{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# MiniCPM-2B 参数高效微调（LoRA）消费级单卡示例\n",
    "\n",
    "本 notebook 是一个使用 `AdvertiseGen` 数据集对 MiniCPM-2B 进行 LoRA 微调，使其具备专业的广告生成能力的代码示例。\n",
    "\n",
    "## 硬件需求\n",
    "- 显存：12GB\n",
    "- 显卡架构：安培架构（推荐）\n",
    "- 内存：16GB"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 1. 准备数据集\n",
    "\n",
    "下载 AdvertiseGen 数据集\n",
    "- [Google Drive](https://drive.google.com/file/d/13_vf0xRTQsyneRKdD1bZIr93vBGOczrk/view?usp=sharing)\n",
    "- [Tsinghua Cloud](https://cloud.tsinghua.edu.cn/f/b3f119a008264b1cabd1/?dl=1)\n",
    "\n",
    "下载后的数据集格式为 `.tar.gz` 的压缩格式，接下来的操作中，假设该压缩包被置于 `finetune/data/`。\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# 校验文件完整性\n",
    "!md5sum data/AdvertiseGen.tar.gz "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# 解压数据集\n",
    "!tar xvf data/AdvertiseGen.tar.gz "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# 转换为 ChatML 格式\n",
    "import os\n",
    "import shutil\n",
    "import json\n",
    "\n",
    "input_dir = \"data/AdvertiseGen\"\n",
    "output_dir = \"data/AdvertiseGenChatML\"\n",
    "if os.path.exists(output_dir):\n",
    "    shutil.rmtree(output_dir)\n",
    "os.makedirs(output_dir, exist_ok=True)\n",
    "\n",
    "for fn in [\"train.json\", \"dev.json\"]:\n",
    "    data_out_list = []\n",
    "    with open(os.path.join(input_dir, fn), \"r\") as f, open(os.path.join(output_dir, fn), \"w\") as fo:\n",
    "        for line in f:\n",
    "            if len(line.strip()) > 0:\n",
    "                data = json.loads(line)\n",
    "                data_out = {\n",
    "                    \"messages\": [\n",
    "                        {\n",
    "                            \"role\": \"user\",\n",
    "                            \"content\": data[\"content\"],\n",
    "                        },\n",
    "                        {\n",
    "                            \"role\": \"assistant\",\n",
    "                            \"content\": data[\"summary\"],\n",
    "                        },\n",
    "                    ]\n",
    "                }\n",
    "                data_out_list.append(data_out)\n",
    "        json.dump(data_out_list, fo, ensure_ascii=False, indent=4)\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 2. 使用 LoRA 进行微调\n",
    "\n",
    "命令行一键运行"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "!bash lora_finetune.sh"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "base",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.10.13"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}