【Qwen微调保姆级教程】Q-LoRA微调Qwen-7B-Chat-Int4，实现医疗场景的智能问答

Ref: Qwen官方文档

一、数据准备

本文使用的数据集是中文医疗问答数据集，节选其中内科7644条数据（sample_IM5000-6000.csv）。

原始数据示例：

我们需要将数据转换为 Qwen 适用的 json 数据，示例如下：

[
    {
        "id": "identity_0",
        "conversations": [
            {
                "from": "user",
                "value": "心血管科问诊，高血压患者能吃党参吗？,我有高血压这两天女婿来的时候给我拿了些党参泡水喝，您好高血压可以吃党参吗？"
            },
            {
                "from": "assistant",
                "value": "高血压病人可以口服党参的。党参有降血脂，降血压的作用，可以彻底消除血液中的垃圾，从而对冠心病以及心血管疾病的患者都有一定的稳定预防工作作用，因此平时口服党参能远离三高的危害。另外党参除了益气养血，降低中枢神经作用，调整消化系统功能，健脾补肺的功能。感谢您的进行咨询，期望我的解释对你有所帮助。"
            }
        ]
    }
]

具体转换代码（划分80%训练集+20%验证集）：

# coding=gb18030
import pandas as pd
import json
from sklearn.model_selection import train_test_split

# 读取CSV文件
csv_file_path = 'datasets/qwen/sample_IM5000-6000.csv'
df = pd.read_csv(csv_file_path, encoding='ANSI')

# 划分数据集
train_df, val_df = train_test_split(df, test_size=0.2, random_state=42)

# 转换训练集数据
train_data_list = []
for index, row in train_df.iterrows():
    department = row['department'] + '问诊'
    title = row['title']
    ask = row['ask']
    answer = row['answer']

    user_value = f"{department}，{title}{ask}"

    conversations = [
        {"from": "user", "value": user_value},
        {"from": "assistant", "value": answer}
    ]

    entry = {
        "id": f"identity_{index}",
        "conversations": conversations
    }

    train_data_list.append(entry)

# 将训练集数据序列化为JSON格式
train_json_data = json.dumps(train_data_list, indent=4, ensure_ascii=False)

# 写入训练集JSON文件
train_json_file_path = './datasets/qwen/train_sample_IM5000-6000.json'
with open(train_json_file_path, 'w', encoding='utf-8') as json_file:
    json_file.write(train_json_data)

# 转换验证集数据
val_data_list = []
for index, row in val_df.iterrows():
    department = row['department'] + '问诊'
    title = row['title']
    ask = row['ask']
    answer = row['answer']

    user_value = f"{department}，{title}{ask}"

    conversations = [
        {"from": "user", "value": user_value},
        {"from": "assistant", "value": answer}
    ]

    entry = {
        "id": f"identity_{index}",
        "conversations": conversations
    }

    val_data_list.append(entry)

# 将验证集数据序列化为JSON格式
val_json_data = json.dumps(val_data_list, indent=4, ensure_ascii=False)

# 写入验证集JSON文件
val_json_file_path = './datasets/qwen/val_sample_IM5000-6000.json'
with open(val_json_file_path, 'w', encoding='utf-8') as json_file:
    json_file.write(val_json_data)

print(f"Train data has been successfully converted to JSON format and saved to {train_json_file_path}")
print(f"Validation data has been successfully converted to JSON format and saved to {val_json_file_path}")

二、模型下载与微调训练

1. 下载源码

git clone https://github.com/QwenLM/Qwen.git

2. 配置环境

pip install -r requirements.txt

pip install mpi4py -i http://pypi.doubanio.com/simple/ --trusted-host pypi.doubanio.com
# 出现下列错误的话，用 conda 安装 mpi4py 
# ERROR: Could not build wheels for mpi4py, which is required to install pyproject.toml-based projects
sudo apt update
sudo apt-get install libopenmpi-dev
# 执行以上两条命令后， 重新下载安装mpi4py命令，发现还是安装失败，改用conda的方式安装
conda install mpi4py

pip install "peft<0.8.0" deepspeed
pip install optimum
pip install auto-gptq
pip install -i https://pypi.tuna.tsinghua.edu.cn/simple auto-gptq
pip install tiktoken

pip install modelscope

 3. 模型下载（无法科学上网，可以通过 modelscope 下载，需 pip install modelscope）

#模型下载
from modelscope import snapshot_download
model_dir = snapshot_download('Qwen/Qwen-7B-Chat-Int4', local_dir="/root/Qwen-7B-Chat-Int4")

4. 微调训练

修改微调训练脚本 finetune_qlora_single_gpu.sh 中的模型和数据路径，

MODEL="/root/Qwen-7B-Chat-Int4"
DATA="/root/datasets/qwen/val_sample_IM5000-6000.json"

运行微调训练的脚本 。

bash finetune/finetune_qlora_single_gpu.sh

5. 推理

# 推理
from transformers import AutoTokenizer
from peft import AutoPeftModelForCausalLM
import json

# 加载预训练的分词器
tokenizer = AutoTokenizer.from_pretrained("/root/Qwen-7B-Chat-Int4")  # 这里需要替换成实际的Qwen模型ID

# 加载微调后的模型
model = AutoPeftModelForCausalLM.from_pretrained(
    "output_qwen/test1",  # 微调模型的路径
    device_map="auto",
    trust_remote_code=True
).eval()

# 假设你的JSON数据存储在一个名为data.json的文件中
with open('datasets/qwen/test.json', 'r', encoding='utf-8') as file:
    test_set = json.load(file)

for dialog in test_set:
    conversations = dialog["conversations"]
    history = None
    for i, conv in enumerate(conversations):
        if conv["from"] == "user":
            # 使用模型生成回答
            response, history = model.chat(tokenizer, conv["value"], history=history)
            # 将回答存储到下一个assistant字段
            if i + 1 < len(conversations) and conversations[i + 1]["from"] == "assistant":
                conversations[i + 1]["value"] = response

# 打印更新后的测试集
for dialog in test_set:
    print(dialog)
# 将更新后的数据集保存到新的JSON文件中
with open('updated_data.json', 'w', encoding='utf-8') as file:
    json.dump(test_set, file, ensure_ascii=False, indent=4)

print("数据已成功更新并保存到 updated_data.json 文件。")