Fix manual QLoRA notebook label collation bug

EthicalHacking_Stable_QLoRA_ManualLoop_NO_UNSLOTH.ipynb

@@ -6,438 +6,44 @@
    "source": [
     "# ✅ Stable Ethical-Hacking QLoRA Fine-Tuning — NO Unsloth, NO Trainer\n",
     "\n",
-    "This notebook is rewritten from root because Kaggle kept hitting Unsloth's compiled trainer bug: `AttributeError: 'int' object has no attribute 'mean'`.\n",
-    "\n",
-    "## Root fix\n",
-    "- ❌ No `unsloth`\n",
-    "- ❌ No `SFTTrainer`\n",
-    "- ❌ No `Trainer.train()`\n",
-    "- ✅ Vanilla `transformers` + `peft` + `bitsandbytes`\n",
-    "- ✅ Manual PyTorch training loop\n",
-    "- ✅ QLoRA NF4 4-bit training for free Kaggle/Colab T4\n",
-    "\n",
-    "Default model is **LiquidAI/LFM2.5-1.2B-Instruct** because it is the most reliable fit for free 16GB T4. You can switch to Qwen3-4B in the config cell if you want more capability but slower training.\n",
-    "\n",
-    "> Use only for ethical, defensive, authorized security education and research.\n"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": ["## 1. Install dependencies\n"]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "!pip install -q -U \"transformers>=4.56.0\" \"peft>=0.18.0\" \"accelerate>=1.0.0\" \"bitsandbytes>=0.45.0\" \"datasets>=3.0.0\" \"huggingface_hub>=0.25.0\" tqdm\n"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": ["## 2. Optional Hugging Face login\n"]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "from huggingface_hub import login\n",
-    "# login(token=\"hf_YOUR_WRITE_TOKEN\")\n"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": ["## 3. Configure run\n"]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "import os, math, random, gc, time\n",
-    "import torch\n",
-    "\n",
-    "# ======================== MODEL CHOICE ========================\n",
-    "# Recommended for free T4: LiquidAI/LFM2.5-1.2B-Instruct\n",
-    "MODEL_ID = \"LiquidAI/LFM2.5-1.2B-Instruct\"\n",
-    "RUN_NAME = \"lfm25-cyber-stable-qlora\"\n",
-    "\n",
-    "# Stronger but slower/tighter on T4. Uncomment to use Qwen3-4B:\n",
-    "# MODEL_ID = \"Qwen/Qwen3-4B-Instruct-2507\"\n",
-    "# RUN_NAME = \"qwen3-4b-cyber-stable-qlora\"\n",
-    "\n",
-    "# ======================== DATA ========================\n",
-    "DATASET_CHOICE = \"cybersecurity\"  # cybersecurity | ultrachat | openhermes | code_corpus\n",
-    "SAMPLE_SIZE = 30000                 # set lower for faster test, e.g. 2000\n",
-    "MAX_SEQ_LENGTH = 2048               # T4-safe. Try 4096 only for LFM2.5 if VRAM allows\n",
-    "\n",
-    "# ======================== QLoRA ========================\n",
-    "LORA_R = 64\n",
-    "LORA_ALPHA = 128\n",
-    "LORA_DROPOUT = 0.05\n",
-    "TARGET_MODULES = \"all-linear\"     # QLoRA-style: target all linear layers\n",
-    "\n",
-    "# ======================== TRAINING ========================\n",
-    "BATCH_SIZE = 1                      # micro-batch per GPU. Stable on free T4\n",
-    "GRAD_ACCUM = 8                      # effective batch = 8\n",
-    "MAX_STEPS = 1000                    # raise to 2000-4000 for longer training\n",
-    "LEARNING_RATE = 2e-4\n",
-    "WARMUP_STEPS = 50\n",
-    "WEIGHT_DECAY = 0.01\n",
-    "LOGGING_STEPS = 10\n",
-    "SAVE_STEPS = 250\n",
-    "SEED = 3407\n",
-    "\n",
-    "OUTPUT_DIR = f\"./{RUN_NAME}-adapter\"\n",
-    "HUB_MODEL_ID = \"your-username/\" + RUN_NAME\n",
-    "PUSH_TO_HUB = False\n",
-    "\n",
-    "# T4 supports fp16, not bf16\n",
-    "COMPUTE_DTYPE = torch.float16\n",
-    "\n",
-    "random.seed(SEED)\n",
-    "torch.manual_seed(SEED)\n",
-    "if torch.cuda.is_available():\n",
-    "    torch.cuda.manual_seed_all(SEED)\n",
-    "    print(f\"GPU: {torch.cuda.get_device_name(0)}\")\n",
-    "    print(f\"VRAM: {torch.cuda.get_device_properties(0).total_memory/1e9:.2f} GB\")\n",
-    "else:\n",
-    "    print(\"WARNING: No GPU detected. Enable GPU runtime.\")\n",
-    "print(\"MODEL_ID =\", MODEL_ID)\n",
-    "print(\"DATASET_CHOICE =\", DATASET_CHOICE)\n"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": ["## 4. Load 4-bit model and attach QLoRA adapter\n"]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig\n",
-    "from peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training, TaskType\n",
-    "\n",
-    "bnb_config = BitsAndBytesConfig(\n",
-    "    load_in_4bit=True,\n",
-    "    bnb_4bit_quant_type=\"nf4\",\n",
-    "    bnb_4bit_use_double_quant=True,\n",
-    "    bnb_4bit_compute_dtype=COMPUTE_DTYPE,\n",
-    ")\n",
-    "\n",
-    "tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True, use_fast=True)\n",
-    "if tokenizer.pad_token is None:\n",
-    "    tokenizer.pad_token = tokenizer.eos_token\n",
-    "tokenizer.padding_side = \"right\"\n",
-    "\n",
-    "model = AutoModelForCausalLM.from_pretrained(\n",
-    "    MODEL_ID,\n",
-    "    quantization_config=bnb_config,\n",
-    "    device_map={\"\": 0},\n",
-    "    trust_remote_code=True,\n",
-    "    low_cpu_mem_usage=True,\n",
-    ")\n",
-    "\n",
-    "model.config.use_cache = False\n",
-    "model = prepare_model_for_kbit_training(model, use_gradient_checkpointing=True)\n",
-    "\n",
-    "lora_config = LoraConfig(\n",
-    "    r=LORA_R,\n",
-    "    lora_alpha=LORA_ALPHA,\n",
-    "    lora_dropout=LORA_DROPOUT,\n",
-    "    bias=\"none\",\n",
-    "    task_type=TaskType.CAUSAL_LM,\n",
-    "    target_modules=TARGET_MODULES,\n",
-    "    use_rslora=True,\n",
-    ")\n",
-    "\n",
-    "model = get_peft_model(model, lora_config)\n",
-    "model.print_trainable_parameters()\n",
-    "\n",
-    "if torch.cuda.is_available():\n",
-    "    print(f\"VRAM after model load: {torch.cuda.memory_allocated()/1e9:.2f} GB / {torch.cuda.get_device_properties(0).total_memory/1e9:.2f} GB\")\n"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": ["## 5. Load and convert dataset\n"]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "from datasets import load_dataset, concatenate_datasets\n",
-    "\n",
-    "SYSTEM_SAFE = \"You are a cybersecurity education assistant. Provide defensive, ethical, and authorized security guidance only. Refuse harmful or unauthorized requests.\"\n",
-    "\n",
-    "def convert_fenrir(example):\n",
-    "    return {\"messages\": [\n",
-    "        {\"role\": \"system\", \"content\": example.get(\"system\") or SYSTEM_SAFE},\n",
-    "        {\"role\": \"user\", \"content\": example[\"user\"]},\n",
-    "        {\"role\": \"assistant\", \"content\": example[\"assistant\"]},\n",
-    "    ]}\n",
-    "\n",
-    "def convert_trendyol(example):\n",
-    "    return {\"messages\": [\n",
-    "        {\"role\": \"system\", \"content\": example.get(\"system\") or SYSTEM_SAFE},\n",
-    "        {\"role\": \"user\", \"content\": example[\"user\"]},\n",
-    "        {\"role\": \"assistant\", \"content\": example[\"assistant\"]},\n",
-    "    ]}\n",
-    "\n",
-    "def convert_ultrachat(example):\n",
-    "    return {\"messages\": example[\"messages\"]}\n",
-    "\n",
-    "def convert_conversations(example):\n",
-    "    msgs = []\n",
-    "    sys = example.get(\"system_prompt\", \"\") or example.get(\"system\", \"\")\n",
-    "    if sys:\n",
-    "        msgs.append({\"role\": \"system\", \"content\": sys})\n",
-    "    for turn in example[\"conversations\"]:\n",
-    "        role = \"user\" if turn.get(\"from\") in (\"human\", \"user\") else \"assistant\"\n",
-    "        msgs.append({\"role\": role, \"content\": turn.get(\"value\", \"\")})\n",
-    "    return {\"messages\": msgs}\n",
-    "\n",
-    "def convert_code(example):\n",
-    "    code = example[\"text\"]\n",
-    "    lang = example.get(\"language\", \"\")\n",
-    "    prompt = f\"Explain and improve this {lang} code snippet. Focus on correctness, safety, and clarity.\"\n",
-    "    return {\"messages\": [\n",
-    "        {\"role\": \"user\", \"content\": prompt},\n",
-    "        {\"role\": \"assistant\", \"content\": code},\n",
-    "    ]}\n",
-    "\n",
-    "if DATASET_CHOICE == \"cybersecurity\":\n",
-    "    ds1 = load_dataset(\"AlicanKiraz0/Cybersecurity-Dataset-Fenrir-v2.1\", split=\"train\")\n",
-    "    ds1 = ds1.map(convert_fenrir, remove_columns=ds1.column_names)\n",
-    "    ds2 = load_dataset(\"Trendyol/Trendyol-Cybersecurity-Instruction-Tuning-Dataset\", split=\"train\")\n",
-    "    ds2 = ds2.map(convert_trendyol, remove_columns=ds2.column_names)\n",
-    "    dataset = concatenate_datasets([ds1, ds2])\n",
-    "elif DATASET_CHOICE == \"ultrachat\":\n",
-    "    dataset = load_dataset(\"HuggingFaceH4/ultrachat_200k\", split=\"train_sft\")\n",
-    "    dataset = dataset.map(convert_ultrachat, remove_columns=dataset.column_names)\n",
-    "elif DATASET_CHOICE == \"openhermes\":\n",
-    "    dataset = load_dataset(\"teknium/OpenHermes-2.5\", split=\"train\")\n",
-    "    dataset = dataset.map(convert_conversations, remove_columns=dataset.column_names)\n",
-    "elif DATASET_CHOICE == \"code_corpus\":\n",
-    "    dataset = load_dataset(\"krystv/code-corpus-llm-training\", split=\"train\")\n",
-    "    dataset = dataset.map(convert_code, remove_columns=dataset.column_names)\n",
-    "else:\n",
-    "    raise ValueError(DATASET_CHOICE)\n",
-    "\n",
-    "print(f\"Loaded rows: {len(dataset):,}\")\n",
-    "if SAMPLE_SIZE and len(dataset) > SAMPLE_SIZE:\n",
-    "    dataset = dataset.shuffle(seed=SEED).select(range(SAMPLE_SIZE))\n",
-    "    print(f\"Subsampled rows: {len(dataset):,}\")\n",
-    "\n",
-    "sample = dataset[0][\"messages\"]\n",
-    "print(\"Sample roles:\", [m[\"role\"] for m in sample])\n",
-    "for m in sample[:3]:\n",
-    "    print(m[\"role\"], \":\", m[\"content\"][:140].replace(\"\\n\", \" \"))\n"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": ["## 6. Apply chat template and tokenize\n"]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "def messages_to_text(example):\n",
-    "    try:\n",
-    "        text = tokenizer.apply_chat_template(example[\"messages\"], tokenize=False, add_generation_prompt=False)\n",
-    "    except Exception:\n",
-    "        parts = []\n",
-    "        for m in example[\"messages\"]:\n",
-    "            parts.append(f\"<{m['role']}>\\n{m['content']}\\n</{m['role']}>\")\n",
-    "        text = \"\\n\".join(parts) + tokenizer.eos_token\n",
-    "    return {\"text\": text}\n",
-    "\n",
-    "text_dataset = dataset.map(messages_to_text, remove_columns=[\"messages\"])\n",
-    "print(text_dataset[0][\"text\"][:500])\n",
-    "\n",
-    "def tokenize_fn(batch):\n",
-    "    out = tokenizer(batch[\"text\"], truncation=True, max_length=MAX_SEQ_LENGTH, padding=False, add_special_tokens=False)\n",
-    "    out[\"labels\"] = [ids.copy() for ids in out[\"input_ids\"]]\n",
-    "    return out\n",
-    "\n",
-    "tokenized = text_dataset.map(tokenize_fn, batched=True, remove_columns=[\"text\"], num_proc=2, desc=\"Tokenizing\")\n",
-    "tokenized = tokenized.filter(lambda x: len(x[\"input_ids\"]) > 8)\n",
-    "print(f\"Tokenized rows: {len(tokenized):,}\")\n",
-    "print(\"First length:\", len(tokenized[0][\"input_ids\"]))\n"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": ["## 7. Build DataLoader\n"]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "from torch.utils.data import DataLoader\n",
-    "from transformers import DataCollatorForLanguageModeling\n",
-    "\n",
-    "collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False, pad_to_multiple_of=8)\n",
-    "tokenized.set_format(type=\"torch\", columns=[\"input_ids\", \"attention_mask\", \"labels\"])\n",
-    "train_loader = DataLoader(tokenized, batch_size=BATCH_SIZE, shuffle=True, collate_fn=collator, num_workers=0)\n",
-    "print(\"Batches per epoch:\", len(train_loader))\n",
-    "print(\"Effective batch size:\", BATCH_SIZE * GRAD_ACCUM)\n"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": ["## 8. Manual training loop — no Trainer, no Unsloth\n"]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "import bitsandbytes as bnb\n",
-    "from transformers import get_cosine_schedule_with_warmup\n",
-    "from tqdm.auto import tqdm\n",
-    "\n",
-    "trainable_params = [p for p in model.parameters() if p.requires_grad]\n",
-    "optimizer = bnb.optim.AdamW8bit(trainable_params, lr=LEARNING_RATE, weight_decay=WEIGHT_DECAY)\n",
-    "scheduler = get_cosine_schedule_with_warmup(optimizer, num_warmup_steps=WARMUP_STEPS, num_training_steps=MAX_STEPS)\n",
-    "\n",
-    "model.train()\n",
-    "global_step = 0\n",
-    "micro_step = 0\n",
-    "running_loss = 0.0\n",
-    "start_time = time.time()\n",
-    "progress = tqdm(total=MAX_STEPS, desc=\"Training\", dynamic_ncols=True)\n",
-    "optimizer.zero_grad(set_to_none=True)\n",
-    "\n",
-    "while global_step < MAX_STEPS:\n",
-    "    for batch in train_loader:\n",
-    "        batch = {k: v.to(model.device) for k, v in batch.items()}\n",
-    "        outputs = model(**batch)\n",
-    "        loss = outputs.loss / GRAD_ACCUM\n",
-    "        loss.backward()\n",
-    "        running_loss += loss.item() * GRAD_ACCUM\n",
-    "        micro_step += 1\n",
-    "\n",
-    "        if micro_step % GRAD_ACCUM == 0:\n",
-    "            torch.nn.utils.clip_grad_norm_(trainable_params, 1.0)\n",
-    "            optimizer.step()\n",
-    "            scheduler.step()\n",
-    "            optimizer.zero_grad(set_to_none=True)\n",
-    "            global_step += 1\n",
-    "            progress.update(1)\n",
-    "\n",
-    "            if global_step % LOGGING_STEPS == 0:\n",
-    "                avg_loss = running_loss / LOGGING_STEPS\n",
-    "                running_loss = 0.0\n",
-    "                lr = scheduler.get_last_lr()[0]\n",
-    "                elapsed = (time.time() - start_time) / 60\n",
-    "                if torch.cuda.is_available():\n",
-    "                    vram = torch.cuda.memory_allocated() / 1e9\n",
-    "                    total = torch.cuda.get_device_properties(0).total_memory / 1e9\n",
-    "                    print(f\"step={global_step} loss={avg_loss:.4f} lr={lr:.2e} vram={vram:.2f}/{total:.2f}GB elapsed={elapsed:.1f}m\")\n",
-    "                else:\n",
-    "                    print(f\"step={global_step} loss={avg_loss:.4f} lr={lr:.2e} elapsed={elapsed:.1f}m\")\n",
-    "\n",
-    "            if global_step % SAVE_STEPS == 0:\n",
-    "                ckpt = f\"{OUTPUT_DIR}/checkpoint-{global_step}\"\n",
-    "                model.save_pretrained(ckpt)\n",
-    "                tokenizer.save_pretrained(ckpt)\n",
-    "                print(f\"Saved checkpoint: {ckpt}\")\n",
-    "\n",
-    "            if global_step >= MAX_STEPS:\n",
-    "                break\n",
-    "\n",
-    "progress.close()\n",
-    "print(\"Training complete\")\n"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": ["## 9. Save adapter / optional push\n"]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "model.save_pretrained(OUTPUT_DIR)\n",
-    "tokenizer.save_pretrained(OUTPUT_DIR)\n",
-    "print(f\"Saved final LoRA adapter to: {OUTPUT_DIR}\")\n",
-    "\n",
-    "if PUSH_TO_HUB:\n",
-    "    model.push_to_hub(HUB_MODEL_ID)\n",
-    "    tokenizer.push_to_hub(HUB_MODEL_ID)\n",
-    "    print(f\"Pushed to: https://huggingface.co/{HUB_MODEL_ID}\")\n"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": ["## 10. Inference sanity check\n"]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "model.eval()\n",
-    "test_prompt = \"Explain how parameterized queries prevent SQL injection. Give a safe Python example.\"\n",
-    "messages = [{\"role\": \"system\", \"content\": SYSTEM_SAFE}, {\"role\": \"user\", \"content\": test_prompt}]\n",
-    "try:\n",
-    "    prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)\n",
-    "except Exception:\n",
-    "    prompt = f\"<system>\\n{SYSTEM_SAFE}\\n</system>\\n<user>\\n{test_prompt}\\n</user>\\n<assistant>\\n\"\n",
-    "inputs = tokenizer(prompt, return_tensors=\"pt\").to(model.device)\n",
-    "with torch.no_grad():\n",
-    "    out = model.generate(**inputs, max_new_tokens=384, do_sample=True, temperature=0.7, top_p=0.9, pad_token_id=tokenizer.pad_token_id, eos_token_id=tokenizer.eos_token_id)\n",
-    "print(tokenizer.decode(out[0][inputs[\"input_ids\"].shape[1]:], skip_special_tokens=True))\n"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "## Troubleshooting\n",
-    "\n",
-    "If you OOM on T4, use this order:\n",
-    "1. Keep `BATCH_SIZE=1`; increase/decrease `GRAD_ACCUM` only.\n",
-    "2. Lower `MAX_SEQ_LENGTH` to 1024.\n",
-    "3. Lower `LORA_R` to 32.\n",
-    "4. Lower `SAMPLE_SIZE` and `MAX_STEPS` for a fast smoke test.\n",
-    "\n",
-    "The previous `int.mean()` error cannot happen here because this notebook never calls Unsloth or `Trainer.train()`.\n"
-   ]
-  }
+    "Backup stable notebook using vanilla `transformers` + `peft` + `bitsandbytes` and a manual PyTorch loop.\n",
+    "\n",
+    "## Fix in this revision\n",
+    "\n",
+    "The previous manual notebook created a variable-length `labels` column before collation. That caused:\n",
+    "\n",
+    "```text\n",
+    "ValueError: expected sequence of length 2048 at dim 1 (got 232)\n",
+    "features (`labels` in this case) have excessive nesting\n",
+    "```\n",
+    "\n",
+    "This revision **does not create labels during tokenization**. `DataCollatorForLanguageModeling` now creates and pads labels correctly.\n",
+    "\n",
+    "> For lowest VRAM, use the Unsloth LFM2.5 notebook. This is the no-Unsloth fallback.\n"
+   ]
+  },
+  {"cell_type":"markdown","metadata":{},"source":["## 1. Install dependencies\n"]},
+  {"cell_type":"code","execution_count":null,"metadata":{},"outputs":[],"source":["!pip install -q -U \"transformers>=4.56.0\" \"peft>=0.18.0\" \"accelerate>=1.0.0\" \"bitsandbytes>=0.45.0\" \"datasets>=3.0.0\" \"huggingface_hub>=0.25.0\" tqdm\n"]},
+  {"cell_type":"markdown","metadata":{},"source":["## 2. Optional Hugging Face login\n"]},
+  {"cell_type":"code","execution_count":null,"metadata":{},"outputs":[],"source":["from huggingface_hub import login\n","# login(token=\"hf_YOUR_WRITE_TOKEN\")\n"]},
+  {"cell_type":"markdown","metadata":{},"source":["## 3. Configure run\n"]},
+  {"cell_type":"code","execution_count":null,"metadata":{},"outputs":[],"source":["import os, math, random, gc, time\n","import torch\n","MODEL_ID = \"LiquidAI/LFM2.5-1.2B-Instruct\"\n","RUN_NAME = \"lfm25-cyber-stable-qlora\"\n","# MODEL_ID = \"Qwen/Qwen3-4B-Instruct-2507\"\n","# RUN_NAME = \"qwen3-4b-cyber-stable-qlora\"\n","DATASET_CHOICE = \"cybersecurity\"\n","SAMPLE_SIZE = 30000\n","MAX_SEQ_LENGTH = 2048\n","LORA_R = 64\n","LORA_ALPHA = 128\n","LORA_DROPOUT = 0.05\n","TARGET_MODULES = \"all-linear\"\n","BATCH_SIZE = 1\n","GRAD_ACCUM = 8\n","MAX_STEPS = 1000\n","LEARNING_RATE = 2e-4\n","WARMUP_STEPS = 50\n","WEIGHT_DECAY = 0.01\n","LOGGING_STEPS = 10\n","SAVE_STEPS = 250\n","SEED = 3407\n","OUTPUT_DIR = f\"./{RUN_NAME}-adapter\"\n","HUB_MODEL_ID = \"your-username/\" + RUN_NAME\n","PUSH_TO_HUB = False\n","COMPUTE_DTYPE = torch.float16\n","random.seed(SEED)\n","torch.manual_seed(SEED)\n","if torch.cuda.is_available():\n","    torch.cuda.manual_seed_all(SEED)\n","    print(f\"GPU: {torch.cuda.get_device_name(0)}\")\n","    print(f\"VRAM: {torch.cuda.get_device_properties(0).total_memory/1e9:.2f} GB\")\n","else:\n","    print(\"WARNING: No GPU detected. Enable GPU runtime.\")\n","print(\"MODEL_ID =\", MODEL_ID)\n"]},
+  {"cell_type":"markdown","metadata":{},"source":["## 4. Load 4-bit model and attach QLoRA adapter\n"]},
+  {"cell_type":"code","execution_count":null,"metadata":{},"outputs":[],"source":["from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig\n","from peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training, TaskType\n","bnb_config = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type=\"nf4\", bnb_4bit_use_double_quant=True, bnb_4bit_compute_dtype=COMPUTE_DTYPE)\n","tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True, use_fast=True)\n","if tokenizer.pad_token is None:\n","    tokenizer.pad_token = tokenizer.eos_token\n","tokenizer.padding_side = \"right\"\n","model = AutoModelForCausalLM.from_pretrained(MODEL_ID, quantization_config=bnb_config, device_map={\"\": 0}, trust_remote_code=True, low_cpu_mem_usage=True)\n","model.config.use_cache = False\n","model = prepare_model_for_kbit_training(model, use_gradient_checkpointing=True)\n","lora_config = LoraConfig(r=LORA_R, lora_alpha=LORA_ALPHA, lora_dropout=LORA_DROPOUT, bias=\"none\", task_type=TaskType.CAUSAL_LM, target_modules=TARGET_MODULES, use_rslora=True)\n","model = get_peft_model(model, lora_config)\n","model.print_trainable_parameters()\n","if torch.cuda.is_available():\n","    print(f\"VRAM after model load: {torch.cuda.memory_allocated()/1e9:.2f} GB / {torch.cuda.get_device_properties(0).total_memory/1e9:.2f} GB\")\n"]},
+  {"cell_type":"markdown","metadata":{},"source":["## 5. Load and convert dataset\n"]},
+  {"cell_type":"code","execution_count":null,"metadata":{},"outputs":[],"source":["from datasets import load_dataset, concatenate_datasets\n","SYSTEM_SAFE = \"You are a cybersecurity education assistant. Provide defensive, ethical, and authorized security guidance only. Refuse harmful or unauthorized requests.\"\n","def convert_sua(example):\n","    return {\"messages\": [{\"role\": \"system\", \"content\": example.get(\"system\") or SYSTEM_SAFE}, {\"role\": \"user\", \"content\": example[\"user\"]}, {\"role\": \"assistant\", \"content\": example[\"assistant\"]}]}\n","if DATASET_CHOICE == \"cybersecurity\":\n","    ds1 = load_dataset(\"AlicanKiraz0/Cybersecurity-Dataset-Fenrir-v2.1\", split=\"train\")\n","    ds1 = ds1.map(convert_sua, remove_columns=ds1.column_names)\n","    ds2 = load_dataset(\"Trendyol/Trendyol-Cybersecurity-Instruction-Tuning-Dataset\", split=\"train\")\n","    ds2 = ds2.map(convert_sua, remove_columns=ds2.column_names)\n","    dataset = concatenate_datasets([ds1, ds2])\n","else:\n","    raise ValueError(\"This compact fallback notebook currently defaults to cybersecurity only.\")\n","print(f\"Loaded rows: {len(dataset):,}\")\n","if SAMPLE_SIZE and len(dataset) > SAMPLE_SIZE:\n","    dataset = dataset.shuffle(seed=SEED).select(range(SAMPLE_SIZE))\n","    print(f\"Subsampled rows: {len(dataset):,}\")\n","print(dataset[0][\"messages\"])\n"]},
+  {"cell_type":"markdown","metadata":{},"source":["## 6. Apply chat template and tokenize — fixed labels\n"]},
+  {"cell_type":"code","execution_count":null,"metadata":{},"outputs":[],"source":["def messages_to_text(example):\n","    try:\n","        text = tokenizer.apply_chat_template(example[\"messages\"], tokenize=False, add_generation_prompt=False)\n","    except Exception:\n","        parts = [f\"<{m['role']}>\\n{m['content']}\\n</{m['role']}>\" for m in example[\"messages\"]]\n","        text = \"\\n\".join(parts) + tokenizer.eos_token\n","    return {\"text\": text}\n","text_dataset = dataset.map(messages_to_text, remove_columns=[\"messages\"])\n","print(text_dataset[0][\"text\"][:500])\n","def tokenize_fn(batch):\n","    # IMPORTANT: Do NOT create labels here. The data collator creates/pads labels safely.\n","    return tokenizer(batch[\"text\"], truncation=True, max_length=MAX_SEQ_LENGTH, padding=False, add_special_tokens=False)\n","tokenized = text_dataset.map(tokenize_fn, batched=True, remove_columns=[\"text\"], num_proc=2, desc=\"Tokenizing\")\n","tokenized = tokenized.filter(lambda x: len(x[\"input_ids\"]) > 8)\n","print(f\"Tokenized rows: {len(tokenized):,}\")\n","print(\"First length:\", len(tokenized[0][\"input_ids\"]))\n"]},
+  {"cell_type":"markdown","metadata":{},"source":["## 7. Build DataLoader\n"]},
+  {"cell_type":"code","execution_count":null,"metadata":{},"outputs":[],"source":["from torch.utils.data import DataLoader\n","from transformers import DataCollatorForLanguageModeling\n","collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False, pad_to_multiple_of=8)\n","tokenized.set_format(type=\"torch\", columns=[\"input_ids\", \"attention_mask\"])\n","train_loader = DataLoader(tokenized, batch_size=BATCH_SIZE, shuffle=True, collate_fn=collator, num_workers=0)\n","print(\"Batches per epoch:\", len(train_loader))\n","print(\"Effective batch size:\", BATCH_SIZE * GRAD_ACCUM)\n"]},
+  {"cell_type":"markdown","metadata":{},"source":["## 8. Manual training loop\n"]},
+  {"cell_type":"code","execution_count":null,"metadata":{},"outputs":[],"source":["import bitsandbytes as bnb\n","from transformers import get_cosine_schedule_with_warmup\n","from tqdm.auto import tqdm\n","trainable_params = [p for p in model.parameters() if p.requires_grad]\n","optimizer = bnb.optim.AdamW8bit(trainable_params, lr=LEARNING_RATE, weight_decay=WEIGHT_DECAY)\n","scheduler = get_cosine_schedule_with_warmup(optimizer, num_warmup_steps=WARMUP_STEPS, num_training_steps=MAX_STEPS)\n","model.train()\n","global_step = 0\n","micro_step = 0\n","running_loss = 0.0\n","start_time = time.time()\n","progress = tqdm(total=MAX_STEPS, desc=\"Training\", dynamic_ncols=True)\n","optimizer.zero_grad(set_to_none=True)\n","while global_step < MAX_STEPS:\n","    for batch in train_loader:\n","        batch = {k: v.to(model.device) for k, v in batch.items()}\n","        outputs = model(**batch)\n","        loss = outputs.loss / GRAD_ACCUM\n","        loss.backward()\n","        running_loss += loss.item() * GRAD_ACCUM\n","        micro_step += 1\n","        if micro_step % GRAD_ACCUM == 0:\n","            torch.nn.utils.clip_grad_norm_(trainable_params, 1.0)\n","            optimizer.step(); scheduler.step(); optimizer.zero_grad(set_to_none=True)\n","            global_step += 1; progress.update(1)\n","            if global_step % LOGGING_STEPS == 0:\n","                avg_loss = running_loss / LOGGING_STEPS; running_loss = 0.0\n","                lr = scheduler.get_last_lr()[0]\n","                elapsed = (time.time() - start_time) / 60\n","                if torch.cuda.is_available():\n","                    vram = torch.cuda.memory_allocated() / 1e9; total = torch.cuda.get_device_properties(0).total_memory / 1e9\n","                    print(f\"step={global_step} loss={avg_loss:.4f} lr={lr:.2e} vram={vram:.2f}/{total:.2f}GB elapsed={elapsed:.1f}m\")\n","                else:\n","                    print(f\"step={global_step} loss={avg_loss:.4f} lr={lr:.2e} elapsed={elapsed:.1f}m\")\n","            if global_step % SAVE_STEPS == 0:\n","                ckpt = f\"{OUTPUT_DIR}/checkpoint-{global_step}\"\n","                model.save_pretrained(ckpt); tokenizer.save_pretrained(ckpt)\n","                print(f\"Saved checkpoint: {ckpt}\")\n","            if global_step >= MAX_STEPS: break\n","progress.close()\n","print(\"Training complete\")\n"]},
+  {"cell_type":"markdown","metadata":{},"source":["## 9. Save adapter / optional push\n"]},
+  {"cell_type":"code","execution_count":null,"metadata":{},"outputs":[],"source":["model.save_pretrained(OUTPUT_DIR)\n","tokenizer.save_pretrained(OUTPUT_DIR)\n","print(f\"Saved final LoRA adapter to: {OUTPUT_DIR}\")\n","if PUSH_TO_HUB:\n","    model.push_to_hub(HUB_MODEL_ID); tokenizer.push_to_hub(HUB_MODEL_ID)\n"]},
+  {"cell_type":"markdown","metadata":{},"source":["## 10. Inference sanity check\n"]},
+  {"cell_type":"code","execution_count":null,"metadata":{},"outputs":[],"source":["model.eval()\n","test_prompt = \"Explain how parameterized queries prevent SQL injection. Give a safe Python example.\"\n","messages = [{\"role\": \"system\", \"content\": SYSTEM_SAFE}, {\"role\": \"user\", \"content\": test_prompt}]\n","try:\n","    prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)\n","except Exception:\n","    prompt = f\"<system>\\n{SYSTEM_SAFE}\\n</system>\\n<user>\\n{test_prompt}\\n</user>\\n<assistant>\\n\"\n","inputs = tokenizer(prompt, return_tensors=\"pt\").to(model.device)\n","with torch.no_grad():\n","    out = model.generate(**inputs, max_new_tokens=384, do_sample=True, temperature=0.7, top_p=0.9, pad_token_id=tokenizer.pad_token_id, eos_token_id=tokenizer.eos_token_id)\n","print(tokenizer.decode(out[0][inputs[\"input_ids\"].shape[1]:], skip_special_tokens=True))\n"]}
  ],
- "metadata": {
-  "kernelspec": {"display_name": "Python 3", "language": "python", "name": "python3"},
-  "language_info": {"name": "python", "version": "3.10"}
- },
+ "metadata": {"kernelspec": {"display_name": "Python 3", "language": "python", "name": "python3"}, "language_info": {"name": "python", "version": "3.10"}},
  "nbformat": 4,
  "nbformat_minor": 5
 }