LoRA微调7B模型全流程：数据构建、训练配置与Loss曲线分析

- GPU: NVIDIA RTX 4090 (24GB)
- CUDA: 12.1
- PyTorch: 2.3.1
- transformers: 4.44.2
- peft: 0.12.0
- bitsandbytes: 0.43.3
- accelerate: 0.34.2
- trl: 0.9.6
- datasets: 2.21.0

{
  "instruction": "变频器显示E07故障码，应该如何处理？",
  "input": "设备型号：ATV320，运行环境：粉尘车间",
  "output": "E07表示直流母线过压。处理步骤：1. 检查输入电压是否超过额定值10%；2. 确认减速时间参数dEC是否设置过短，建议从3.0s调整到5.0s；3. 若带大惯量负载，需加装制动电阻。若上述无效，测量母线电容容量，低于标称值80%需更换。"
}

from datasets import load_dataset

def format_prompt(example):
    if example["input"]:
        user_content = f"{example['instruction']}\n补充信息：{example['input']}"
    else:
        user_content = example["instruction"]

    messages = [
        {"role": "system", "content": "你是工业设备售后技术支持工程师，回答要具体、可操作，给出明确步骤和参数。"},
        {"role": "user", "content": user_content},
        {"role": "assistant", "content": example["output"]}
    ]
    text = tokenizer.apply_chat_template(messages, tokenize=False)
    return {"text": text}

dataset = load_dataset("json", data_files="train_data.jsonl", split="train")
dataset = dataset.map(format_prompt, remove_columns=dataset.column_names)

from peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
import torch

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=True,
)

model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-7B-Instruct",
    quantization_config=bnb_config,
    device_map="auto",
    trust_remote_code=True,
)
model = prepare_model_for_kbit_training(model)

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()
# trainable params: 40,370,176 || all params: 7,655,986,688 || trainable%: 0.5273

from transformers import TrainingArguments
from trl import SFTTrainer

training_args = TrainingArguments(
    output_dir="./qwen2.5-7b-lora-cs",
    num_train_epochs=3,
    per_device_train_batch_size=2,
    gradient_accumulation_steps=8,
    learning_rate=2e-4,
    lr_scheduler_type="cosine",
    warmup_ratio=0.03,
    logging_steps=10,
    save_strategy="epoch",
    evaluation_strategy="epoch",
    bf16=True,
    gradient_checkpointing=True,
    optim="paged_adamw_8bit",
    report_to="none",
)

trainer = SFTTrainer(
    model=model,
    args=training_args,
    train_dataset=train_dataset,
    eval_dataset=eval_dataset,
    tokenizer=tokenizer,
    dataset_text_field="text",
    max_seq_length=1024,
    packing=False,
)

trainer.train()

from peft import PeftModel

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-cs/checkpoint-1800")
model = model.merge_and_unload()
model.save_pretrained("./qwen2.5-7b-cs-merged")