassitant"--model TinyPixel/Llama-2-7B-bf16-sharded--data_path timdettmers/openassistant-guanaco--text_column text--use_peft--use_int4--learning_rate 0.4--train_batch_size 3--num_train_epochs 2--trainer sft--model_max_length 1048--push_to_hub--repo_id trojrobert/llama2-autotrain-openassistant--block_size 1048 > training.log

二、使用TRL微调LLAMA-2

       TRL是一个全栈库，提供了通过强化学习来训练transformer语言模型一系列工具，包括从监督微调步骤（SFT）、奖励建模步骤（RM）到近端策略优化（PPO）步骤。

1）安装相关的库

!pip install -q -U trl peft transformers  datasets bitsandbytes wandb

2）从Huggingface导入数据集

from datasets import load_dataset​dataset_name = "timdettmers/openassistant-guanaco"​dataset = load_dataset(dataset_name, split="train")

3）量化配置，从Huggingface下载模型

import torchfrom transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig​​# quantizition configurationbnb_config = BitsAndBytesConfig(    load_in_4bit=True,    bnb_4bit_quant_type="nf4",    bnb_4bit_compute_dtype=torch.float16,)​​# download model​model_name = "TinyPixel/Llama-2-7B-bf16-sharded"model = AutoModelForCausalLM.from_pretrained(    model_name,    quantization_config=bnb_config,    trust_remote_code=True)model.config.use_cache = False

4）下载Tokenizer

tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)tokenizer.pad_token = tokenizer.eos_token

5）创建PEFT配置

from peft import LoraConfig, get_peft_model​lora_alpha = 16lora_dropout = 0.1lora_r = 64​peft_config = LoraConfig(    lora_alpha=lora_alpha,    lora_dropout=lora_dropout,    r=lora_r,    bias="none",    task_type="CAUSAL_LM")

6）创建微调和训练配置

from transformers import TrainingArguments​output_dir = "./results"per_device_train_batch_size = 4gradient_accumulation_steps = 4optim = "paged_adamw_32bit"save_steps = 100logging_steps = 10learning_rate = 2e-4max_grad_norm = 0.3max_steps = 100warmup_ratio = 0.03lr_scheduler_type = "constant"​training_arguments = TrainingArguments(    output_dir=output_dir,    per_device_train_batch_size=per_device_train_batch_size,    gradient_accumulation_steps=gradient_accumulation_steps,    optim=optim,    save_steps=save_steps,    logging_steps=logging_steps,    learning_rate=learning_rate,    fp16=True,    max_grad_norm=max_grad_norm,    max_steps=max_steps,    warmup_ratio=warmup_ratio,    group_by_length=True,    lr_scheduler_type=lr_scheduler_type,)

7）创建SFTTrainer配置

from trl import SFTTrainer​max_seq_length = 512​trainer = SFTTrainer(    model=model,    train_dataset=dataset,    peft_config=peft_config,    dataset_text_field="text",    max_seq_length=max_seq_length,    tokenizer=tokenizer,    args=training_arguments,)

8）在微调的时候，对LN层使用float 32训练更稳定

for name, module in trainer.model.named_modules():    if "norm" in name:        module = module.to(torch.float32)

9）开始微调

trainer.train()

10）保存微调好的模型

model_to_save = trainer.model.module if hasattr(trainer.model, 'module') else trainer.model  # Take care of distributed/parallel trainingmodel_to_save.save_pretrained("outputs")

11）加载微调好的模型

lora_config = LoraConfig.from_pretrained('outputs')tuned_model = get_peft_model(model, lora_config)

12）测试微调好的模型效果

​text = "What is a large language model?"device = "cuda:0"​inputs = tokenizer(text, return_tensors="pt").to(device)outputs = tuned_model.generate(**inputs, max_new_tokens=50)print(tokenizer.decode(outputs[0], skip_special_tokens=True))

参考文献：

[1] https://trojrobert.medium.com/4-easier-ways-for-fine-tuning-llama-2-and-other-open-source-llms-eb3218657f6e

[2] https://colab.research.google.com/drive/1JMEi2VMNGMOTyfEcQZyp23EISUrWg5cg?usp=sharing

[3] https://colab.research.google.com/drive/1ctevXhrE60s7o9RzsxpIqq37EjyU9tBn?usp=sharing#scrollTo=bsbdrb5p2ONa