大模型微调入门（Transformers + Pytorch）

目标

输入：你是谁？

输出：我们预训练的名字。

训练

为了性能好下载小参数模型，普通机器都能运行。

下载模型

python 复制代码

# 方式1：使用魔搭社区SDK 下载
# down_deepseek.py
from modelscope import snapshot_download
model_dir = snapshot_download('deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B')

# 方式2：git lfs 
# 需要提前安装git大文件存储 git-lfs
# 在线查看 https://www.modelscope.cn/models/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B
git lfs install
git clone https://www.modelscope.cn/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B.git

训练模型

python 复制代码

# finetune_deepseek.py
from datasets import Dataset
from transformers import (
    AutoModelForCausalLM,
    AutoTokenizer,
    TrainingArguments,
    Trainer,
    DataCollatorForLanguageModeling
)

# 加载模型和分词器
model_name = "deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B"
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=True)

# 准备训练数据
train_data = [
    {
        "question": "你是谁?",
        "answer": "我是黄登峰。"
    },
    {
        "question": "你的名字是什么？",
        "answer": "黄登峰"
    },
    {
        "question": "你是做什么的？",
        "answer": "我是深圳一家公司打工的牛马程序员。"
    },
    # 在这里添加更多的问答对
]

test_data = [
    {
        "question": "你的名字是什么?",
        "answer": "我的名字是黄登峰。"
    }
]
def format_instruction(example):
    """格式化输入输出对"""
    return f"Human: {example['question']}\n\nAssistant: {example['answer']}"

# 转换数据格式
train_formatted_data = [{"text": format_instruction(item)} for item in train_data]
test_formatted_data = [{"text": format_instruction(item)} for item in test_data]
train_dataset = Dataset.from_list(train_formatted_data)
test_dataset = Dataset.from_list(test_formatted_data)

# 数据预处理函数
def preprocess_function(examples):
    return tokenizer(examples["text"], truncation=True, padding="max_length", max_length=512)

# 对数据集进行预处理
train_tokenized_dataset = train_dataset.map(
    preprocess_function,
    batched=True,
    remove_columns=train_dataset.column_names
)

test_tokenized_dataset = test_dataset.map(
    preprocess_function,
    batched=True,
    remove_columns=test_dataset.column_names
)
output_dir = "deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B_CUSTOM"

# 训练参数设置
training_args = TrainingArguments(
    output_dir=output_dir,
    num_train_epochs=3,
    per_device_train_batch_size=4,
    save_steps=100,
    save_total_limit=2,
    learning_rate=2e-5,
    weight_decay=0.01,
    logging_dir="./logs",
    logging_steps=10,
)

# 创建训练器
trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=train_tokenized_dataset,
    eval_dataset=test_tokenized_dataset,
    data_collator=DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False),
)

# 开始训练
trainer.train()

# 保存模型
trainer.save_model()
# 保存tokenizer
tokenizer.save_pretrained(output_dir)

模型格式

训练后的模型输出格式是Hugging Face格式，vllm 可以直接使用，ollama,llama.cpp默认是GGUF格式。

bash 复制代码

# 需要用llama.cpp仓库的convert_hf_to_gguf.py脚本来转换
git clone https://github.com/ggerganov/llama.cpp.git
pip install -r llama.cpp/requirements.txt
# 如果不量化，保留模型的效果
python llama.cpp/convert_hf_to_gguf.py ./DeepSeek-R1-Distill-Qwen-1.5B  --outtype f16 --verbose --outfile DeepSeek-R1-Distill-Qwen-1.5B.gguf
# 如果需要量化（加速并有损效果），直接执行下面脚本就可以
python llama.cpp/convert_hf_to_gguf.py ./DeepSeek-R1-Distill-Qwen-1.5B  --outtype q8_0 --verbose --outfile DeepSeek-R1-Distill-Qwen-1.5B.gguf

验证

python 复制代码

# test_model.py
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

def generate_response(prompt, model, tokenizer, max_length=512):
    # 将输入格式化为训练时的格式
    formatted_prompt = f"Human: {prompt}\n\nAssistant:"
    
    # 对输入进行编码
    inputs = tokenizer(formatted_prompt, return_tensors="pt", padding=True, truncation=True)
    
    # 生成回答
    with torch.no_grad():
        outputs = model.generate(
            inputs.input_ids,
            max_length=max_length,
            num_return_sequences=1,
            temperature=0.7,
            do_sample=True,
            pad_token_id=tokenizer.pad_token_id,
            eos_token_id=tokenizer.eos_token_id,
        )
    
    # 解码输出
    response = tokenizer.decode(outputs[0], skip_special_tokens=True)
    
    # 提取Assistant的回答部分
    response = response.split("Assistant:")[-1].strip()
    return response

def main():
    # 加载微调后的模型和分词器
    model_path = "deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B_CUSTOM"
    tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
    model = AutoModelForCausalLM.from_pretrained(model_path, trust_remote_code=True)
    
    # 准备测试问题
    test_questions = [
        "你是谁?",
        "你的名字是什么?",
        "你是做什么的?",
    ]
    
    # 测试模型回答
    print("开始测试模型回答：")
    print("-" * 50)
    for question in test_questions:
        print(f"问题: {question}")
        response = generate_response(question, model, tokenizer)
        print(f"回答: {response}")
        print("-" * 50)

if __name__ == "__main__":
    main()