LoRA微调7B大模型全流程：从数据准备到推理效果对比（QLoRA实测）

pip install torch==2.3.1 transformers==4.44.2 peft==0.12.0 bitsandbytes==0.43.3 trl==0.9.6 accelerate==0.33.0 datasets==2.20.0

from datasets import load_dataset
import json

def format_data(example):
    messages = [
        {"role": "user", "content": example["question"]},
        {"role": "assistant", "content": example["answer"]}
    ]
    text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=False)
    return {"text": text}

dataset = load_dataset("json", data_files="medical_qa.jsonl", split="train")
dataset = dataset.map(format_data, remove_columns=dataset.column_names)
dataset = dataset.train_test_split(test_size=0.05, seed=42)

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training
from trl import SFTTrainer, SFTConfig
from datasets import load_dataset

model_name = "Qwen/Qwen2.5-7B-Instruct"

# 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,
)

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

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    quantization_config=bnb_config,
    device_map="auto",
    trust_remote_code=True,
)
model = prepare_model_for_kbit_training(model)

# LoRA配置
lora_config = LoraConfig(
    r=16,
    lora_alpha=32,
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj",
                    "gate_proj", "up_proj", "down_proj"],
    lora_dropout=0.05,
    bias="none",
    task_type="CAUSAL_LM",
)
model = get_peft_model(model, lora_config)
model.print_trainable_parameters()
# 输出: trainable params: 20,185,088 || all params: 7,635,000,000 || trainable%: 0.264

# 数据加载与格式化
dataset = load_dataset("json", data_files="medical_qa.jsonl", split="train")
def format_data(example):
    messages = [
        {"role": "user", "content": example["question"]},
        {"role": "assistant", "content": example["answer"]}
    ]
    return {"text": tokenizer.apply_chat_template(messages, tokenize=False)}
dataset = dataset.map(format_data, remove_columns=dataset.column_names)
split = dataset.train_test_split(test_size=0.05, seed=42)

# 训练配置
training_args = SFTConfig(
    output_dir="./qwen2.5-7b-lora-medical",
    per_device_train_batch_size=4,
    gradient_accumulation_steps=4,
    num_train_epochs=3,
    learning_rate=2e-4,
    lr_scheduler_type="cosine",
    warmup_ratio=0.03,
    logging_steps=10,
    save_strategy="epoch",
    bf16=True,
    optim="paged_adamw_8bit",
    gradient_checkpointing=True,
    max_seq_length=1024,
    dataset_text_field="text",
    report_to="none",
)

trainer = SFTTrainer(
    model=model,
    args=training_args,
    train_dataset=split["train"],
    eval_dataset=split["test"],
    tokenizer=tokenizer,
)

trainer.train()
trainer.save_model("./qwen2.5-7b-lora-medical/final")

from peft import PeftModel
from transformers import AutoModelForCausalLM

base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-7B-Instruct",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
model = PeftModel.from_pretrained(base_model, "./qwen2.5-7b-lora-medical/final")
model = model.merge_and_unload()
model.save_pretrained("./qwen2.5-7b-medical-merged")