Skill蒸馏完整指南：可复制的实战案例

pip install openai transformers datasets torch
export OPENAI_API_KEY="sk-xxxx"  # 你的API Key

from openai import OpenAI
import json

client = OpenAI()

# 定义发布WordPress文章的Prompt模板
PROMPT_TEMPLATE = """你是一个WordPress运营专家。用户给你文章信息，你需要输出完整的发布流程。

文章信息：
- 标题：{title}
- 正文：{content}
- 分类：{category}
- 标签：{tags}

请按以下格式输出发布步骤（只输出步骤，不要解释）：

步骤1: 连接XML-RPC接口
地址: https://helloai.jp/xmlrpc.php
用户名: ai-agent

步骤2: 上传封面图片
图片路径: {image_path}

步骤3: 创建文章
标题: {title}
正文: {content}
分类: {category}

步骤4: 发布文章"""

# 准备训练数据
articles = [
    {
        "title": "AI如何改变2026年的工作方式",
        "content": "本文探讨人工智能对未来职场的影响...",
        "category": "科技",
        "tags": "AI,职场,未来",
        "image_path": "./images/ai-work.jpg"
    },
    # 可以添加更多文章样本
]

# 生成训练数据
training_data = []
for article in articles:
    response = client.chat.completions.create(
        model="gpt-4",
        messages=[
            {"role": "system", "content": "你是一个WordPress运营专家。"},
            {"role": "user", "content": PROMPT_TEMPLATE.format(**article)}
        ],
        temperature=0.7
    )
    
    training_data.append({
        "input": f"帮我发布文章：{article['title']}",
        "output": response.choices[0].message.content
    })

# 保存训练数据
with open("wp_skill_training_data.json", "w", encoding="utf-8") as f:
    json.dump(training_data, f, ensure_ascii=False, indent=2)

print(f"生成了 {len(training_data)} 条训练数据")

python generate_data.py

import json

# 读取原始数据
with open("wp_skill_training_data.json", "r") as f:
    data = json.load(f)

# 转换为训练格式
train_data = []
for item in data:
    # 构造成指令微调格式
    train_data.append({
        "instruction": item["input"],
        "input": "",
        "output": item["output"]
    })

# 保存为JSONL（方便后续训练）
with open("train.jsonl", "w", encoding="utf-8") as f:
    for item in train_data:
        f.write(json.dumps(item, ensure_ascii=False) + "\n")

print(f"转换完成，共 {len(train_data)} 条")

python prepare_dataset.py

from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments, Trainer
from datasets import load_dataset

# 使用小型模型作为基础
model_name = "gpt2"  # 可以换成 Qwen/Qwen2-0.5B 等更小的模型

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

# 添加pad token
tokenizer.pad_token = tokenizer.eos_token

# 加载数据
dataset = load_dataset("json", data_files="train.jsonl", split="train")

# 数据预处理
def preprocess(example):
    prompt = f"指令: {example['instruction']}\n输入: {example['input']}\n输出: {example['output']}"
    result = tokenizer(prompt, truncation=True, max_length=512)
    result["labels"] = result["input_ids"]
    return result

dataset = dataset.map(preprocess, remove_columns=dataset.column_names)

# 训练参数
training_args = TrainingArguments(
    output_dir="./wp-skill-model",
    num_train_epochs=3,
    per_device_train_batch_size=2,
    learning_rate=1e-4,
    save_steps=100,
    logging_steps=50,
)

# 开始训练
trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=dataset,
)

trainer.train()
model.save_pretrained("./wp-skill-model")
tokenizer.save_pretrained("./wp-skill-model")

python finetune.py

from transformers import AutoModelForCausalLM, AutoTokenizer

# 加载蒸馏后的模型
model_path = "./wp-skill-model"
tokenizer = AutoTokenizer.from_pretrained(model_path)
model = AutoModelForCausalLM.from_pretrained(model_path)

# 测试输入
test_input = "帮我发布文章：GPT-5新特性解析"

# 生成
inputs = tokenizer(test_input, return_tensors="pt")
outputs = model.generate(
    inputs["input_ids"],
    max_new_tokens=200,
    temperature=0.7,
    do_sample=True
)

# 解析结果
result = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(result)

python test_model.py

步骤1: 连接XML-RPC接口
地址: https://helloai.jp/xmlrpc.php
用户名: ai-agent

步骤2: 上传封面图片
...

步骤3: 创建文章
...

import xmlrpc.client
from transformers import AutoModelForCausalLM, AutoTokenizer

class WordPressPublisher:
    def __init__(self, model_path="./wp-skill-model"):
        self.tokenizer = AutoTokenizer.from_pretrained(model_path)
        self.model = AutoModelForCausalLM.from_pretrained(model_path)
        
        # WordPress连接配置
        self.wp_url = "https://helloai.jp/xmlrpc.php"
        self.username = "ai-agent"
        self.password = "你的应用密码"
    
    def publish(self, title, content, category="未分类"):
        # 1. 用蒸馏模型生成发布步骤
        steps = self._generate_steps(title, content, category)
        
        # 2. 执行步骤
        for step in steps:
            self._execute_step(step)
        
        return "发布成功"
    
    def _generate_steps(self, title, content, category):
        prompt = f"帮我发布文章：{title}"
        inputs = self.tokenizer(prompt, return_tensors="pt")
        outputs = self.model.generate(inputs["input_ids"], max_new_tokens=200)
        return self.tokenizer.decode(outputs[0], skip_special_tokens=True)
    
    def _execute_step(self, step):
        # 执行XML-RPC操作
        server = xmlrpc.client.ServerProxy(self.wp_url)
        # 根据step内容执行对应操作...
        pass

# 使用
wp = WordPressPublisher()
wp.publish("测试文章", "这是正文内容")