7B模型LoRA微调实录：从数据清洗到loss收敛的完整链路

Python 3.10.12
torch 2.1.2+cu121
transformers 4.36.2
peft 0.7.1
bitsandbytes 0.41.3
accelerate 0.25.0
datasets 2.16.1
trl 0.7.10

conda create -n lora_finetune python=3.10 -y
conda activate lora_finetune
pip install torch==2.1.2 --index-url https://download.pytorch.org/whl/cu121
pip install transformers==4.36.2 peft==0.7.1 bitsandbytes==0.41.3 accelerate==0.25.0 datasets==2.16.1 trl==0.7.10

system
你是一个专业的医疗助手。请根据用户描述提供症状分析、可能诊断和就医建议。
user
{用户问题}
assistant
{结构化答案}

from datasets import Dataset, load_dataset
import json

# 假设数据是jsonl格式，每行一个对话
def load_and_format(data_path):
    with open(data_path, 'r', encoding='utf-8') as f:
        lines = [json.loads(line) for line in f]

    # 格式化函数
    def format_example(example):
        system_msg = "你是一个专业的医疗助手。请根据用户描述提供症状分析、可能诊断和就医建议。"
        user_msg = example['question']
        assistant_msg = json.dumps(example['answer'], ensure_ascii=False)

        # ChatML格式
        text = f"system\n{system_msg}\nuser\n{user_msg}\nassistant\n{assistant_msg}"
        return {'text': text}

    dataset = Dataset.from_list(lines)
    dataset = dataset.map(format_example, remove_columns=dataset.column_names)
    return dataset

train_dataset = load_and_format('medical_qa_30w.jsonl')
print(f"训练数据量: {len(train_dataset)}")  # 输出: 300000

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

model_name = "Qwen/Qwen2-7B"

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=True  # 双重量化，进一步节省显存
)

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

tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
tokenizer.pad_token = tokenizer.eos_token  # Qwen没有pad token，用eos代替

# 准备k-bit训练：冻结所有参数，仅保留lm_head可训练（实际不需要，但保留以防万一）
model = prepare_model_for_kbit_training(model)

lora_config = LoraConfig(
    r=16,                     # 秩，实验发现8-32之间效果差异不大，16是平衡点
    lora_alpha=32,            # 缩放因子，通常设置为r的2倍
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
    lora_dropout=0.1,         # dropout防止过拟合，小数据集建议0.1
    bias="none",
    task_type="CAUSAL_LM",
    modules_to_save=["lm_head"]  # 保留lm_head可训练，稳定输出分布
)

model = get_peft_model(model, lora_config)
model.print_trainable_parameters()
# 输出: trainable params: 7,077,888 || all params: 7,073,686,528 || trainable%: 0.1000

from transformers import TrainingArguments, Trainer
from trl import DataCollatorForCompletionOnlyLM

# 数据整理器：只计算assistant部分的loss，忽略system和user
response_template = "assistant\n"
collator = DataCollatorForCompletionOnlyLM(
    response_template=response_template,
    tokenizer=tokenizer,
    mlm=False
)

training_args = TrainingArguments(
    output_dir="./qwen2-lora-medical",
    num_train_epochs=3,
    per_device_train_batch_size=8,    # 4090 24G刚好
    gradient_accumulation_steps=4,    # 等效batch size = 8*4 = 32
    learning_rate=2e-4,               # LoRA通常用1e-4到5e-4，比全参数微调大
    lr_scheduler_type="cosine",
    warmup_ratio=0.03,
    logging_steps=10,
    save_steps=500,
    evaluation_strategy="steps",
    eval_steps=500,
    save_total_limit=2,
    fp16=True,                        # 混合精度训练
    gradient_checkpointing=True,      # 梯度检查点，节省显存的关键
    report_to="tensorboard",
    dataloader_num_workers=4,
)

trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=train_dataset,
    eval_dataset=eval_dataset,  # 另外准备的验证集，约5000条
    data_collator=collator,
    tokenizer=tokenizer,
)

trainer.train()

根据您的描述，头晕和恶心可能与多种因素有关，包括低血糖、颈椎问题、内耳疾病等。建议您去医院做进一步检查。

{
  "症状分析": "患者主诉头晕伴恶心，血压正常，排除高血压相关头晕。可能原因包括：1. 颈椎病（颈性眩晕）——长时间低头工作常见；2. 前庭系统疾病——如耳石症、梅尼埃病；3. 低血糖或贫血——需结合血糖和血常规检查。",
  "可能诊断": "1. 颈性眩晕（可能性较大，占60%）；2. 耳石症（20%）；3. 其他（20%）。",
  "就医建议": "1. 建议挂神经内科或耳鼻喉科；2. 做颈椎MRI和耳石诱发试验；3. 注意休息，避免突然转头。"
}