LoRA微调7B模型全过程：数据准备、训练配置、loss曲线与推理效果对比

{"instruction": "标准版套餐可以开发票吗？", "input": "", "output": "标准版套餐不支持开具发票。专业版及以上套餐支持开票，请在「订单中心-发票管理」中提交申请，电子发票一般1个工作日内开出。"}

from datasets import load_dataset
from transformers import AutoTokenizer

model_path = "Qwen/Qwen2.5-7B-Instruct"
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)

dataset = load_dataset("json", data_files="data/train.json", split="train")
dataset = dataset.train_test_split(test_size=0.1, seed=42)

def format_sample(ex):
    messages = [
        {"role": "system", "content": "你是XX公司的客服助手，回答需严格依据公司业务规则。"},
        {"role": "user", "content": ex["instruction"] + ("\n" + ex["input"] if ex["input"] else "")},
        {"role": "assistant", "content": ex["output"]},
    ]
    text = tokenizer.apply_chat_template(messages, tokenize=False)
    return {"text": text}

train_ds = dataset["train"].map(format_sample, remove_columns=dataset["train"].column_names)
eval_ds = dataset["test"].map(format_sample, remove_columns=dataset["test"].column_names)
print(train_ds[0]["text"][:300])

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

model = AutoModelForCausalLM.from_pretrained(
    model_path,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)
model.gradient_checkpointing_enable()
model.enable_input_require_grads()  # gradient checkpointing + LoRA 必须

lora_config = LoraConfig(
    r=16,
    lora_alpha=32,
    lora_dropout=0.05,
    bias="none",
    task_type="CAUSAL_LM",
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj",
                    "gate_proj", "up_proj", "down_proj"],
)
model = get_peft_model(model, lora_config)
model.print_trainable_parameters()
# trainable params: 20,185,088 || all params: 7,635,689,472 || trainable%: 0.2644

args = TrainingArguments(
    output_dir="./output_lora",
    num_train_epochs=3,
    per_device_train_batch_size=2,
    per_device_eval_batch_size=2,
    gradient_accumulation_steps=8,
    learning_rate=2e-4,
    lr_scheduler_type="cosine",
    warmup_ratio=0.03,
    logging_steps=10,
    eval_strategy="steps",
    eval_steps=100,
    save_strategy="steps",
    save_steps=200,
    save_total_limit=2,
    bf16=True,
    gradient_checkpointing=True,
    optim="adamw_torch",
    report_to="none",
    max_grad_norm=1.0,
)

trainer = SFTTrainer(
    model=model,
    args=args,
    train_dataset=train_ds,
    eval_dataset=eval_ds,
    dataset_text_field="text",
    max_seq_length=1024,
    tokenizer=tokenizer,
    packing=False,
)
trainer.train()
trainer.save_model("./output_lora/final")

from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained(model_path, torch_dtype=torch.bfloat16, device_map="auto")
model = PeftModel.from_pretrained(base, "./output_lora/final")
model = model.merge_and_unload()  # 合并后推理更快，无额外延迟
model.eval()

def chat(q):
    msgs = [{"role": "system", "content": "你是XX公司的客服助手，回答需严格依据公司业务规则。"},
            {"role": "user", "content": q}]
    text = tokenizer.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
    ids = tokenizer(text, return_tensors="pt").to(model.device)
    out = model.generate(**ids, max_new_tokens=256, temperature=0.1, do_sample=False)
    return tokenizer.decode(out[0][ids["input_ids"].shape[1]:], skip_special_tokens=True)

print(chat("标准版套餐可以开发票吗？"))