fix: update Kaggle notebook with self-contained cells

- Each cell defines its own paths (no NameError)
- Auto-downloads model only if not exists
- 9 steps: clone → install → model → config → train → merge → done

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

kaggle_train_stack29.ipynb

@@ -16,10 +16,9 @@
     "---\n",
     "\n",
     "**Instructions:**\n",
-    "1. Kaggle → New Notebook\n",
-    "2. Add this notebook's code OR clone from GitHub\n",
-    "3. Enable GPU (Settings → Accelerator → GPU P100)\n",
-    "4. Run cells in order\n",
+    "1. Enable GPU: Settings → Accelerator → GPU P100\n",
+    "2. Run cells in order from the top\n",
+    "3. Model auto-downloads if not present\n",
     "\n",
     "---"
    ]
@@ -30,7 +29,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "# Check GPU\n",
+    "# STEP 1: Check GPU\n",
     "!nvidia-smi"
    ]
   },
@@ -40,13 +39,12 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "# STEP 1: Clone the repo\n",
+    "# STEP 2: Clone repo\n",
     "import os\n",
     "import shutil\n",
     "\n",
     "REPO_DIR = \"/kaggle/working/stack-2.9\"\n",
     "\n",
-    "# Remove old if exists\n",
     "if os.path.exists(REPO_DIR):\n",
     "    shutil.rmtree(REPO_DIR)\n",
     "\n",
@@ -62,7 +60,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "# STEP 2: Install dependencies\n",
+    "# STEP 3: Install dependencies\n",
     "!pip install -q torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118\n",
     "!pip install -q transformers peft accelerate datasets pyyaml tqdm scipy bitsandbytes\n",
     "print(\"✅ Dependencies installed\")"
@@ -74,22 +72,37 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "# STEP 3: Download Base Model\n",
-    "from transformers import AutoModelForCausalLM, AutoTokenizer\n",
+    "# STEP 4: Setup paths (MODEL_DIR, OUTPUT_DIR)\n",
+    "import os\n",
     "\n",
-    "MODEL_NAME = \"Qwen/Qwen2.5-Coder-7B\"\n",
+    "REPO_DIR = \"/kaggle/working/stack-2.9\"\n",
     "MODEL_DIR = os.path.join(REPO_DIR, \"base_model_qwen7b\")\n",
+    "OUTPUT_DIR = os.path.join(REPO_DIR, \"training_output\")\n",
+    "\n",
+    "print(f\"REPO_DIR: {REPO_DIR}\")\n",
+    "print(f\"MODEL_DIR: {MODEL_DIR}\")\n",
+    "print(f\"OUTPUT_DIR: {OUTPUT_DIR}\")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# STEP 5: Download model (if not exists)\n",
+    "from transformers import AutoModelForCausalLM, AutoTokenizer\n",
     "\n",
-    "if not os.path.exists(os.path.join(MODEL_DIR, \"config.json\")):\n",
-    "    print(f\"Downloading {MODEL_NAME}...\")\n",
+    "if os.path.exists(os.path.join(MODEL_DIR, \"config.json\")):\n",
+    "    print(\"✅ Model already exists, skipping download!\")\n",
+    "else:\n",
+    "    print(\"⬇️ Downloading model (Qwen2.5-Coder-7B)...\")\n",
     "    print(\"This takes ~10-15 minutes...\")\n",
-    "    tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, trust_remote_code=True)\n",
+    "    tokenizer = AutoTokenizer.from_pretrained(\"Qwen/Qwen2.5-Coder-7B\", trust_remote_code=True)\n",
     "    tokenizer.save_pretrained(MODEL_DIR)\n",
-    "    model = AutoModelForCausalLM.from_pretrained(MODEL_NAME, trust_remote_code=True)\n",
+    "    model = AutoModelForCausalLM.from_pretrained(\"Qwen/Qwen2.5-Coder-7B\", trust_remote_code=True)\n",
     "    model.save_pretrained(MODEL_DIR)\n",
     "    print(\"✅ Model downloaded!\")\n",
-    "else:\n",
-    "    print(\"✅ Model already exists\")\n",
     "\n",
     "!ls -lh {MODEL_DIR} | head -5"
    ]
@@ -100,30 +113,34 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "# STEP 4: Setup paths and config\n",
+    "# STEP 6: Create config\n",
     "import yaml\n",
-    "\n",
-    "config_path = os.path.join(REPO_DIR, \"stack/training/train_config_local.yaml\")\n",
-    "\n",
-    "with open(config_path, 'r') as f:\n",
-    "    config = yaml.safe_load(f)\n",
-    "\n",
-    "# Update for Kaggle GPU\n",
-    "config['model']['name'] = MODEL_DIR\n",
-    "config['hardware']['device'] = \"cuda\"\n",
-    "config['hardware']['num_gpus'] = 1\n",
-    "\n",
-    "OUTPUT_DIR = os.path.join(REPO_DIR, \"training_output\")\n",
-    "config['output']['lora_dir'] = os.path.join(OUTPUT_DIR, \"lora\")\n",
-    "config['output']['merged_dir'] = os.path.join(OUTPUT_DIR, \"merged\")\n",
+    "import os\n",
     "\n",
     "os.makedirs(OUTPUT_DIR, exist_ok=True)\n",
-    "updated_config = os.path.join(OUTPUT_DIR, \"train_config.yaml\")\n",
     "\n",
-    "with open(updated_config, 'w') as f:\n",
+    "config = {\n",
+    "    'model': {'name': MODEL_DIR, 'trust_remote_code': True, 'torch_dtype': 'float16'},\n",
+    "    'data': {'input_path': './data/final/train.jsonl', 'max_length': 2048},\n",
+    "    'lora': {'r': 16, 'alpha': 32, 'dropout': 0.05,\n",
+    "             'target_modules': ['q_proj', 'k_proj', 'v_proj', 'o_proj'],\n",
+    "             'bias': 'none', 'task_type': 'CAUSAL_LM'},\n",
+    "    'training': {'num_epochs': 1, 'batch_size': 2, 'gradient_accumulation': 4,\n",
+    "                 'learning_rate': 2e-4, 'warmup_steps': 50, 'weight_decay': 0.01,\n",
+    "                 'max_grad_norm': 1.0, 'logging_steps': 5, 'eval_steps': 100,\n",
+    "                 'save_steps': 200, 'save_total_limit': 2, 'fp16': True, 'bf16': False,\n",
+    "                 'gradient_checkpointing': True},\n",
+    "    'output': {'lora_dir': os.path.join(OUTPUT_DIR, 'lora'),\n",
+    "               'merged_dir': os.path.join(OUTPUT_DIR, 'merged')},\n",
+    "    'quantization': {'enabled': False},\n",
+    "    'hardware': {'device': 'cuda', 'num_gpus': 1, 'use_4bit': False, 'use_8bit': False}\n",
+    "}\n",
+    "\n",
+    "config_path = os.path.join(OUTPUT_DIR, \"train_config.yaml\")\n",
+    "with open(config_path, 'w') as f:\n",
     "    yaml.dump(config, f)\n",
     "\n",
-    "print(f\"✅ Config saved to: {updated_config}\")\n",
+    "print(f\"✅ Config saved to: {config_path}\")\n",
     "print(f\"   Device: {config['hardware']['device']}\")"
    ]
   },
@@ -133,7 +150,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "# STEP 5: Train LoRA\n",
+    "# STEP 7: Train LoRA\n",
     "import sys\n",
     "sys.path.insert(0, os.path.join(REPO_DIR, \"stack/training\"))\n",
     "\n",
@@ -142,10 +159,10 @@
     "print(\"=\"*60)\n",
     "\n",
     "from train_lora import train_lora\n",
-    "trainer = train_lora(updated_config)\n",
+    "trainer = train_lora(config_path)\n",
     "\n",
     "print(\"=\"*60)\n",
-    "print(\"TRAINING COMPLETED\")\n",
+    "print(\"TRAINING COMPLETED!\")\n",
     "print(\"=\"*60)"
    ]
   },
@@ -155,7 +172,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "# STEP 6: Merge and save\n",
+    "# STEP 8: Merge model\n",
     "import sys\n",
     "sys.path.insert(0, os.path.join(REPO_DIR, \"stack/training\"))\n",
     "from merge_adapter import merge_adapter\n",
@@ -175,7 +192,7 @@
     "\n",
     "merge_adapter(merge_cfg_path, os.path.join(OUTPUT_DIR, \"lora\"), merged_dir)\n",
     "\n",
-    "print(f\"✅ Model saved to: {merged_dir}\")\n",
+    "print(f\"✅ Merged model saved to: {merged_dir}\")\n",
     "!ls -lh {merged_dir}"
    ]
   },
@@ -185,15 +202,13 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "# STEP 7: Download the trained model (for saving)\n",
-    "# The model is saved at OUTPUT_DIR/merged/\n",
-    "# You can download it from the Kaggle outputs\n",
-    "\n",
-    "print(\"Training complete!\")\n",
-    "print(f\"Model saved at: {merged_dir}\")\n",
-    "print(\"\\nTo download:\")\n",
-    "print(\"1. Click 'Output' tab in Kaggle\")\n",
-    "print(\"2. Download the files from training_output/merged/\")"
+    "# STEP 9: Done!\n",
+    "print(\"=\"*60)\n",
+    "print(\"🎉 TRAINING COMPLETE!\")\n",
+    "print(\"=\"*60)\n",
+    "print(f\"LoRA adapter: {os.path.join(OUTPUT_DIR, 'lora')}\")\n",
+    "print(f\"Merged model: {os.path.join(OUTPUT_DIR, 'merged')}\")\n",
+    "print(\"\\n📥 Download from: Kaggle → Output tab\")"
    ]
   }
  ],