Add phase1_dora_training.ipynb with auto-resume feature

phase1_dora_training.ipynb

@@ -0,0 +1,673 @@
+{
+ "nbformat": 4,
+ "nbformat_minor": 0,
+ "metadata": {
+  "colab": {
+   "provenance": [],
+   "gpuType": "A100"
+  },
+  "kernelspec": {
+   "name": "python3",
+   "display_name": "Python 3"
+  },
+  "accelerator": "GPU"
+ },
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "source": [
+    "# 🔮 YEJI Phase 1: QDoRA Fine-tuning (PEFT + TRL v0.26+)\n",
+    "\n",
+    "Qwen3-8B-Base 모델을 **PEFT + QDoRA**로 Fine-tuning합니다.\n",
+    "\n",
+    "## 주요 구성\n",
+    "- **모델**: Qwen/Qwen3-8B-Base (4-bit NF4)\n",
+    "- **방법**: QDoRA via PEFT (`use_dora=True`)\n",
+    "- **데이터**: 밸런싱 40K + 멀티턴 500건\n",
+    "- **환경**: Colab A100 40GB\n",
+    "- **TRL**: v0.26+ (SFTConfig 방식)"
+   ],
+   "metadata": {
+    "id": "cell-0"
+   }
+  },
+  {
+   "cell_type": "markdown",
+   "source": [
+    "---\n",
+    "## 1️⃣ 환경 설정"
+   ],
+   "metadata": {
+    "id": "cell-1"
+   }
+  },
+  {
+   "cell_type": "code",
+   "source": [
+    "# GPU 확인\n",
+    "!nvidia-smi"
+   ],
+   "metadata": {
+    "id": "cell-2"
+   },
+   "execution_count": null,
+   "outputs": []
+  },
+  {
+   "cell_type": "code",
+   "source": [
+    "# 패키지 설치 (TRL v0.26+)\n",
+    "!pip install --no-cache-dir -q transformers accelerate peft bitsandbytes\n",
+    "!pip install --no-cache-dir -q \"trl>=0.26\" datasets huggingface_hub\n",
+    "\n",
+    "print(\"✅ 패키지 설치 완료!\")"
+   ],
+   "metadata": {
+    "id": "cell-3"
+   },
+   "execution_count": null,
+   "outputs": []
+  },
+  {
+   "cell_type": "code",
+   "source": [
+    "# 버전 확인 및 임포트\n",
+    "import json\n",
+    "import gc\n",
+    "import time\n",
+    "from datetime import datetime\n",
+    "\n",
+    "import torch\n",
+    "import transformers\n",
+    "import peft\n",
+    "import trl\n",
+    "\n",
+    "print(f\"PyTorch: {torch.__version__}\")\n",
+    "print(f\"Transformers: {transformers.__version__}\")\n",
+    "print(f\"PEFT: {peft.__version__}\")\n",
+    "print(f\"TRL: {trl.__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_memory / 1e9:.1f} GB\")"
+   ],
+   "metadata": {
+    "id": "cell-4"
+   },
+   "execution_count": null,
+   "outputs": []
+  },
+  {
+   "cell_type": "markdown",
+   "source": [
+    "---\n",
+    "## 2️⃣ 설정"
+   ],
+   "metadata": {
+    "id": "cell-5"
+   }
+  },
+  {
+   "cell_type": "code",
+   "source": [
+    "# ============================================================\n",
+    "# 전역 설정\n",
+    "# ============================================================\n",
+    "CONFIG = {\n",
+    "    # 데이터\n",
+    "    \"balanced_dataset\": \"tellang/yeji-fortune-telling-ko-balanced\",\n",
+    "    \"multiturn_dataset\": \"tellang/yeji-fortune-telling-ko-multiturn\",\n",
+    "    \n",
+    "    # 모델\n",
+    "    \"base_model\": \"Qwen/Qwen3-8B-Base\",\n",
+    "    \"output_repo\": \"tellang/yeji-8b-qdora-v1\",\n",
+    "    \n",
+    "    # QDoRA 설정\n",
+    "    \"lora_r\": 8,\n",
+    "    \"lora_alpha\": 16,\n",
+    "    \"use_dora\": True,\n",
+    "    \n",
+    "    # 학습\n",
+    "    \"num_epochs\": 3,\n",
+    "    \"batch_size\": 2,\n",
+    "    \"grad_accum_steps\": 4,\n",
+    "    \"learning_rate\": 2e-4,\n",
+    "    \"max_seq_length\": 2048,\n",
+    "    \n",
+    "    # 체크포인트\n",
+    "    \"save_steps\": 500,\n",
+    "    \"eval_steps\": 500,\n",
+    "    \"save_total_limit\": 3,\n",
+    "    \n",
+    "    # 완료 후 자동 종료\n",
+    "    \"auto_shutdown\": \"unassign\",  # None / \"unassign\" / \"terminate\"\n",
+    "}\n",
+    "\n",
+    "SYSTEM_PROMPT = \"\"\"당신은 전문 점술가 '예지'입니다. 사주팔자, 타로, 호로스코프를 전문적으로 해석합니다.\n",
+    "친근하고 따뜻한 말투로 상담하며, 구체적이고 실용적인 조언을 제공합니다.\"\"\"\n",
+    "\n",
+    "print(\"✅ 설정 완료\")\n",
+    "print(f\"   모델: {CONFIG['base_model']}\")\n",
+    "print(f\"   DoRA: {CONFIG['use_dora']}\")\n",
+    "print(f\"   r={CONFIG['lora_r']}, alpha={CONFIG['lora_alpha']}\")"
+   ],
+   "metadata": {
+    "id": "cell-6"
+   },
+   "execution_count": null,
+   "outputs": []
+  },
+  {
+   "cell_type": "code",
+   "source": [
+    "# ============================================================\n",
+    "# 유틸리티 함수\n",
+    "# ============================================================\n",
+    "def shutdown_colab(mode):\n",
+    "    \"\"\"Colab 세션 종료\"\"\"\n",
+    "    if mode is None:\n",
+    "        return\n",
+    "    try:\n",
+    "        from google.colab import runtime\n",
+    "        if mode == \"unassign\":\n",
+    "            print(\"🔌 GPU 할당 해제 중...\")\n",
+    "            runtime.unassign()\n",
+    "        elif mode == \"terminate\":\n",
+    "            import os\n",
+    "            os._exit(0)\n",
+    "    except Exception as e:\n",
+    "        print(f\"⚠️ 종료 실패: {e}\")\n",
+    "\n",
+    "print(\"✅ 유틸리티 함수 정의 완료\")"
+   ],
+   "metadata": {
+    "id": "cell-7"
+   },
+   "execution_count": null,
+   "outputs": []
+  },
+  {
+   "cell_type": "markdown",
+   "source": [
+    "---\n",
+    "## 3️⃣ HuggingFace 로그인"
+   ],
+   "metadata": {
+    "id": "cell-8"
+   }
+  },
+  {
+   "cell_type": "code",
+   "source": "# HuggingFace 로그인\nfrom huggingface_hub import login\n\ndef extract_token(obj):\n    \"\"\"재귀적으로 토큰 추출\"\"\"\n    if isinstance(obj, str) and obj.startswith('hf_'):\n        return obj\n    if isinstance(obj, dict):\n        for key in ['token', 'HF_TOKEN', 'hf_token']:\n            if key in obj:\n                result = extract_token(obj[key])\n                if result:\n                    return result\n        for v in obj.values():\n            result = extract_token(v)\n            if result:\n                return result\n    return None\n\nHF_TOKEN = None\n\n# 1. Colab secrets\ntry:\n    from google.colab import userdata\n    raw = userdata.get('HF_TOKEN')\n    HF_TOKEN = extract_token(raw) if isinstance(raw, dict) else raw\nexcept Exception:\n    pass\n\n# 2. 환경변수\nif not HF_TOKEN:\n    import os\n    HF_TOKEN = os.environ.get('HF_TOKEN')\n\n# 3. 수동 입력\nif not HF_TOKEN or not isinstance(HF_TOKEN, str):\n    HF_TOKEN = input(\"HuggingFace 토큰 입력: \")\n\nlogin(token=HF_TOKEN)\nprint(\"✅ HuggingFace 로그인 완료!\")",
+   "metadata": {
+    "id": "cell-9"
+   },
+   "execution_count": null,
+   "outputs": []
+  },
+  {
+   "cell_type": "markdown",
+   "source": [
+    "---\n",
+    "## 4️⃣ 데이터 준비"
+   ],
+   "metadata": {
+    "id": "cell-10"
+   }
+  },
+  {
+   "cell_type": "code",
+   "source": [
+    "# 토크나이저 로드\n",
+    "from transformers import AutoTokenizer\n",
+    "\n",
+    "tokenizer = AutoTokenizer.from_pretrained(\n",
+    "    CONFIG[\"base_model\"],\n",
+    "    trust_remote_code=True,\n",
+    ")\n",
+    "\n",
+    "if tokenizer.pad_token is None:\n",
+    "    tokenizer.pad_token = tokenizer.eos_token\n",
+    "    tokenizer.pad_token_id = tokenizer.eos_token_id\n",
+    "\n",
+    "print(f\"✅ 토크나이저 로드 완료\")\n",
+    "print(f\"   vocab_size: {tokenizer.vocab_size}\")"
+   ],
+   "metadata": {
+    "id": "cell-11"
+   },
+   "execution_count": null,
+   "outputs": []
+  },
+  {
+   "cell_type": "code",
+   "source": [
+    "# 데이터셋 로드 및 변환\n",
+    "from datasets import load_dataset, concatenate_datasets\n",
+    "\n",
+    "print(\"📥 데이터셋 로드 중...\")\n",
+    "\n",
+    "# 밸런싱 데이터 로드\n",
+    "balanced_ds = load_dataset(CONFIG[\"balanced_dataset\"], split=\"train\")\n",
+    "print(f\"   밸런싱: {len(balanced_ds):,}건\")\n",
+    "\n",
+    "# 멀티턴 데이터 로드\n",
+    "multiturn_ds = load_dataset(CONFIG[\"multiturn_dataset\"], split=\"train\")\n",
+    "print(f\"   멀티턴: {len(multiturn_ds):,}건\")\n",
+    "\n",
+    "# Alpaca → text 변환\n",
+    "def format_alpaca(example):\n",
+    "    messages = [\n",
+    "        {\"role\": \"system\", \"content\": SYSTEM_PROMPT},\n",
+    "        {\"role\": \"user\", \"content\": example[\"instruction\"] + (\"\\n\" + example[\"input\"] if example.get(\"input\") else \"\")},\n",
+    "        {\"role\": \"assistant\", \"content\": example[\"output\"]},\n",
+    "    ]\n",
+    "    return {\"text\": tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=False)}\n",
+    "\n",
+    "# ShareGPT → text 변환\n",
+    "def format_sharegpt(example):\n",
+    "    convs = json.loads(example[\"conversations\"])\n",
+    "    messages = [{\"role\": \"system\", \"content\": SYSTEM_PROMPT}]\n",
+    "    for msg in convs:\n",
+    "        role = \"user\" if msg[\"role\"] == \"user\" else \"assistant\"\n",
+    "        messages.append({\"role\": role, \"content\": msg[\"content\"]})\n",
+    "    return {\"text\": tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=False)}\n",
+    "\n",
+    "print(\"🔄 데이터 변환 중...\")\n",
+    "balanced_formatted = balanced_ds.map(format_alpaca, remove_columns=balanced_ds.column_names, num_proc=4)\n",
+    "multiturn_formatted = multiturn_ds.map(format_sharegpt, remove_columns=multiturn_ds.column_names, num_proc=4)\n",
+    "\n",
+    "# 병합 및 셔플\n",
+    "train_ds = concatenate_datasets([balanced_formatted, multiturn_formatted]).shuffle(seed=42)\n",
+    "print(f\"\\n✅ 데이터 준비 완료: {len(train_ds):,}건\")"
+   ],
+   "metadata": {
+    "id": "cell-12"
+   },
+   "execution_count": null,
+   "outputs": []
+  },
+  {
+   "cell_type": "code",
+   "source": [
+    "# Train/Eval 분리 (95:5)\n",
+    "split = train_ds.train_test_split(test_size=0.05, seed=42)\n",
+    "train_dataset = split[\"train\"]\n",
+    "eval_dataset = split[\"test\"]\n",
+    "\n",
+    "print(f\"📊 데이터 분리:\")\n",
+    "print(f\"   Train: {len(train_dataset):,}건\")\n",
+    "print(f\"   Eval: {len(eval_dataset):,}건\")\n",
+    "\n",
+    "# 샘플 확인\n",
+    "print(f\"\\n📝 샘플:\")\n",
+    "print(train_dataset[0][\"text\"][:300] + \"...\")"
+   ],
+   "metadata": {
+    "id": "cell-13"
+   },
+   "execution_count": null,
+   "outputs": []
+  },
+  {
+   "cell_type": "markdown",
+   "source": [
+    "---\n",
+    "## 5️⃣ 모델 준비 (QDoRA)"
+   ],
+   "metadata": {
+    "id": "cell-14"
+   }
+  },
+  {
+   "cell_type": "code",
+   "source": [
+    "# 4-bit 양자화 모델 로드\n",
+    "from transformers import AutoModelForCausalLM, BitsAndBytesConfig\n",
+    "\n",
+    "print(f\"📥 모델 로드 중: {CONFIG['base_model']}\")\n",
+    "print(\"   (약 2-3분 소요)\")\n",
+    "\n",
+    "try:\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",
+    "    model = AutoModelForCausalLM.from_pretrained(\n",
+    "        CONFIG[\"base_model\"],\n",
+    "        quantization_config=bnb_config,\n",
+    "        device_map=\"auto\",\n",
+    "        trust_remote_code=True,\n",
+    "        attn_implementation=\"eager\",\n",
+    "    )\n",
+    "    \n",
+    "    print(f\"\\n✅ 모델 로드 완료!\")\n",
+    "    print(f\"   VRAM: {torch.cuda.memory_allocated() / 1e9:.1f} GB\")\n",
+    "\n",
+    "except Exception as e:\n",
+    "    print(f\"❌ 모델 로드 실패: {e}\")\n",
+    "    shutdown_colab(\"unassign\")\n",
+    "    raise"
+   ],
+   "metadata": {
+    "id": "cell-15"
+   },
+   "execution_count": null,
+   "outputs": []
+  },
+  {
+   "cell_type": "code",
+   "source": [
+    "# QDoRA 어댑터 적용\n",
+    "from peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training, TaskType\n",
+    "\n",
+    "print(f\"🔧 QDoRA 어댑터 적용 중... (use_dora={CONFIG['use_dora']})\")\n",
+    "\n",
+    "model = prepare_model_for_kbit_training(model)\n",
+    "\n",
+    "lora_config = LoraConfig(\n",
+    "    task_type=TaskType.CAUSAL_LM,\n",
+    "    r=CONFIG[\"lora_r\"],\n",
+    "    lora_alpha=CONFIG[\"lora_alpha\"],\n",
+    "    target_modules=[\"q_proj\", \"k_proj\", \"v_proj\", \"o_proj\", \"gate_proj\", \"up_proj\", \"down_proj\"],\n",
+    "    lora_dropout=0.05,\n",
+    "    bias=\"none\",\n",
+    "    use_dora=CONFIG[\"use_dora\"],\n",
+    ")\n",
+    "\n",
+    "model = get_peft_model(model, lora_config)\n",
+    "\n",
+    "print(f\"\\n✅ QDoRA 적용 완료!\")\n",
+    "model.print_trainable_parameters()"
+   ],
+   "metadata": {
+    "id": "cell-16"
+   },
+   "execution_count": null,
+   "outputs": []
+  },
+  {
+   "cell_type": "markdown",
+   "source": [
+    "---\n",
+    "## 6️⃣ Baseline 측정"
+   ],
+   "metadata": {
+    "id": "cell-17"
+   }
+  },
+  {
+   "cell_type": "code",
+   "source": [
+    "# Baseline 품질 측정\n",
+    "TEST_PROMPTS = {\n",
+    "    \"사주\": \"1985년 12월 25일 자시(23시)에 태어난 사람의 사주팔자를 분석해주세요.\",\n",
+    "    \"타로\": \"취업 운세를 보려고 합니다. 타로 카드 '황제', '세계', '심판'이 나왔어요.\",\n",
+    "    \"호로스코프\": \"사자자리의 2024년 연간 운세를 알려주세요.\",\n",
+    "}\n",
+    "\n",
+    "KEYWORDS = {\n",
+    "    \"사주\": [\"년\", \"월\", \"일\", \"시\"],\n",
+    "    \"타로\": [\"황제\", \"세계\", \"심판\", \"의미\"],\n",
+    "    \"호로스코프\": [\"사자\", \"운\"],\n",
+    "}\n",
+    "\n",
+    "def generate_response(prompt, max_new_tokens=256):\n",
+    "    messages = [\n",
+    "        {\"role\": \"system\", \"content\": SYSTEM_PROMPT},\n",
+    "        {\"role\": \"user\", \"content\": prompt},\n",
+    "    ]\n",
+    "    text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)\n",
+    "    inputs = tokenizer(text, return_tensors=\"pt\").to(model.device)\n",
+    "    \n",
+    "    with torch.no_grad():\n",
+    "        outputs = model.generate(\n",
+    "            **inputs,\n",
+    "            max_new_tokens=max_new_tokens,\n",
+    "            do_sample=True,\n",
+    "            temperature=0.7,\n",
+    "            top_p=0.9,\n",
+    "            pad_token_id=tokenizer.pad_token_id,\n",
+    "        )\n",
+    "    return tokenizer.decode(outputs[0][inputs['input_ids'].shape[1]:], skip_special_tokens=True).strip()\n",
+    "\n",
+    "def evaluate_quality():\n",
+    "    results = {}\n",
+    "    total_score = 0\n",
+    "    \n",
+    "    for domain, prompt in TEST_PROMPTS.items():\n",
+    "        response = generate_response(prompt)\n",
+    "        keywords = KEYWORDS[domain]\n",
+    "        found = [kw for kw in keywords if kw in response]\n",
+    "        score = len(found) / len(keywords) * 100\n",
+    "        \n",
+    "        results[domain] = {\"score\": score, \"found\": found, \"length\": len(response)}\n",
+    "        total_score += score\n",
+    "        \n",
+    "        status = \"✅\" if score >= 50 else \"⚠️\"\n",
+    "        print(f\"{status} {domain}: {score:.0f}%\")\n",
+    "        print(f\"   키워드: {found}\")\n",
+    "        print(f\"   응답 길이: {len(response)}자\")\n",
+    "    \n",
+    "    avg = total_score / len(TEST_PROMPTS)\n",
+    "    print(f\"\\n📊 종합 점수: {avg:.0f}%\")\n",
+    "    return results, avg\n",
+    "\n",
+    "print(\"=\"*50)\n",
+    "print(\"📊 Baseline (학습 전)\")\n",
+    "print(\"=\"*50)\n",
+    "baseline_results, baseline_avg = evaluate_quality()"
+   ],
+   "metadata": {
+    "id": "cell-18"
+   },
+   "execution_count": null,
+   "outputs": []
+  },
+  {
+   "cell_type": "markdown",
+   "source": [
+    "---\n",
+    "## 7️⃣ 학습"
+   ],
+   "metadata": {
+    "id": "cell-19"
+   }
+  },
+  {
+   "cell_type": "code",
+   "source": "# SFTConfig 설정\nfrom trl import SFTConfig\n\nsft_config = SFTConfig(\n    output_dir=\"./yeji-qdora-v1\",\n    \n    # 학습\n    num_train_epochs=CONFIG[\"num_epochs\"],\n    per_device_train_batch_size=CONFIG[\"batch_size\"],\n    per_device_eval_batch_size=CONFIG[\"batch_size\"],\n    gradient_accumulation_steps=CONFIG[\"grad_accum_steps\"],\n    \n    # Optimizer\n    learning_rate=CONFIG[\"learning_rate\"],\n    lr_scheduler_type=\"cosine\",\n    warmup_ratio=0.05,\n    weight_decay=0.01,\n    optim=\"paged_adamw_8bit\",\n    \n    # Precision & Memory\n    bf16=True,\n    gradient_checkpointing=True,\n    max_grad_norm=0.3,\n    \n    # 저장\n    save_strategy=\"steps\",\n    save_steps=CONFIG[\"save_steps\"],\n    save_total_limit=CONFIG[\"save_total_limit\"],\n    logging_steps=50,\n    \n    # 평가\n    eval_strategy=\"steps\",\n    eval_steps=CONFIG[\"eval_steps\"],\n    load_best_model_at_end=True,\n    metric_for_best_model=\"eval_loss\",\n    \n    # Hub\n    push_to_hub=True,\n    hub_model_id=CONFIG[\"output_repo\"],\n    hub_strategy=\"checkpoint\",\n    \n    report_to=\"none\",\n)\n\nprint(\"✅ SFTConfig 설정 완료\")\nprint(f\"   epochs: {sft_config.num_train_epochs}\")\nprint(f\"   effective_batch: {sft_config.per_device_train_batch_size * sft_config.gradient_accumulation_steps}\")",
+   "metadata": {
+    "id": "cell-20"
+   },
+   "execution_count": null,
+   "outputs": []
+  },
+  {
+   "cell_type": "code",
+   "source": "# SFTTrainer 초기화\nfrom trl import SFTTrainer\n\ntrainer = SFTTrainer(\n    model=model,\n    processing_class=tokenizer,\n    args=sft_config,\n    train_dataset=train_dataset,\n    eval_dataset=eval_dataset,\n)\n\nprint(\"✅ SFTTrainer 초기화 완료!\")",
+   "metadata": {
+    "id": "cell-21"
+   },
+   "execution_count": null,
+   "outputs": []
+  },
+  {
+   "cell_type": "code",
+   "source": "# ============================================================\n# [테스트] 체크포인트 감지 테스트 (학습 전 확인용)\n# ============================================================\n# 실제 학습 전에 체크포인트가 제대로 감지되는지 확인\n# 이 셀은 선택 사항 - 건너뛰어도 됨\n\nfrom huggingface_hub import HfApi\n\ndef test_checkpoint_detection():\n    \"\"\"체크포인트 감지 테스트\"\"\"\n    print(\"🧪 체크포인트 감지 테스트\")\n    print(\"=\" * 50)\n    \n    try:\n        api = HfApi()\n        \n        # Repo 존재 확인\n        try:\n            files = api.list_repo_files(CONFIG[\"output_repo\"])\n            print(f\"✅ Repo 존재: {CONFIG['output_repo']}\")\n        except Exception:\n            print(f\"📭 Repo 없음: {CONFIG['output_repo']}\")\n            print(\"   → 처음부터 새로 학습합니다.\")\n            return\n        \n        # 체크포인트 파일 확인\n        checkpoint_files = [f for f in files if f.startswith(\"last-checkpoint/\")]\n        \n        if checkpoint_files:\n            print(f\"✅ 체크포인트 파일 {len(checkpoint_files)}개 발견:\")\n            for f in checkpoint_files[:5]:\n                print(f\"   - {f}\")\n            if len(checkpoint_files) > 5:\n                print(f\"   ... 외 {len(checkpoint_files) - 5}개\")\n            \n            # trainer_state 확인\n            if \"last-checkpoint/trainer_state.json\" in files:\n                import requests\n                url = f\"https://huggingface.co/{CONFIG['output_repo']}/raw/main/last-checkpoint/trainer_state.json\"\n                state = requests.get(url).json()\n                \n                print(f\"\\n📊 학습 진행 상황:\")\n                print(f\"   Step: {state.get('global_step', '?')}\")\n                print(f\"   Epoch: {state.get('epoch', 0):.2f} / {CONFIG['num_epochs']}\")\n                print(f\"   Best Loss: {state.get('best_metric', '?')}\")\n                \n                # 남은 학습량 계산\n                total_steps = len(train_dataset) // (CONFIG[\"batch_size\"] * CONFIG[\"grad_accum_steps\"]) * CONFIG[\"num_epochs\"]\n                current_step = state.get('global_step', 0)\n                remaining = ((total_steps - current_step) / total_steps) * 100\n                \n                print(f\"\\n📈 예상 진행률:\")\n                print(f\"   완료: {100 - remaining:.1f}%\")\n                print(f\"   남음: {remaining:.1f}%\")\n                \n            print(f\"\\n✅ 테스트 통과! 전체 실행하면 이어서 학습됩니다.\")\n        else:\n            print(\"📭 체크포인트 없음\")\n            print(\"   → 처음부터 새로 학습합니다.\")\n            \n    except Exception as e:\n        print(f\"❌ 테스트 실패: {e}\")\n\ntest_checkpoint_detection()",
+   "metadata": {},
+   "execution_count": null,
+   "outputs": []
+  },
+  {
+   "cell_type": "code",
+   "source": "# 학습 시작 (체크포인트 자동 감지)\nfrom huggingface_hub import HfApi, snapshot_download\nimport os\n\ndef check_checkpoint():\n    \"\"\"HuggingFace Hub에서 체크포인트 확인 및 다운로드\"\"\"\n    try:\n        api = HfApi()\n        files = api.list_repo_files(CONFIG[\"output_repo\"])\n        \n        if \"last-checkpoint/trainer_state.json\" in files:\n            print(\"📦 기존 체크포인트 발견!\")\n            \n            # 체크포인트 다운로드\n            print(\"   다운로드 중...\")\n            checkpoint_path = snapshot_download(\n                CONFIG[\"output_repo\"],\n                allow_patterns=[\"last-checkpoint/*\"],\n                local_dir=\"./resume_checkpoint\",\n            )\n            \n            # 진행 상황 확인\n            import json\n            state_path = f\"{checkpoint_path}/last-checkpoint/trainer_state.json\"\n            with open(state_path) as f:\n                state = json.load(f)\n            \n            print(f\"   ✅ Step: {state.get('global_step', '?')}\")\n            print(f\"   ✅ Epoch: {state.get('epoch', '?'):.2f}\")\n            print(f\"   ✅ Best Loss: {state.get('best_metric', '?'):.4f}\")\n            \n            return f\"{checkpoint_path}/last-checkpoint\"\n        else:\n            print(\"📭 체크포인트 없음 - 처음부터 학습\")\n            return None\n            \n    except Exception as e:\n        print(f\"📭 체크포인트 확인 실패: {e}\")\n        print(\"   처음부터 학습합니다.\")\n        return None\n\n# 체크포인트 확인\nprint(\"=\" * 50)\nprint(\"🔍 체크포인트 확인 중...\")\nprint(\"=\" * 50)\nresume_path = check_checkpoint()\n\n# 학습 시작\nprint(\"\\n\" + \"=\" * 50)\nif resume_path:\n    print(\"🔄 이어서 학습 시작!\")\nelse:\n    print(\"🚀 새로 학습 시작!\")\nprint(f\"   시작: {datetime.now().strftime('%Y-%m-%d %H:%M:%S')}\")\nprint(f\"   데이터: {len(train_dataset):,}건\")\nprint(f\"   Baseline: {baseline_avg:.0f}%\")\nprint(\"=\" * 50)\n\nstart_time = time.time()\n\ntry:\n    if resume_path:\n        train_result = trainer.train(resume_from_checkpoint=resume_path)\n    else:\n        train_result = trainer.train()\n    \n    elapsed = time.time() - start_time\n    print(f\"\\n✅ 학습 완료!\")\n    print(f\"   소요: {elapsed/60:.1f}분\")\n    print(f\"   Train Loss: {train_result.training_loss:.4f}\")\n\nexcept Exception as e:\n    print(f\"\\n❌ 학습 실패: {e}\")\n    shutdown_colab(\"unassign\")\n    raise",
+   "metadata": {
+    "id": "cell-22"
+   },
+   "execution_count": null,
+   "outputs": []
+  },
+  {
+   "cell_type": "markdown",
+   "source": [
+    "---\n",
+    "## 8️⃣ 평가"
+   ],
+   "metadata": {
+    "id": "cell-23"
+   }
+  },
+  {
+   "cell_type": "code",
+   "source": [
+    "# Eval Loss\n",
+    "eval_result = trainer.evaluate()\n",
+    "print(f\"📊 Eval Loss: {eval_result['eval_loss']:.4f}\")"
+   ],
+   "metadata": {
+    "id": "cell-24"
+   },
+   "execution_count": null,
+   "outputs": []
+  },
+  {
+   "cell_type": "code",
+   "source": [
+    "# 학습 후 품질 측정\n",
+    "print(\"=\"*50)\n",
+    "print(\"📊 Fine-tuned (학습 후)\")\n",
+    "print(\"=\"*50)\n",
+    "finetuned_results, finetuned_avg = evaluate_quality()\n",
+    "\n",
+    "# 비교\n",
+    "print(\"\\n\" + \"=\"*50)\n",
+    "print(\"📊 Baseline vs Fine-tuned\")\n",
+    "print(\"=\"*50)\n",
+    "for domain in TEST_PROMPTS:\n",
+    "    b = baseline_results[domain][\"score\"]\n",
+    "    f = finetuned_results[domain][\"score\"]\n",
+    "    diff = f - b\n",
+    "    print(f\"{domain}: {b:.0f}% → {f:.0f}% ({'+' if diff >= 0 else ''}{diff:.0f}%)\")\n",
+    "\n",
+    "improvement = finetuned_avg - baseline_avg\n",
+    "print(f\"\\n종합: {baseline_avg:.0f}% → {finetuned_avg:.0f}% ({'+' if improvement >= 0 else ''}{improvement:.0f}%)\")"
+   ],
+   "metadata": {
+    "id": "cell-25"
+   },
+   "execution_count": null,
+   "outputs": []
+  },
+  {
+   "cell_type": "markdown",
+   "source": [
+    "---\n",
+    "## 9️⃣ 저장 & 업로드"
+   ],
+   "metadata": {
+    "id": "cell-26"
+   }
+  },
+  {
+   "cell_type": "code",
+   "source": [
+    "# 저장 및 업로드\n",
+    "print(\"💾 저장 중...\")\n",
+    "trainer.save_model(\"./yeji-qdora-v1-final\")\n",
+    "tokenizer.save_pretrained(\"./yeji-qdora-v1-final\")\n",
+    "\n",
+    "print(f\"📤 Hub 업로드 중: {CONFIG['output_repo']}\")\n",
+    "trainer.push_to_hub(\n",
+    "    commit_message=f\"YEJI QDoRA v1 (Loss: {train_result.training_loss:.4f}, {baseline_avg:.0f}%→{finetuned_avg:.0f}%)\"\n",
+    ")\n",
+    "\n",
+    "print(f\"\\n✅ 완료!\")\n",
+    "print(f\"   https://huggingface.co/{CONFIG['output_repo']}\")"
+   ],
+   "metadata": {
+    "id": "cell-27"
+   },
+   "execution_count": null,
+   "outputs": []
+  },
+  {
+   "cell_type": "code",
+   "source": [
+    "# 결과 요약\n",
+    "print(\"\\n\" + \"=\"*50)\n",
+    "print(\"📊 학습 결과 요약\")\n",
+    "print(\"=\"*50)\n",
+    "print(f\"모델: {CONFIG['base_model']}\")\n",
+    "print(f\"방법: QDoRA (r={CONFIG['lora_r']}, alpha={CONFIG['lora_alpha']})\")\n",
+    "print(f\"데이터: {len(train_dataset):,}건\")\n",
+    "print(f\"Train Loss: {train_result.training_loss:.4f}\")\n",
+    "print(f\"Eval Loss: {eval_result['eval_loss']:.4f}\")\n",
+    "print(f\"품질: {baseline_avg:.0f}% → {finetuned_avg:.0f}%\")\n",
+    "print(f\"출력: {CONFIG['output_repo']}\")"
+   ],
+   "metadata": {
+    "id": "cell-28"
+   },
+   "execution_count": null,
+   "outputs": []
+  },
+  {
+   "cell_type": "markdown",
+   "source": [
+    "---\n",
+    "## 🔟 리소스 정리"
+   ],
+   "metadata": {
+    "id": "cell-29"
+   }
+  },
+  {
+   "cell_type": "code",
+   "source": [
+    "# 메모리 정리\n",
+    "del model\n",
+    "del trainer\n",
+    "gc.collect()\n",
+    "torch.cuda.empty_cache()\n",
+    "print(\"✅ 메모리 정리 완료\")"
+   ],
+   "metadata": {
+    "id": "cell-30"
+   },
+   "execution_count": null,
+   "outputs": []
+  },
+  {
+   "cell_type": "code",
+   "source": [
+    "# 자동 종료\n",
+    "if CONFIG[\"auto_shutdown\"]:\n",
+    "    print(f\"⏰ 5초 후 GPU 해제...\")\n",
+    "    time.sleep(5)\n",
+    "    shutdown_colab(CONFIG[\"auto_shutdown\"])"
+   ],
+   "metadata": {
+    "id": "cell-31"
+   },
+   "execution_count": null,
+   "outputs": []
+  },
+  {
+   "cell_type": "code",
+   "source": [
+    "# (유틸) 수동 GPU 해제\n",
+    "# from google.colab import runtime\n",
+    "# runtime.unassign()"
+   ],
+   "metadata": {
+    "id": "cell-32"
+   },
+   "execution_count": null,
+   "outputs": []
+  }
+ ]
+}