W-VII

community-contributions/abdoul/week_seven_exercise.ipynb

@@ -0,0 +1,495 @@
+{
+  "cells": [
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "GHsssBgWM_l0"
+      },
+      "source": [
+        "# QLoRA Fine-Tuning; LLaMA 3.1 8B"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "MDyR63OTNUJ6"
+      },
+      "outputs": [],
+      "source": [
+        "import sys\n",
+        "print(f\"Python: {sys.version}\")\n",
+        "\n",
+        "import torch\n",
+        "print(f\"PyTorch: {torch.__version__}\")\n",
+        "print(f\"CUDA Available: {torch.cuda.is_available()}\")\n",
+        "print(f\"CUDA Version: {torch.version.cuda}\")\n",
+        "print(f\"GPU: {torch.cuda.get_device_name(0)}\")"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "3_8F5m3xxMtz"
+      },
+      "outputs": [],
+      "source": [
+        "!pip install -q --upgrade transformers==4.48.3 accelerate==1.3.0 datasets==3.2.0\n",
+        "!pip install -q --upgrade peft==0.14.0 trl==0.14.0 bitsandbytes==0.46.0\n",
+        "!pip install -q --upgrade matplotlib scipy scikit-learn\n",
+        "!pip install -q --upgrade \"huggingface_hub<1.0,>=0.24.0\"\n",
+        "!pip install -q --upgrade bitsandbytes"
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "CpVdMBUVxMtz"
+      },
+      "source": [
+        "## Environment Setup"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "-yikV8pRBer9"
+      },
+      "outputs": [],
+      "source": [
+        "import os\n",
+        "import re\n",
+        "import math\n",
+        "import torch\n",
+        "import torch.nn.functional as F\n",
+        "import matplotlib.pyplot as plt\n",
+        "from tqdm import tqdm\n",
+        "from huggingface_hub import login\n",
+        "from transformers import (\n",
+        "    AutoModelForCausalLM,\n",
+        "    AutoTokenizer,\n",
+        "    BitsAndBytesConfig,\n",
+        "    set_seed\n",
+        ")\n",
+        "from datasets import load_dataset\n",
+        "from peft import PeftModel\n",
+        "\n",
+        "%matplotlib inline"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "uuTX-xonNeOK"
+      },
+      "outputs": [],
+      "source": [
+        "BASE_MODEL = \"meta-llama/Meta-Llama-3.1-8B\"\n",
+        "DATASET_NAME = \"ed-donner/pricer-data\"\n",
+        "FINETUNED_MODEL = \"ed-donner/pricer-2024-09-13_13.04.39\"\n",
+        "REVISION = \"e8d637df551603dc86cd7a1598a8f44af4d7ae36\"\n",
+        "\n",
+        "TOP_K = 3\n",
+        "TEST_SIZE = 250\n",
+        "\n",
+        "GREEN = \"\\033[92m\"\n",
+        "YELLOW = \"\\033[93m\"\n",
+        "RED = \"\\033[91m\"\n",
+        "RESET = \"\\033[0m\"\n",
+        "COLOR_MAP = {\"red\": RED, \"orange\": YELLOW, \"green\": GREEN}"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "WyFPZeMcM88v"
+      },
+      "outputs": [],
+      "source": [
+        "from google.colab import userdata\n",
+        "hf_token = userdata.get('HF_TOKEN')\n",
+        "\n",
+        "login(hf_token, add_to_git_credential=True)\n",
+        "print(\"Successfully authenticated with HuggingFace\")"
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "qDqXth7MxMt0"
+      },
+      "source": [
+        "## Load Data"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "cvXVoJH8LS6u"
+      },
+      "outputs": [],
+      "source": [
+        "print(f\"Loading dataset: {DATASET_NAME}\")\n",
+        "dataset = load_dataset(DATASET_NAME)\n",
+        "train = dataset['train']\n",
+        "test = dataset['test']\n",
+        "\n",
+        "print(f\"\\nDataset loaded successfully:\")\n",
+        "print(f\"  Training examples: {len(train):,}\")\n",
+        "print(f\"  Test examples: {len(test):,}\")"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "xb86e__Wc7j_"
+      },
+      "outputs": [],
+      "source": [
+        "print(\"Sample test example:\\n\")\n",
+        "sample = test[0]\n",
+        "print(f\"Text: {sample['text'][:200]}...\")\n",
+        "print(f\"\\nGround truth price: ${sample['price']:.2f}\")"
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "qJWQ0a3wZ0Bw"
+      },
+      "source": [
+        "## Quantization & Model Loading\n",
+        "\n",
+        "(4-bit quantization reduces LLaMA 3.1 8B from ~32GB to ~5-6GB VRAM)"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "lAUAAcEC6ido"
+      },
+      "outputs": [],
+      "source": [
+        "quant_config = BitsAndBytesConfig(\n",
+        "    load_in_4bit=True,\n",
+        "    bnb_4bit_use_double_quant=True,\n",
+        "    bnb_4bit_compute_dtype=torch.bfloat16,\n",
+        "    bnb_4bit_quant_type=\"nf4\"\n",
+        ")\n",
+        "\n",
+        "print(\"Using 4-bit NF4 quantization\")"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "R_O04fKxMMT-"
+      },
+      "outputs": [],
+      "source": [
+        "print(f\"Loading base model: {BASE_MODEL}\")\n",
+        "\n",
+        "tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL, trust_remote_code=True)\n",
+        "tokenizer.pad_token = tokenizer.eos_token\n",
+        "tokenizer.padding_side = \"right\"\n",
+        "\n",
+        "base_model = AutoModelForCausalLM.from_pretrained(\n",
+        "    BASE_MODEL,\n",
+        "    quantization_config=quant_config,\n",
+        "    device_map=\"auto\",\n",
+        ")\n",
+        "base_model.generation_config.pad_token_id = tokenizer.pad_token_id\n",
+        "\n",
+        "print(f\"Base model loaded - Memory: {base_model.get_memory_footprint() / 1e9:.2f} GB\")"
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "m7PHUKDVxMt1"
+      },
+      "source": [
+        "## Load PEFT Adapters"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "6_Q1fqluxMt1"
+      },
+      "outputs": [],
+      "source": [
+        "print(f\"Loading fine-tuned adapters: {FINETUNED_MODEL}\")\n",
+        "print(f\"Revision: {REVISION}\")\n",
+        "\n",
+        "fine_tuned_model = PeftModel.from_pretrained(base_model, FINETUNED_MODEL, revision=REVISION)\n",
+        "\n",
+        "print(f\"Fine-tuned model ready - Total memory: {fine_tuned_model.get_memory_footprint() / 1e9:.2f} GB\")"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "Qst1LhBVAB04"
+      },
+      "outputs": [],
+      "source": [
+        "def extract_price(text):\n",
+        "    if \"Price is $\" in text:\n",
+        "        content = text.split(\"Price is $\")[1]\n",
+        "        content = content.replace(',', '').replace('$', '')\n",
+        "        match = re.search(r\"[-+]?\\d*\\.?\\d+\", content)\n",
+        "        return float(match.group()) if match else 0.0\n",
+        "    return 0.0"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "jXFBW_5UeEcp"
+      },
+      "outputs": [],
+      "source": [
+        "test_cases = [\n",
+        "    \"Price is $24.99\",\n",
+        "    \"Price is $1,234.50\",\n",
+        "    \"Price is $a fabulous 899.99 or so\"\n",
+        "]\n",
+        "\n",
+        "for test in test_cases:\n",
+        "    result = extract_price(test)\n",
+        "    print(f\"{test} -> ${result:.2f}\")"
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "TTy_WAGexMt2"
+      },
+      "source": [
+        "## Prediction Function\n",
+        "\n",
+        "Top-K weighted averaging computes probability-weighted average of top K tokens."
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "Je5dR8QEAI1d"
+      },
+      "outputs": [],
+      "source": [
+        "def advanced_predict(prompt, top_k=TOP_K):\n",
+        "    set_seed(42)\n",
+        "    inputs = tokenizer.encode(prompt, return_tensors=\"pt\").to(\"cuda\")\n",
+        "    attention_mask = torch.ones(inputs.shape, device=\"cuda\")\n",
+        "\n",
+        "    with torch.no_grad():\n",
+        "        outputs = fine_tuned_model(inputs, attention_mask=attention_mask)\n",
+        "        next_token_logits = outputs.logits[:, -1, :].to('cpu')\n",
+        "\n",
+        "    next_token_probs = F.softmax(next_token_logits, dim=-1)\n",
+        "    top_probs, top_token_ids = next_token_probs.topk(top_k)\n",
+        "\n",
+        "    prices, weights = [], []\n",
+        "\n",
+        "    for i in range(top_k):\n",
+        "        predicted_token = tokenizer.decode(top_token_ids[0][i])\n",
+        "        probability = top_probs[0][i]\n",
+        "\n",
+        "        try:\n",
+        "            price = float(predicted_token)\n",
+        "            if price > 0:\n",
+        "                prices.append(price)\n",
+        "                weights.append(probability)\n",
+        "        except ValueError:\n",
+        "            continue\n",
+        "\n",
+        "    if not prices:\n",
+        "        return 0.0\n",
+        "\n",
+        "    total_weight = sum(weights)\n",
+        "    weighted_avg = sum(p * w / total_weight for p, w in zip(prices, weights))\n",
+        "\n",
+        "    return weighted_avg.item()"
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "7nI3Ec7exMt2"
+      },
+      "source": [
+        "## Evaluation Framework\n",
+        "\n",
+        "Metrics:\n",
+        "- Dollar Error: |prediction - truth|\n",
+        "- RMSLE: Root Mean Squared Log Error (penalizes relative errors)\n",
+        "- Hit Rate: Percentage in green zone (error < $40 OR < 20% of true price)"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "30lzJXBH7BcK"
+      },
+      "outputs": [],
+      "source": [
+        "class Tester:\n",
+        "\n",
+        "    def __init__(self, predictor, data, title=None, size=TEST_SIZE):\n",
+        "        self.predictor = predictor\n",
+        "        self.data = data\n",
+        "        self.title = title or predictor.__name__.replace(\"_\", \" \").title()\n",
+        "        self.size = min(size, len(data))\n",
+        "        self.guesses = []\n",
+        "        self.truths = []\n",
+        "        self.errors = []\n",
+        "        self.sles = []\n",
+        "        self.colors = []\n",
+        "\n",
+        "    def color_for(self, error, truth):\n",
+        "        if error < 40 or error / truth < 0.2:\n",
+        "            return \"green\"\n",
+        "        elif error < 80 or error / truth < 0.4:\n",
+        "            return \"orange\"\n",
+        "        else:\n",
+        "            return \"red\"\n",
+        "\n",
+        "    def run_datapoint(self, i):\n",
+        "        datapoint = self.data[i]\n",
+        "        guess = self.predictor(datapoint[\"text\"])\n",
+        "        truth = datapoint[\"price\"]\n",
+        "        error = abs(guess - truth)\n",
+        "\n",
+        "        log_error = math.log(truth + 1) - math.log(guess + 1)\n",
+        "        sle = log_error ** 2\n",
+        "\n",
+        "        color = self.color_for(error, truth)\n",
+        "        title = datapoint[\"text\"].split(\"\\n\\n\")[1][:30] + \"...\"\n",
+        "\n",
+        "        self.guesses.append(guess)\n",
+        "        self.truths.append(truth)\n",
+        "        self.errors.append(error)\n",
+        "        self.sles.append(sle)\n",
+        "        self.colors.append(color)\n",
+        "\n",
+        "        print(f\"{COLOR_MAP[color]}{i+1}: Guess: ${guess:,.2f} | Truth: ${truth:,.2f} | Error: ${error:,.2f} | SLE: {sle:,.3f} | {title}{RESET}\")\n",
+        "\n",
+        "    def chart(self, title):\n",
+        "        plt.figure(figsize=(14, 10))\n",
+        "        max_val = max(max(self.truths), max(self.guesses))\n",
+        "\n",
+        "        plt.plot([0, max_val], [0, max_val], color='deepskyblue', lw=3, alpha=0.7, label='Perfect prediction')\n",
+        "        plt.scatter(self.truths, self.guesses, s=20, c=self.colors, alpha=0.6)\n",
+        "\n",
+        "        plt.xlabel('Ground Truth Price ($)', fontsize=12)\n",
+        "        plt.ylabel('Model Prediction ($)', fontsize=12)\n",
+        "        plt.xlim(0, max_val)\n",
+        "        plt.ylim(0, max_val)\n",
+        "        plt.title(title, fontsize=14, fontweight='bold')\n",
+        "        plt.grid(alpha=0.3)\n",
+        "        plt.legend()\n",
+        "        plt.tight_layout()\n",
+        "        plt.show()\n",
+        "\n",
+        "    def report(self):\n",
+        "        average_error = sum(self.errors) / self.size\n",
+        "        rmsle = math.sqrt(sum(self.sles) / self.size)\n",
+        "        hits = sum(1 for color in self.colors if color == \"green\")\n",
+        "        hit_rate = hits / self.size * 100\n",
+        "\n",
+        "        title = f\"{self.title} | Avg Error: ${average_error:,.2f} | RMSLE: {rmsle:.3f} | Hit Rate: {hit_rate:.1f}%\"\n",
+        "\n",
+        "        print(f\"\\n{'='*80}\")\n",
+        "        print(f\"EVALUATION SUMMARY\")\n",
+        "        print(f\"{'='*80}\")\n",
+        "        print(f\"Model: {self.title}\")\n",
+        "        print(f\"Test Size: {self.size}\")\n",
+        "        print(f\"Average Dollar Error: ${average_error:,.2f}\")\n",
+        "        print(f\"RMSLE: {rmsle:.4f}\")\n",
+        "        print(f\"Hit Rate (Green): {hit_rate:.2f}% ({hits}/{self.size})\")\n",
+        "        print(f\"{'='*80}\\n\")\n",
+        "\n",
+        "        self.chart(title)\n",
+        "\n",
+        "    def run(self):\n",
+        "        print(f\"Running evaluation on {self.size} examples...\\n\")\n",
+        "        for i in tqdm(range(self.size), desc=\"Evaluating\"):\n",
+        "            self.run_datapoint(i)\n",
+        "        self.report()\n",
+        "\n",
+        "    @classmethod\n",
+        "    def test(cls, function, data, **kwargs):\n",
+        "        cls(function, data, **kwargs).run()"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "krjnRyHO4Fp6"
+      },
+      "outputs": [],
+      "source": [
+        "print(f\"Loading dataset: {DATASET_NAME}\")\n",
+        "dataset = load_dataset(DATASET_NAME)\n",
+        "train = dataset['train']\n",
+        "test = dataset['test']\n",
+        "\n",
+        "print(f\"\\nDataset loaded successfully:\")\n",
+        "print(f\"  Training examples: {len(train):,}\")\n",
+        "print(f\"  Test examples: {len(test):,}\")"
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "5JjNwXMDxMt2"
+      },
+      "source": [
+        "## Run Evaluation"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "W_KcLvyt6kbb"
+      },
+      "outputs": [],
+      "source": [
+        "Tester.test(advanced_predict, test, title=\"LLaMA 3.1 8B QLoRA (400K)\")"
+      ]
+    }
+  ],
+  "metadata": {
+    "accelerator": "GPU",
+    "colab": {
+      "gpuType": "T4",
+      "provenance": []
+    },
+    "kernelspec": {
+      "display_name": "Python 3",
+      "name": "python3"
+    },
+    "language_info": {
+      "name": "python"
+    }
+  },
+  "nbformat": 4,
+  "nbformat_minor": 0
+}