week6/day2.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "id": "28a0673e-96b5-43f2-8a8b-bd033bf851b0",
   "metadata": {},
   "source": [
    "# The Product Pricer Continued\n",
    "\n",
    "A model that can estimate how much something costs, from its description.\n",
    "\n",
    "## Data Curation Part 2\n",
    "\n",
    "Today we'll extend our dataset to a greater coverage, and craft it into an excellent dataset for training.  \n",
    "Data curation can seem less exciting than other things we work on, but it's a crucial part of the LLM engineers' responsibility and an important craft to hone, so that you can build your own commercial solutions with high quality datasets.\n",
    "\n",
    "The dataset is here:  \n",
    "https://huggingface.co/datasets/McAuley-Lab/Amazon-Reviews-2023\n",
    "\n",
    "And the folder with all the product datasets is here:  \n",
    "https://huggingface.co/datasets/McAuley-Lab/Amazon-Reviews-2023/tree/main/raw/meta_categories\n",
    "\n",
    "## Important Note - read me first please\n",
    "\n",
    "We are about to craft a massive dataset of 400,000 items covering multiple types of product. In Week 7 we will be using this data to train our own model. It's a pretty big dataset, and depending on the GPU you select, training could take 20+ hours. It will be really good fun, but it could cost a few dollars in compute units.\n",
    "\n",
    "As an alternative, if you want to keep things quick & low cost, you can work with a smaller dataset focused only on Home Appliances. You'll be able to cover the same learning points; the results will be good -- not quite as good as the full dataset, but still pretty amazing! If you'd prefer to do this, I've set up an alternative jupyter notebook in this folder called `lite.ipynb` that you should use in place of this one.\n",
    "\n",
    "Also, if you'd prefer, you can shortcut running all this data curation by downloading the pickle files that we save in the last cell. The pickle files are available here: https://drive.google.com/drive/folders/1f_IZGybvs9o0J5sb3xmtTEQB3BXllzrW"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "67cedf85-8125-4322-998e-9375fe745597",
   "metadata": {},
   "outputs": [],
   "source": [
    "# imports\n",
    "\n",
    "import os\n",
    "import random\n",
    "from dotenv import load_dotenv\n",
    "from huggingface_hub import login\n",
    "from datasets import load_dataset, Dataset, DatasetDict\n",
    "import matplotlib.pyplot as plt\n",
    "from collections import Counter, defaultdict\n",
    "import numpy as np\n",
    "import pickle"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "7390a6aa-79cb-4dea-b6d7-de7e4b13e472",
   "metadata": {},
   "outputs": [],
   "source": [
    "# environment\n",
    "\n",
    "load_dotenv(override=True)\n",
    "os.environ['OPENAI_API_KEY'] = os.getenv('OPENAI_API_KEY', 'your-key-if-not-using-env')\n",
    "os.environ['ANTHROPIC_API_KEY'] = os.getenv('ANTHROPIC_API_KEY', 'your-key-if-not-using-env')\n",
    "os.environ['HF_TOKEN'] = os.getenv('HF_TOKEN', 'your-key-if-not-using-env')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "0732274a-aa6a-44fc-aee2-40dc8a8e4451",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Log in to HuggingFace\n",
    "\n",
    "hf_token = os.environ['HF_TOKEN']\n",
    "login(hf_token, add_to_git_credential=True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "6746144c-2e19-485a-8086-368c144722b4",
   "metadata": {},
   "outputs": [],
   "source": [
    "# More imports after HF login\n",
    "\n",
    "from loaders import ItemLoader\n",
    "from items import Item"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "1adcf323-de9d-4c24-a9c3-d7ae554d06ca",
   "metadata": {},
   "outputs": [],
   "source": [
    "%matplotlib inline"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "01065d69-765c-42c8-9f90-68b8c8754068",
   "metadata": {},
   "source": [
    "## The ItemLoader code\n",
    "\n",
    "Look in loaders.py - there's some useful code to make life easier for us"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "049885d4-fdfa-4ff0-a932-4a2ed73928e2",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Load in the same dataset as last time\n",
    "\n",
    "items = ItemLoader(\"Appliances\").load()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "ffba41b5-ddb6-4359-9790-9b2db900eee1",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Look for a familiar item..\n",
    "print(items[1].prompt)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "e2b6dc50-ac5c-4cf2-af2e-968ed8ef86d7",
   "metadata": {},
   "source": [
    "## Now to SCALE UP\n",
    "\n",
    "Let's look at all datasets of all the items that you might find in a large home retail store - electrical, electronic, office and related, but not clothes / beauty / books."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "d1d06cd3-f3c2-44f0-a9f2-13b54ff8be5c",
   "metadata": {},
   "outputs": [],
   "source": [
    "dataset_names = [\n",
    "    \"Automotive\",\n",
    "    \"Electronics\",\n",
    "    \"Office_Products\",\n",
    "    \"Tools_and_Home_Improvement\",\n",
    "    \"Cell_Phones_and_Accessories\",\n",
    "    \"Toys_and_Games\",\n",
    "    \"Appliances\",\n",
    "    \"Musical_Instruments\",\n",
    "]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "aa8fd0f0-509a-4298-8fcc-e499a061e1be",
   "metadata": {},
   "outputs": [],
   "source": [
    "items = []\n",
    "for dataset_name in dataset_names:\n",
    "    loader = ItemLoader(dataset_name)\n",
    "    items.extend(loader.load())\n",
    "\n",
    "# Now, time for a coffee break!!\n",
    "# By the way, I put the biggest datasets first.. it gets faster."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "3e29a5ab-ca61-41cc-9b33-22d374681b85",
   "metadata": {},
   "outputs": [],
   "source": [
    "print(f\"A grand total of {len(items):,} items\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "89078cb1-9679-4eb0-b295-599b8586bcd1",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Plot the distribution of token counts again\n",
    "\n",
    "tokens = [item.token_count for item in items]\n",
    "plt.figure(figsize=(15, 6))\n",
    "plt.title(f\"Token counts: Avg {sum(tokens)/len(tokens):,.1f} and highest {max(tokens):,}\\n\")\n",
    "plt.xlabel('Length (tokens)')\n",
    "plt.ylabel('Count')\n",
    "plt.hist(tokens, rwidth=0.7, color=\"skyblue\", bins=range(0, 300, 10))\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "c38e0c43-9f7a-450e-a911-c94d37d9b9c3",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Plot the distribution of prices\n",
    "\n",
    "prices = [item.price for item in items]\n",
    "plt.figure(figsize=(15, 6))\n",
    "plt.title(f\"Prices: Avg {sum(prices)/len(prices):,.1f} and highest {max(prices):,}\\n\")\n",
    "plt.xlabel('Price ($)')\n",
    "plt.ylabel('Count')\n",
    "plt.hist(prices, rwidth=0.7, color=\"blueviolet\", bins=range(0, 1000, 10))\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "eabc7c61-0cd2-41f4-baa1-b85400bbf87f",
   "metadata": {},
   "outputs": [],
   "source": [
    "category_counts = Counter()\n",
    "for item in items:\n",
    "    category_counts[item.category]+=1\n",
    "\n",
    "categories = category_counts.keys()\n",
    "counts = [category_counts[category] for category in categories]\n",
    "\n",
    "# Bar chart by category\n",
    "plt.figure(figsize=(15, 6))\n",
    "plt.bar(categories, counts, color=\"goldenrod\")\n",
    "plt.title('How many in each category')\n",
    "plt.xlabel('Categories')\n",
    "plt.ylabel('Count')\n",
    "\n",
    "plt.xticks(rotation=30, ha='right')\n",
    "\n",
    "# Add value labels on top of each bar\n",
    "for i, v in enumerate(counts):\n",
    "    plt.text(i, v, f\"{v:,}\", ha='center', va='bottom')\n",
    "\n",
    "# Display the chart\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "e5b6e987-83ba-4262-a082-57c6b0741062",
   "metadata": {},
   "source": [
    "# Objective\n",
    "\n",
    "Craft a dataset which is more balanced in terms of prices. Less heavily scewed to cheap items, with an average that's higher than $60. Try to balance out the categories - fewer Automotive items."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "3b9424c1-44e0-499a-b45e-a35246655469",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Create a dict with a key of each price from $1 to $999\n",
    "# And in the value, put a list of items with that price (to nearest round number)\n",
    "\n",
    "slots = defaultdict(list)\n",
    "for item in items:\n",
    "    slots[round(item.price)].append(item)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "7805a7f1-4ad8-48f6-bea3-d64b64894804",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Create a dataset called \"sample\" which tries to more evenly take from the range of prices\n",
    "# And gives more weight to items from categories other than Automotive\n",
    "# Set random seed for reproducibility\n",
    "\n",
    "np.random.seed(42)\n",
    "random.seed(42)\n",
    "sample = []\n",
    "for i in range(1, 1000):\n",
    "    slot = slots[i]\n",
    "    if i>=240:\n",
    "        sample.extend(slot)\n",
    "    elif len(slot) <= 1200:\n",
    "        sample.extend(slot)\n",
    "    else:\n",
    "        weights = np.array([1 if item.category=='Automotive' else 5 for item in slot])\n",
    "        weights = weights / np.sum(weights)\n",
    "        selected_indices = np.random.choice(len(slot), size=1200, replace=False, p=weights)\n",
    "        selected = [slot[i] for i in selected_indices]\n",
    "        sample.extend(selected)\n",
    "\n",
    "print(f\"There are {len(sample):,} items in the sample\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "430b432f-b769-41da-9506-a238cb5cf1b6",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Plot the distribution of prices in sample\n",
    "\n",
    "prices = [float(item.price) for item in sample]\n",
    "plt.figure(figsize=(15, 10))\n",
    "plt.title(f\"Avg {sum(prices)/len(prices):.2f} and highest {max(prices):,.2f}\\n\")\n",
    "plt.xlabel('Price ($)')\n",
    "plt.ylabel('Count')\n",
    "plt.hist(prices, rwidth=0.7, color=\"darkblue\", bins=range(0, 1000, 10))\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "0d570794-6f1d-462e-b567-a46bae3556a1",
   "metadata": {},
   "outputs": [],
   "source": [
    "# OK, we did well in terms of raising the average price and having a smooth-ish population of prices\n",
    "# Let's see the categories\n",
    "\n",
    "category_counts = Counter()\n",
    "for item in sample:\n",
    "    category_counts[item.category]+=1\n",
    "\n",
    "categories = category_counts.keys()\n",
    "counts = [category_counts[category] for category in categories]\n",
    "\n",
    "# Create bar chart\n",
    "plt.figure(figsize=(15, 6))\n",
    "plt.bar(categories, counts, color=\"lightgreen\")\n",
    "\n",
    "# Customize the chart\n",
    "plt.title('How many in each category')\n",
    "plt.xlabel('Categories')\n",
    "plt.ylabel('Count')\n",
    "\n",
    "plt.xticks(rotation=30, ha='right')\n",
    "\n",
    "# Add value labels on top of each bar\n",
    "for i, v in enumerate(counts):\n",
    "    plt.text(i, v, f\"{v:,}\", ha='center', va='bottom')\n",
    "\n",
    "# Display the chart\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "6609d77c-3e0a-4679-9129-c7cdc3273070",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Automotive still in the lead, but improved somewhat\n",
    "# For another perspective, let's look at a pie\n",
    "\n",
    "plt.figure(figsize=(12, 10))\n",
    "plt.pie(counts, labels=categories, autopct='%1.0f%%', startangle=90)\n",
    "\n",
    "# Add a circle at the center to create a donut chart (optional)\n",
    "centre_circle = plt.Circle((0,0), 0.70, fc='white')\n",
    "fig = plt.gcf()\n",
    "fig.gca().add_artist(centre_circle)\n",
    "plt.title('Categories')\n",
    "\n",
    "# Equal aspect ratio ensures that pie is drawn as a circle\n",
    "plt.axis('equal')  \n",
    "\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "ac046cc1-2717-415b-96ad-b73b2950d235",
   "metadata": {},
   "source": [
    "# Dataset Curated!\n",
    "\n",
    "We've crafted an excellent dataset.\n",
    "\n",
    "Let's do some final checks"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "70219e99-22cc-4e08-9121-51f9707caef0",
   "metadata": {},
   "outputs": [],
   "source": [
    "# How does the price vary with the character count of the prompt?\n",
    "\n",
    "sizes = [len(item.prompt) for item in sample]\n",
    "prices = [item.price for item in sample]\n",
    "\n",
    "# Create the scatter plot\n",
    "plt.figure(figsize=(15, 8))\n",
    "plt.scatter(sizes, prices, s=0.2, color=\"red\")\n",
    "\n",
    "# Add labels and title\n",
    "plt.xlabel('Size')\n",
    "plt.ylabel('Price')\n",
    "plt.title('Is there a simple correlation?')\n",
    "\n",
    "# Display the plot\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "30ae1453-b9fc-40db-8310-65d850c4b1da",
   "metadata": {},
   "outputs": [],
   "source": [
    "def report(item):\n",
    "    prompt = item.prompt\n",
    "    tokens = Item.tokenizer.encode(item.prompt)\n",
    "    print(prompt)\n",
    "    print(tokens[-10:])\n",
    "    print(Item.tokenizer.batch_decode(tokens[-10:]))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "d9998b8d-d746-4541-9ac2-701108e0e8fb",
   "metadata": {},
   "outputs": [],
   "source": [
    "report(sample[398000])"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "7aa0a3fc-d2fe-4e6e-8fdb-96913df2f588",
   "metadata": {},
   "source": [
    "## Observation\n",
    "\n",
    "An interesting thing about the Llama tokenizer is that every number from 1 to 999 gets mapped to 1 token, much as we saw with gpt-4o. The same is not true of qwen2, gemma and phi3, which all map individual digits to tokens. This does turn out to be a bit useful for our project, although it's not an essential requirement."
   ]
  },
  {
   "cell_type": "markdown",
   "id": "0f03c0ee-3103-4603-af5c-b484884a3aa2",
   "metadata": {},
   "source": [
    "# Finally\n",
    "\n",
    "It's time to break down our data into a training, test and validation dataset.\n",
    "\n",
    "It's typical to use 5%-10% of your data for testing purposes, but actually we have far more than we need at this point. We'll take 400,000 points for training, and we'll reserve 2,000 for testing, although we won't use all of them.\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "3b163ca2-18ef-4c26-8e9d-88eb55f114f6",
   "metadata": {},
   "outputs": [],
   "source": [
    "random.seed(42)\n",
    "random.shuffle(sample)\n",
    "train = sample[:400_000]\n",
    "test = sample[400_000:402_000]\n",
    "print(f\"Divided into a training set of {len(train):,} items and test set of {len(test):,} items\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "299b9816-8885-4798-829a-69d66d60eb01",
   "metadata": {},
   "outputs": [],
   "source": [
    "print(train[0].prompt)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "97222da3-9f2c-4d15-a5cd-5e5f8dbde6cc",
   "metadata": {},
   "outputs": [],
   "source": [
    "print(test[0].test_prompt())"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "7a116369-335a-412b-b70c-2add6675c2e3",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Plot the distribution of prices in the first 250 test points\n",
    "\n",
    "prices = [float(item.price) for item in test[:250]]\n",
    "plt.figure(figsize=(15, 6))\n",
    "plt.title(f\"Avg {sum(prices)/len(prices):.2f} and highest {max(prices):,.2f}\\n\")\n",
    "plt.xlabel('Price ($)')\n",
    "plt.ylabel('Count')\n",
    "plt.hist(prices, rwidth=0.7, color=\"darkblue\", bins=range(0, 1000, 10))\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "d522d752-6f66-4786-a4dc-8ef51842558c",
   "metadata": {},
   "source": [
    "# Finally - upload your brand new dataset\n",
    "\n",
    "Convert to prompts and upload to HuggingFace hub"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "fa11b3e5-fcf4-4efc-a573-f6f67fec3e73",
   "metadata": {},
   "outputs": [],
   "source": [
    "train_prompts = [item.prompt for item in train]\n",
    "train_prices = [item.price for item in train]\n",
    "test_prompts = [item.test_prompt() for item in test]\n",
    "test_prices = [item.price for item in test]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "b020ab1b-7153-4e5f-b8a3-d5bc2fafb6df",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Create a Dataset from the lists\n",
    "\n",
    "train_dataset = Dataset.from_dict({\"text\": train_prompts, \"price\": train_prices})\n",
    "test_dataset = Dataset.from_dict({\"text\": test_prompts, \"price\": test_prices})\n",
    "dataset = DatasetDict({\n",
    "    \"train\": train_dataset,\n",
    "    \"test\": test_dataset\n",
    "})"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "17639641-fb55-44e2-a463-b0b394d00f32",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Uncomment these lines if you're ready to push to the hub, and replace my name with your HF username\n",
    "\n",
    "# HF_USER = \"ed-donner\"\n",
    "# DATASET_NAME = f\"{HF_USER}/pricer-data\"\n",
    "# dataset.push_to_hub(DATASET_NAME, private=True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "b85733ba-d165-4f07-b055-46803543edfe",
   "metadata": {},
   "outputs": [],
   "source": [
    "# One more thing!\n",
    "# Let's pickle the training and test dataset so we don't have to execute all this code next time!\n",
    "\n",
    "with open('train.pkl', 'wb') as file:\n",
    "    pickle.dump(train, file)\n",
    "\n",
    "with open('test.pkl', 'wb') as file:\n",
    "    pickle.dump(test, file)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "2b58dc61-747f-46f7-b9e0-c205db4f3e5e",
   "metadata": {},
   "source": [
    "## Todos for you:\n",
    "\n",
    "- Investigate the dataset more!\n",
    "- Confirm that the tokenizer tokenizes all 3 digit prices into 1 token"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3 (ipykernel)",
   "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": "3.11.12"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}