LoRA微调7B大模型：从数据到推理，loss降了68%的完整记录

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

import json
import pandas as pd

df = pd.read_csv("fault_qa.csv")  # 3800行
data = []
for _, row in df.iterrows():
    item = {
        "instruction": "你是一个工业设备故障诊断专家，请根据用户描述的报警代码，给出可能原因和处理步骤。",
        "input": row["question"],
        "output": row["answer"]
    }
    data.append(item)

# 划分训练集和验证集，9:1
split = int(len(data) * 0.9)
train_data = data[:split]
val_data = data[split:]

with open("train.jsonl", "w", encoding="utf-8") as f:
    for item in train_data:
        f.write(json.dumps(item, ensure_ascii=False) + "\n")

with open("val.jsonl", "w", encoding="utf-8") as f:
    for item in val_data:
        f.write(json.dumps(item, ensure_ascii=False) + "\n")

print(f"训练集：{len(train_data)}条，验证集：{len(val_data)}条")
# 输出：训练集：3420条，验证集：380条

import torch
from datasets import load_dataset
from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments
from peft import LoraConfig, get_peft_model
from trl import SFTTrainer

# 1. 加载模型和tokenizer
model_name = "Qwen/Qwen2.5-7B-Instruct"
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
tokenizer.pad_token = tokenizer.eos_token

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)

# 2. LoRA配置
lora_config = LoraConfig(
    r=16,
    lora_alpha=32,
    lora_dropout=0.05,
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj",
                    "gate_proj", "up_proj", "down_proj"],
    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: 7,635,801,600 || trainable%: 0.2643

# 3. 数据格式化
def format_data(example):
    text = f"### Instruction:\n{example['instruction']}\n\n### Input:\n{example['input']}\n\n### Output:\n{example['output']}"
    return {"text": text}

train_dataset = load_dataset("json", data_files="train.jsonl", split="train")
val_dataset = load_dataset("json", data_files="val.jsonl", split="train")
train_dataset = train_dataset.map(format_data)
val_dataset = val_dataset.map(format_data)

# 4. 训练参数
training_args = TrainingArguments(
    output_dir="./qwen2.5-7b-lora-fault",
    num_train_epochs=3,
    per_device_train_batch_size=4,
    gradient_accumulation_steps=4,
    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="adamw_torch",
    report_to="none",
    save_total_limit=2,
    load_best_model_at_end=True,
    metric_for_best_model="eval_loss"
)

# 5. 训练
trainer = SFTTrainer(
    model=model,
    args=training_args,
    train_dataset=train_dataset,
    eval_dataset=val_dataset,
    tokenizer=tokenizer,
    dataset_text_field="text",
    max_seq_length=512,
    packing=False
)

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

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

base_model = "Qwen/Qwen2.5-7B-Instruct"
lora_path = "./qwen2.5-7b-lora-fault/final"

tokenizer = AutoTokenizer.from_pretrained(base_model, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    base_model, torch_dtype=torch.bfloat16, device_map="auto", trust_remote_code=True
)
model = PeftModel.from_pretrained(model, lora_path)
model.eval()

prompt = "### Instruction:\n你是一个工业设备故障诊断专家，请根据用户描述的报警代码，给出可能原因和处理步骤。\n\n### Input:\nE104报警怎么处理？\n\n### Output:\n"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=256, temperature=0.1, do_sample=True)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))