Add new Colab notebook with proper format

Network_Admin_LLM_Training.ipynb

@@ -0,0 +1,443 @@
+{
+  "nbformat": 4,
+  "nbformat_minor": 5,
+  "metadata": {
+    "colab": {
+      "provenance": [],
+      "gpuType": "T4",
+      "name": "Network_Admin_LLM_Training.ipynb"
+    },
+    "kernelspec": {
+      "display_name": "Python 3",
+      "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": "3.10.12"
+    },
+    "accelerator": "GPU"
+  },
+  "cells": [
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "title"
+      },
+      "source": [
+        "# 🚀 Network Admin LLM - QLoRA Fine-tuning\n",
+        "\n",
+        "**基礎模型**: microsoft/Phi-4-mini-instruct (3.8B)\n",
+        "\n",
+        "**訓練方法**: QLoRA SFT (4-bit 量化 + LoRA)\n",
+        "\n",
+        "**數據集**: NetEval (5732題) + Telecom Intent Config (10K) ≈ 15K 樣本\n",
+        "\n",
+        "**訓練時間**: ~2-3 小時 (T4 GPU)\n",
+        "\n",
+        "---\n",
+        "\n",
+        "⚠️ **請先設定 GPU Runtime**: Runtime → Change runtime type → T4 GPU"
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "step1"
+      },
+      "source": [
+        "## Step 1. 安裝依賴"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "metadata": {
+        "id": "install"
+      },
+      "execution_count": null,
+      "outputs": [],
+      "source": [
+        "!pip install -q transformers trl peft bitsandbytes accelerate datasets trackio\n",
+        "\n",
+        "import torch\n",
+        "print(f'PyTorch: {torch.__version__}')\n",
+        "print(f'CUDA: {torch.cuda.is_available()}')\n",
+        "if torch.cuda.is_available():\n",
+        "    print(f'GPU: {torch.cuda.get_device_name(0)}')\n",
+        "    print(f'VRAM: {torch.cuda.get_device_properties(0).total_mem / 1024**3:.1f} GB')\n",
+        "print('✅ Dependencies installed')"
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "step2"
+      },
+      "source": [
+        "## Step 2. HuggingFace 登入\n",
+        "\n",
+        "請先到 https://huggingface.co/settings/tokens 取得你的 Access Token（需要 Write 權限）"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "metadata": {
+        "id": "login"
+      },
+      "execution_count": null,
+      "outputs": [],
+      "source": [
+        "from huggingface_hub import login\n",
+        "\n",
+        "# 方法 1: 直接輸入 token\n",
+        "# login(token=\"hf_xxxxxxxxxxxxxxxxxxxx\")\n",
+        "\n",
+        "# 方法 2: 互動式登入 (推薦)\n",
+        "login()"
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "step3"
+      },
+      "source": [
+        "## Step 3. 配置設定\n",
+        "\n",
+        "請修改 `HF_USERNAME` 為你的 HuggingFace 用戶名"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "metadata": {
+        "id": "config"
+      },
+      "execution_count": null,
+      "outputs": [],
+      "source": [
+        "# ======== 請修改這裡 ========\n",
+        "HF_USERNAME = \"YOUR_HF_USERNAME\"  # 改成你的 HuggingFace 用戶名\n",
+        "# ============================\n",
+        "\n",
+        "MODEL_NAME = \"microsoft/Phi-4-mini-instruct\"\n",
+        "OUTPUT_DIR = f\"{HF_USERNAME}/network-admin-phi4-mini\"\n",
+        "\n",
+        "TRAINING_CONFIG = {\n",
+        "    \"learning_rate\": 2e-4,\n",
+        "    \"num_epochs\": 3,\n",
+        "    \"batch_size\": 4,\n",
+        "    \"gradient_accumulation\": 4,\n",
+        "    \"max_seq_length\": 2048,\n",
+        "    \"lora_r\": 16,\n",
+        "    \"lora_alpha\": 32,\n",
+        "}\n",
+        "\n",
+        "print(f'模型: {MODEL_NAME}')\n",
+        "print(f'輸出: https://huggingface.co/{OUTPUT_DIR}')"
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "step4"
+      },
+      "source": [
+        "## Step 4. 載入數據集"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "metadata": {
+        "id": "data"
+      },
+      "execution_count": null,
+      "outputs": [],
+      "source": [
+        "import os\n",
+        "import torch\n",
+        "from datasets import load_dataset, concatenate_datasets\n",
+        "\n",
+        "# --- NetEval 網管考試題庫 ---\n",
+        "print('📚 載入 NetEval 網管考試題庫...')\n",
+        "neteval_dataset = load_dataset('NASP/neteval-exam', split='train')\n",
+        "print(f'   NetEval: {len(neteval_dataset)} 題')\n",
+        "\n",
+        "def convert_neteval(example):\n",
+        "    question = example['Question']\n",
+        "    options = (\n",
+        "        f\"\\nA. {example.get('A', '')}\"\n",
+        "        f\"\\nB. {example.get('B', '')}\"\n",
+        "        f\"\\nC. {example.get('C', '')}\"\n",
+        "        f\"\\nD. {example.get('D', '')}\"\n",
+        "    )\n",
+        "    answer = f\"正確答案是: {example['Answer']}\"\n",
+        "    if example.get('Explanation'):\n",
+        "        answer += f\"\\n\\n📖 解說: {example['Explanation']}\"\n",
+        "    return {\n",
+        "        'messages': [\n",
+        "            {'role': 'system', 'content': '你是一位網路管理專家。請回答關於網路、安全、路由、交換機、VLAN、防火牆等IT基礎設施的問題。'},\n",
+        "            {'role': 'user', 'content': f'{question}{options}'},\n",
+        "            {'role': 'assistant', 'content': answer}\n",
+        "        ]\n",
+        "    }\n",
+        "\n",
+        "neteval_converted = neteval_dataset.map(\n",
+        "    convert_neteval,\n",
+        "    remove_columns=neteval_dataset.column_names\n",
+        ")\n",
+        "\n",
+        "# --- Telecom 電信意圖配置數據 ---\n",
+        "print('📚 載入 Telecom 電信意圖配置數據...')\n",
+        "telecom_dataset = load_dataset('nraptisss/telecom-intent-config-sft-10k', split='train')\n",
+        "telecom_messages = telecom_dataset.map(\n",
+        "    lambda x: {'messages': x['messages']},\n",
+        "    remove_columns=[c for c in telecom_dataset.column_names if c != 'messages']\n",
+        ")\n",
+        "print(f'   Telecom: {len(telecom_messages)} 條')\n",
+        "\n",
+        "# --- 合併數據集 ---\n",
+        "combined = concatenate_datasets([neteval_converted, telecom_messages])\n",
+        "split_data = combined.train_test_split(test_size=0.1, seed=42)\n",
+        "train_ds = split_data['train']\n",
+        "eval_ds = split_data['test']\n",
+        "\n",
+        "print(f'\\n📊 總計: {len(train_ds)} train / {len(eval_ds)} eval')"
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "step5"
+      },
+      "source": [
+        "## Step 5. 載入模型與 QLoRA"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "metadata": {
+        "id": "model"
+      },
+      "execution_count": null,
+      "outputs": [],
+      "source": [
+        "from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig\n",
+        "from peft import LoraConfig, prepare_model_for_kbit_training, get_peft_model\n",
+        "\n",
+        "# Tokenizer\n",
+        "tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)\n",
+        "tokenizer.pad_token = tokenizer.eos_token\n",
+        "print(f'Tokenizer vocab: {len(tokenizer):,}')\n",
+        "\n",
+        "# 4-bit 量化配置\n",
+        "bnb_config = BitsAndBytesConfig(\n",
+        "    load_in_4bit=True,\n",
+        "    bnb_4bit_quant_type='nf4',\n",
+        "    bnb_4bit_compute_dtype=torch.bfloat16,\n",
+        "    bnb_4bit_use_double_quant=True,\n",
+        ")\n",
+        "\n",
+        "# 載入模型\n",
+        "print('📥 載入模型 (4-bit)...')\n",
+        "model = AutoModelForCausalLM.from_pretrained(\n",
+        "    MODEL_NAME,\n",
+        "    quantization_config=bnb_config,\n",
+        "    device_map='auto',\n",
+        "    trust_remote_code=True,\n",
+        ")\n",
+        "model = prepare_model_for_kbit_training(model)\n",
+        "print('✅ 模型載入完成')\n",
+        "\n",
+        "# LoRA 配置\n",
+        "lora_config = LoraConfig(\n",
+        "    r=TRAINING_CONFIG['lora_r'],\n",
+        "    lora_alpha=TRAINING_CONFIG['lora_alpha'],\n",
+        "    lora_dropout=0.05,\n",
+        "    bias='none',\n",
+        "    task_type='CAUSAL_LM',\n",
+        "    target_modules=['q_proj', 'k_proj', 'v_proj', 'o_proj', 'gate_proj', 'up_proj', 'down_proj'],\n",
+        "    modules_to_save=['lm_head', 'embed_tokens'],\n",
+        ")\n",
+        "\n",
+        "model = get_peft_model(model, lora_config)\n",
+        "model.print_trainable_parameters()"
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "step6"
+      },
+      "source": [
+        "## Step 6. 開始訓練\n",
+        "\n",
+        "☕ 這步驟約需 2-3 小時，可以去喝杯咖啡"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "metadata": {
+        "id": "train"
+      },
+      "execution_count": null,
+      "outputs": [],
+      "source": [
+        "from trl import SFTTrainer, SFTConfig\n",
+        "\n",
+        "training_args = SFTConfig(\n",
+        "    learning_rate=TRAINING_CONFIG['learning_rate'],\n",
+        "    lr_scheduler_type='cosine',\n",
+        "    warmup_ratio=0.1,\n",
+        "    num_train_epochs=TRAINING_CONFIG['num_epochs'],\n",
+        "    per_device_train_batch_size=TRAINING_CONFIG['batch_size'],\n",
+        "    gradient_accumulation_steps=TRAINING_CONFIG['gradient_accumulation'],\n",
+        "    max_seq_length=TRAINING_CONFIG['max_seq_length'],\n",
+        "    gradient_checkpointing=True,\n",
+        "    bf16=True,\n",
+        "    output_dir='./output',\n",
+        "    logging_steps=10,\n",
+        "    save_steps=500,\n",
+        "    eval_steps=500,\n",
+        "    push_to_hub=True,\n",
+        "    hub_model_id=OUTPUT_DIR,\n",
+        "    logging_strategy='steps',\n",
+        "    logging_first_step=True,\n",
+        ")\n",
+        "\n",
+        "trainer = SFTTrainer(\n",
+        "    model=model,\n",
+        "    args=training_args,\n",
+        "    train_dataset=train_ds,\n",
+        "    eval_dataset=eval_ds,\n",
+        "    processing_class=tokenizer,\n",
+        "    peft_config=lora_config,\n",
+        ")\n",
+        "\n",
+        "print('🚀 開始訓練...')\n",
+        "print(f'訓練完成後模型將保存到: https://huggingface.co/{OUTPUT_DIR}')\n",
+        "trainer.train()"
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "step7"
+      },
+      "source": [
+        "## Step 7. 上傳模型到 HuggingFace Hub"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "metadata": {
+        "id": "push"
+      },
+      "execution_count": null,
+      "outputs": [],
+      "source": [
+        "trainer.push_to_hub()\n",
+        "print(f'✅ 模型已上傳: https://huggingface.co/{OUTPUT_DIR}')"
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "step8"
+      },
+      "source": [
+        "## Step 8. 測試推理\n",
+        "\n",
+        "訓練完成後，直接在這裡測試模型："
+      ]
+    },
+    {
+      "cell_type": "code",
+      "metadata": {
+        "id": "inference"
+      },
+      "execution_count": null,
+      "outputs": [],
+      "source": [
+        "# 測試推理\n",
+        "from transformers import pipeline\n",
+        "\n",
+        "pipe = pipeline(\n",
+        "    'text-generation',\n",
+        "    model=model,\n",
+        "    tokenizer=tokenizer,\n",
+        "    max_new_tokens=512,\n",
+        "    do_sample=True,\n",
+        "    temperature=0.7,\n",
+        ")\n",
+        "\n",
+        "# 測試問題\n",
+        "test_questions = [\n",
+        "    '什麼是 VLAN？如何配置？',\n",
+        "    'OSPF 和 BGP 的區別是什麼？',\n",
+        "    '如何排查網路連接問題？',\n",
+        "]\n",
+        "\n",
+        "for q in test_questions:\n",
+        "    print(f'\\n{\"=\"*50}')\n",
+        "    print(f'💬 問: {q}')\n",
+        "    print(f'{\"=\"*50}')\n",
+        "    messages = [\n",
+        "        {'role': 'system', 'content': '你是一位網路管理專家。'},\n",
+        "        {'role': 'user', 'content': q}\n",
+        "    ]\n",
+        "    output = pipe(messages)\n",
+        "    reply = output[0]['generated_text'][-1]['content']\n",
+        "    print(f'🤖 答: {reply}')"
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "done"
+      },
+      "source": [
+        "---\n",
+        "\n",
+        "## ✅ 完成！\n",
+        "\n",
+        "你的網管 LLM 已經訓練完成並上傳到 HuggingFace Hub。\n",
+        "\n",
+        "### 在其他地方使用這個模型：\n",
+        "\n",
+        "```python\n",
+        "from peft import PeftModel\n",
+        "from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig\n",
+        "import torch\n",
+        "\n",
+        "bnb_config = BitsAndBytesConfig(\n",
+        "    load_in_4bit=True,\n",
+        "    bnb_4bit_quant_type='nf4',\n",
+        "    bnb_4bit_compute_dtype=torch.bfloat16,\n",
+        ")\n",
+        "\n",
+        "base_model = AutoModelForCausalLM.from_pretrained(\n",
+        "    'microsoft/Phi-4-mini-instruct',\n",
+        "    quantization_config=bnb_config,\n",
+        "    device_map='auto',\n",
+        ")\n",
+        "model = PeftModel.from_pretrained(base_model, 'YOUR_USERNAME/network-admin-phi4-mini')\n",
+        "tokenizer = AutoTokenizer.from_pretrained('microsoft/Phi-4-mini-instruct')\n",
+        "\n",
+        "messages = [{'role': 'user', 'content': '什麼是 VLAN？'}]\n",
+        "text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)\n",
+        "inputs = tokenizer(text, return_tensors='pt').to('cuda')\n",
+        "outputs = model.generate(**inputs, max_new_tokens=512)\n",
+        "print(tokenizer.decode(outputs[0], skip_special_tokens=True))\n",
+        "```"
+      ]
+    }
+  ]
+}