Qwen2 大模型微调实战：基于指令微调的文本分类

4. 配置训练可视化工具

我们使用 SwanLab 来监控整个训练过程，并评估最终的模型效果。

这里直接使用 SwanLab 和 Transformers 的集成来实现：

from swanlab.integration.huggingface import SwanLabCallback
from transformers import Trainer

swanlab_callback = SwanLabCallback(...)

trainer = Trainer(
    ...
    callbacks=[swanlab_callback],
)

5. 完整代码与训练

开始训练时的目录结构：

|--- train.py
|--- train.jsonl
|--- test.jsonl

train.py:

import json
import pandas as pd
import torch
from datasets import Dataset
from modelscope import snapshot_download, AutoTokenizer
from swanlab.integration.huggingface import SwanLabCallback
from peft import LoraConfig, TaskType, get_peft_model
from transformers import AutoModelForCausalLM, TrainingArguments, Trainer, DataCollatorForSeq2Seq
import os
import swanlab

def dataset_jsonl_transfer(origin_path, new_path):
    """
    将原始数据集转换为大模型微调所需数据格式的新数据集
    """
    messages = []
    # 读取旧的 JSONL 文件
    with open(origin_path, "r") as file:
        for line in file:
            # 解析每一行的 json 数据
            data = json.loads(line)
            context = data["text"]
            catagory = data["category"]
            label = data["output"]
            message = {
                "instruction": "你是一个文本分类领域的专家，你会接收到一段文本和几个潜在的分类选项，请输出文本内容的正确类型",
                "input": f"文本:{context},类型选型:{catagory}",
                "output": label,
            }
            messages.append(message)
    # 保存重构后的 JSONL 文件
    with open(new_path, "w", encoding="utf-8") as file:
        for message in messages:
            file.write(json.dumps(message, ensure_ascii=False) + "\n")

def process_func(example):
    """
    将数据集进行预处理
    """
    MAX_LENGTH = 384
    input_ids, attention_mask, labels = [], [], []
    instruction = tokenizer(
        f"<|im_start|>system\n你是一个文本分类领域的专家，你会接收到一段文本和几个潜在的分类选项，请输出文本内容的正确类型<|im_end|>\n<|im_start|>user\n{example['input']}<|im_end|>\n<|im_start|>assistant\n",
        add_special_tokens=False,
    )
    response = tokenizer(f"{example['output']}", add_special_tokens=False)
    input_ids = instruction["input_ids"] + response["input_ids"] + [tokenizer.pad_token_id]
    attention_mask = (
        instruction["attention_mask"] + response["attention_mask"] + [1]
    )
    labels = [-100] * len(instruction["input_ids"]) + response["input_ids"] + [tokenizer.pad_token_id]
    if len(input_ids) > MAX_LENGTH:  # 做一个截断
        input_ids = input_ids[:MAX_LENGTH]
        attention_mask = attention_mask[:MAX_LENGTH]
        labels = labels[:MAX_LENGTH]
    return {"input_ids": input_ids, "attention_mask": attention_mask, "labels": labels}

def predict(messages, model, tokenizer):
    device = "cuda"
    text = tokenizer.apply_chat_template(
        messages,
        tokenize=False,
        add_generation_prompt=True
    )
    model_inputs = tokenizer([text], return_tensors="pt").to(device)
    generated_ids = model.generate(
        model_inputs.input_ids,
        max_new_tokens=512
    )
    generated_ids = [
        output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
    ]
    response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
    print(response)
    return response

# 在 modelscope 上下载 Qwen 模型到本地目录下
model_dir = snapshot_download("qwen/Qwen2-1.5B-Instruct", cache_dir="./", revision="master")

# Transformers 加载模型权重
tokenizer = AutoTokenizer.from_pretrained("./qwen/Qwen2-1.5B-Instruct/", use_fast=False, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("./qwen/Qwen2-1.5B-Instruct/", device_map="auto", torch_dtype=torch.bfloat16)
model.enable_input_require_grads()  # 开启梯度检查点时，要执行该方法

# 加载、处理数据集和测试集
train_dataset_path = "train.jsonl"
test_dataset_path = "test.jsonl"

train_jsonl_new_path = "new_train.jsonl"
test_jsonl_new_path = "new_test.jsonl"

if not os.path.exists(train_jsonl_new_path):
    dataset_jsonl_transfer(train_dataset_path, train_jsonl_new_path)
if not os.path.exists(test_jsonl_new_path):
    dataset_jsonl_transfer(test_dataset_path, test_jsonl_new_path)

# 得到训练集
train_df = pd.read_json(train_jsonl_new_path, lines=True)
train_ds = Dataset.from_pandas(train_df)
train_dataset = train_ds.map(process_func, remove_columns=train_ds.column_names)

config = LoraConfig(
    task_type=TaskType.CAUSAL_LM,
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
    inference_mode=False,  # 训练模式
    r=8,  # Lora 秩
    lora_alpha=32,  # Lora alaph，具体作用参见 Lora 原理
    lora_dropout=0.1,  # Dropout 比例
)

model = get_peft_model(model, config)

args = TrainingArguments(
    output_dir="./output/Qwen2",
    per_device_train_batch_size=4,
    gradient_accumulation_steps=4,
    logging_steps=10,
    num_train_epochs=2,
    save_steps=100,
    learning_rate=1e-4,
    save_on_each_node=True,
    gradient_checkpointing=True,
    report_to="none",
)

swanlab_callback = SwanLabCallback(
    project="Qwen2-fintune",
    experiment_name="Qwen2-1.5B-Instruct",
    description="使用通义千问 Qwen2-1.5B-Instruct 模型在 zh_cls_fudan-news 数据集上微调。",
    config={
        "model": "qwen/Qwen2-1.5B-Instruct",
        "dataset": "huangjintao/zh_cls_fudan-news",
    }
)

trainer = Trainer(
    model=model,
    args=args,
    train_dataset=train_dataset,
    data_collator=DataCollatorForSeq2Seq(tokenizer=tokenizer, padding=True),
    callbacks=[swanlab_callback],
)

trainer.train()

# 用测试集的前 10 条，测试模型
test_df = pd.read_json(test_jsonl_new_path, lines=True)[:10]

test_text_list = []
for index, row in test_df.iterrows():
    instruction = row['instruction']
    input_value = row['input']

    messages = [
        {"role": "system", "content": f"{instruction}"},
        {"role": "user", "content": f"{input_value}"
    ]

    response = predict(messages, model, tokenizer)
    messages.append({"role": "assistant", "content": f"{response}"})
    result_text = f"{messages[0]}\n\n{messages[1]}\n\n{messages[2]}"
    test_text_list.append(swanlab.Text(result_text, caption=response))

swanlab.log({"Prediction": test_text_list})
swanlab.finish()

看到下面的进度条即代表训练开始。

6. 训练结果演示

在 SwanLab 上查看最终的训练结果：

可以看到在 2 个 epoch 之后，微调后的 qwen2 的 loss 降低到了不错的水平——当然对于大模型来说，真正的效果评估还得看主观效果。

可以看到在一些测试样例上，微调后的 qwen2 能够给出准确的文本类型。

至此，你已经完成了 qwen2 指令微调的训练！

7. 推理训练好的模型

训好的模型默认被保存在 ./output/Qwen2 文件夹下。

推理模型的代码如下：

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

def predict(messages, model, tokenizer):
    device = "cuda"
    text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    model_inputs = tokenizer([text], return_tensors="pt").to(device)
    generated_ids = model.generate(model_inputs.input_ids, max_new_tokens=512)
    generated_ids = [output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)]
    response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
    return response

# 加载原下载路径的 tokenizer
tokenizer = AutoTokenizer.from_pretrained("./qwen/Qwen2-1.5B-Instruct/", use_fast=False, trust_remote_code=True)

# 加载训练好的模型路径的 model，将下面的 checkpointXXX 替换为实际的 checkpoint 文件名名称
model = AutoModelForCausalLM.from_pretrained("./output/Qwen2/checkpointXXX", device_map="auto", torch_dtype=torch.bfloat16)

test_texts = {
    'instruction': "你是一个文本分类领域的专家，你会接收到一段文本和几个潜在的分类选项，请输出文本内容的正确类型",
    'input': "文本：航空动力学报 JOURNAL OF AEROSPACE POWER1998 年 第 4 期 No.4 1998 科技期刊管路系统敷设的并行工程模型研究*陈志英* * 马 枚北京航空航天大学【摘要】提出了一种应用于并行工程模型转换研究的标号法，该法是将现行串行设计过程 (As-is) 转换为并行设计过程 (To-be)。本文应用该法将发动机外部管路系统敷设过程模型进行了串并行转换，应用并行工程过程重构的手段，得到了管路敷设并行过程模型。"
}

instruction = test_texts['instruction']
input_value = test_texts['input']

messages = [
    {"role": "system", "content": f"{instruction}"},
    {"role": "user", "content": f"{input_value}"}
]

response = predict(messages, model, tokenizer)
print(response)