LoRA微调7B模型实战：从数据准备到推理效果对比

Python 3.10.12
torch 2.0.1+cu118
transformers 4.31.0
peft 0.5.0
bitsandbytes 0.41.0
accelerate 0.23.0
datasets 2.14.0

{
  "instruction": "请根据以下医学知识回答问题",
  "input": "患者出现持续性胸痛，可能是什么疾病？",
  "output": "持续性胸痛可能由心绞痛、心肌梗死、肺栓塞等引起..."
}

from datasets import Dataset
import json

def preprocess_data(file_path):
    with open(file_path, 'r', encoding='utf-8') as f:
        raw_data = json.load(f)

    processed = []
    for item in raw_data:
        text = f"### 指令：{item['instruction']}\n### 输入：{item['input']}\n### 输出：{item['output']}"
        processed.append({"text": text})

    dataset = Dataset.from_list(processed)
    return dataset

# 分词处理
def tokenize_function(examples, tokenizer, max_length=512):
    tokenized = tokenizer(
        examples["text"],
        truncation=True,
        padding="max_length",
        max_length=max_length
    )
    tokenized["labels"] = tokenized["input_ids"].copy()
    return tokenized

# 使用方式
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-hf")
tokenizer.pad_token = tokenizer.eos_token
dataset = preprocess_data("medical_qa.json")
dataset = dataset.map(lambda x: tokenize_function(x, tokenizer), batched=True)

import torch
from transformers import AutoModelForCausalLM, BitsAndBytesConfig, TrainingArguments
from peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training
from trl import SFTTrainer

# 4-bit量化配置
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=True
)

# 加载基座模型
base_model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-2-7b-hf",
    quantization_config=bnb_config,
    device_map="auto",
    trust_remote_code=True
)
base_model = prepare_model_for_kbit_training(base_model)

# LoRA配置
lora_config = LoraConfig(
    r=8,
    lora_alpha=16,
    target_modules=["q_proj", "v_proj", "k_proj", "o_proj", 
                   "gate_proj", "up_proj", "down_proj"],
    lora_dropout=0.05,
    bias="none",
    task_type="CAUSAL_LM"
)

model = get_peft_model(base_model, lora_config)
model.print_trainable_parameters()  # 输出：trainable params: 4,194,304 || all params: 6,745,776,128 || trainable%: 0.0622

# 训练参数
training_args = TrainingArguments(
    output_dir="./llama2-lora-medical",
    per_device_train_batch_size=2,
    gradient_accumulation_steps=8,
    num_train_epochs=3,
    learning_rate=2e-4,
    fp16=True,
    logging_steps=10,
    save_steps=500,
    gradient_checkpointing=True,
    optim="paged_adamw_8bit",
    lr_scheduler_type="cosine",
    warmup_ratio=0.03,
    report_to="tensorboard"
)

trainer = SFTTrainer(
    model=model,
    args=training_args,
    train_dataset=dataset,
    tokenizer=tokenizer,
    max_seq_length=512,
    dataset_text_field="text"
)

trainer.train()