7B模型LoRA微调全流程：数据、训练、Loss与推理对比

原始数据（PDF/文本） → 清洗 → JSONL（instruction + output） → 分词 → 训练（QLoRA） → 保存adapter → 推理对比

{
  "instruction": "请回答：阿托伐他汀钙片的主要适应症是什么？",
  "output": "阿托伐他汀钙片主要用于高胆固醇血症和冠心病的治疗，通过抑制HMG-CoA还原酶降低血脂。"
}

import json
from datasets import Dataset

def load_medical_data(file_path):
    data = []
    with open(file_path, 'r', encoding='utf-8') as f:
        for line in f:
            item = json.loads(line)
            # 格式化为chat模板：user\n...\nassistant\n...
            text = f"user\n{item['instruction']}\nassistant\n{item['output']}"
            data.append({"text": text})
    return Dataset.from_list(data)

dataset = load_medical_data("medical_qa.jsonl")
dataset = dataset.train_test_split(test_size=0.05)
print(f"训练集: {len(dataset['train'])} 条, 验证集: {len(dataset['test'])} 条")
# 输出：训练集: 4750 条, 验证集: 250 条

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

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

model_name = "Qwen/Qwen-7B-Chat"
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
tokenizer.pad_token = tokenizer.eos_token  # 重要：设置pad_token

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

# 准备k-bit训练（必要的梯度检查点设置）
model = prepare_model_for_kbit_training(model)

# LoRA配置
lora_config = LoraConfig(
    r=8,
    lora_alpha=16,
    target_modules=["q_proj", "v_proj"],  # 只微调这两个模块
    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: 4,194,304 || all params: 6,738,415,616 || trainable%: 0.0622

from transformers import TrainingArguments, Trainer, DataCollatorForSeq2Seq

training_args = TrainingArguments(
    output_dir="./qwen-medical-lora",
    per_device_train_batch_size=4,
    gradient_accumulation_steps=2,  # 等效batch size = 8
    num_train_epochs=3,
    learning_rate=2e-4,
    fp16=True,
    logging_steps=10,
    save_steps=200,
    eval_steps=200,
    evaluation_strategy="steps",
    save_total_limit=2,
    remove_unused_columns=False,
    report_to="tensorboard",
    load_best_model_at_end=True,
    metric_for_best_model="loss",
)

data_collator = DataCollatorForSeq2Seq(tokenizer, model=model, padding=True)

trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=dataset["train"],
    eval_dataset=dataset["test"],
    data_collator=data_collator,
    tokenizer=tokenizer,
)

trainer.train()

from peft import PeftModel
base_model = AutoModelForCausalLM.from_pretrained(...)
model = PeftModel.from_pretrained(base_model, "./qwen-medical-lora/checkpoint-1200")
model = model.merge_and_unload()  # 合并权重