{
 "cells": [
  {
   "cell_type": "markdown",
   "id": "17c55a41",
   "metadata": {},
   "source": [
    "<br>\n",
    "<hr>\n",
    "<hr>\n",
    "<h1>Conversational AI: Assignment-2_PS1_SM</h1>\n",
    "<h3>Group 85</h3>\n",
    "<hr>\n",
    "<hr>\n",
    "<br>"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "433726b7",
   "metadata": {},
   "source": [
    "<h2>TEAM MEMBERS</h2>\n",
    "\n",
    "<table>\n",
    "  <tr>\n",
    "    <th>Name</th>\n",
    "    <th>BITS ID</th>\n",
    "  </tr>\n",
    "  <tr>\n",
    "    <td>R. Piji Rajendran</td>\n",
    "    <td>2024AD05222</td>\n",
    "  </tr>\n",
    "  <tr>\n",
    "    <td>Reeve Chaitanya</td>\n",
    "    <td>2024AD05225</td>\n",
    "  </tr>\n",
    "  <tr>\n",
    "    <td>Sahil Verma</td>\n",
    "    <td>2024AD05230</td>\n",
    "  </tr>\n",
    "  <tr>\n",
    "    <td>Vankala N Sai Krishna Kumar</td>\n",
    "    <td>2024AD05334</td>\n",
    "  </tr>\n",
    "</table>"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "f1944b2f",
   "metadata": {},
   "source": [
    "# Knowledge Distillation on `PolyAI/banking77`\n",
    "\n",
    "**Assignment:** Compress a high-capacity fine-tuned Transformer (\"Teacher\") into a compact, CPU-friendly\n",
    "Transformer (\"Student\") via knowledge distillation, and quantify what is gained and lost in the process.\n",
    "\n",
    "**Dataset:** [`PolyAI/banking77`](https://huggingface.co/datasets/PolyAI/banking77) — 77-way fine-grained\n",
    "banking-intent classification, 10,003 train / 3,080 test utterances.\n",
    "\n",
    "**Environment:** conda env `agn_env` (Python 3.12), Apple Silicon (MPS acceleration used for teacher\n",
    "fine-tuning; the student is trained and benchmarked on **CPU only**, matching its target deployment profile).\n",
    "\n",
    "| Module | Tasks |\n",
    "|---|---|\n",
    "| 1. Teacher Labeling & Student Setup | Task 1: Teacher fine-tuning & soft-label generation · Task 2: Student tokenizer alignment |\n",
    "| 2. Distillation Architecture & Training | Task 3: Compact student transformer · Task 4: Distillation loss & training |\n",
    "| 3. Comparative Analysis & Benchmarking | Task 5: Accuracy vs. compression · Task 6: Deployment metrics |\n"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "5f8c80fe",
   "metadata": {},
   "source": [
    "## Setup\n",
    "\n",
    "Install dependencies directly into the active `agn_env` environment (safe to re-run — a no-op if already satisfied), then import everything used below."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "id": "00ea94f7",
   "metadata": {
    "execution": {
     "iopub.execute_input": "2026-08-12T16:11:09.681972Z",
     "iopub.status.busy": "2026-08-12T16:11:09.681599Z",
     "iopub.status.idle": "2026-08-12T16:11:10.400082Z",
     "shell.execute_reply": "2026-08-12T16:11:10.399587Z"
    }
   },
   "outputs": [],
   "source": [
    "# Ensure required libraries are present in the active (agn_env) kernel.\n",
    "import sys\n",
    "!{sys.executable} -m pip install -q torch transformers datasets scikit-learn psutil evaluate accelerate tokenizers ipykernel matplotlib\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "id": "7ba294de",
   "metadata": {
    "execution": {
     "iopub.execute_input": "2026-08-12T16:11:10.401717Z",
     "iopub.status.busy": "2026-08-12T16:11:10.401620Z",
     "iopub.status.idle": "2026-08-12T16:11:15.250467Z",
     "shell.execute_reply": "2026-08-12T16:11:15.250099Z"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Teacher training device : mps\n",
      "Student device (always) : cpu\n"
     ]
    }
   ],
   "source": [
    "import os\n",
    "import gc\n",
    "import json\n",
    "import time\n",
    "import random\n",
    "import tempfile\n",
    "import subprocess\n",
    "from collections import OrderedDict\n",
    "\n",
    "import numpy as np\n",
    "import pandas as pd\n",
    "import torch\n",
    "import torch.nn as nn\n",
    "import torch.nn.functional as F\n",
    "from torch.utils.data import Dataset, DataLoader\n",
    "import matplotlib.pyplot as plt\n",
    "from IPython.display import display, Markdown\n",
    "\n",
    "from datasets import load_dataset\n",
    "from transformers import (\n",
    "    AutoTokenizer,\n",
    "    BertForSequenceClassification,\n",
    "    TrainingArguments,\n",
    "    Trainer,\n",
    "    DataCollatorWithPadding,\n",
    ")\n",
    "from tokenizers import Tokenizer as HFTokenizer\n",
    "from tokenizers.models import WordPiece\n",
    "from tokenizers.trainers import WordPieceTrainer\n",
    "from tokenizers.pre_tokenizers import Whitespace\n",
    "from tokenizers.normalizers import BertNormalizer\n",
    "from tokenizers.processors import TemplateProcessing\n",
    "\n",
    "from sklearn.metrics import accuracy_score, f1_score\n",
    "\n",
    "import psutil\n",
    "\n",
    "SEED = 42\n",
    "random.seed(SEED)\n",
    "np.random.seed(SEED)\n",
    "torch.manual_seed(SEED)\n",
    "\n",
    "TRAIN_DEVICE = torch.device(\"mps\") if torch.backends.mps.is_available() else torch.device(\"cpu\")\n",
    "CPU_DEVICE = torch.device(\"cpu\")\n",
    "STUDENT_DEVICE = CPU_DEVICE  # student is trained + benchmarked on CPU throughout: this is the point of the exercise\n",
    "\n",
    "print(f\"Teacher training device : {TRAIN_DEVICE}\")\n",
    "print(f\"Student device (always) : {STUDENT_DEVICE}\")\n"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "0b317e8b",
   "metadata": {},
   "source": [
    "<br>\n",
    "<hr>\n",
    "<h1>Module 1: Teacher Labeling & Student Setup</h1>\n",
    "<hr>\n",
    "<br>"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "50527830",
   "metadata": {},
   "source": [
    "### Task 1 — Teacher Integration & Soft-Label Generation\n",
    "\n",
    "We fine-tune `bert-base-uncased` (110M parameters) end-to-end on banking77 as our **Teacher**. Once\n",
    "fine-tuned, we run one no-grad forward pass over the *entire* training set to cache the Teacher's raw\n",
    "77-dimensional **logits** for every example. These cached logits are the \"soft labels\" used by the\n",
    "distillation loss in Task 4 — they are computed once, up front, so the (comparatively expensive) Teacher\n",
    "never needs to run again during Student training."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "id": "77f6b15b",
   "metadata": {
    "execution": {
     "iopub.execute_input": "2026-08-12T16:11:15.251761Z",
     "iopub.status.busy": "2026-08-12T16:11:15.251589Z",
     "iopub.status.idle": "2026-08-12T16:11:18.822451Z",
     "shell.execute_reply": "2026-08-12T16:11:18.821517Z"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Could not load PolyAI/banking77 directly (RuntimeError); falling back to the parquet mirror legacy-datasets/banking77.\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Loaded from : legacy-datasets/banking77 (parquet mirror of PolyAI/banking77)\n",
      "Train examples : 10003\n",
      "Test examples  : 3080\n",
      "Classes        : 77\n",
      "\n",
      "Sample row: {'text': 'I am still waiting on my card?', 'label': 11}\n",
      "First 10 intents: ['activate_my_card', 'age_limit', 'apple_pay_or_google_pay', 'atm_support', 'automatic_top_up', 'balance_not_updated_after_bank_transfer', 'balance_not_updated_after_cheque_or_cash_deposit', 'beneficiary_not_allowed', 'cancel_transfer', 'card_about_to_expire']\n"
     ]
    }
   ],
   "source": [
    "# Load banking77. The original PolyAI/banking77 repo only ships a loading *script*, which recent\n",
    "# versions of `datasets` (>=4.0) no longer execute. We fall back to a verified parquet mirror with\n",
    "# identical contents (10,003 train / 3,080 test / same 77 ClassLabel names) if the script path fails.\n",
    "try:\n",
    "    raw_datasets = load_dataset(\"PolyAI/banking77\")\n",
    "    dataset_source = \"PolyAI/banking77\"\n",
    "except Exception as e:\n",
    "    print(f\"Could not load PolyAI/banking77 directly ({type(e).__name__}); \"\n",
    "          f\"falling back to the parquet mirror legacy-datasets/banking77.\")\n",
    "    raw_datasets = load_dataset(\"legacy-datasets/banking77\")\n",
    "    dataset_source = \"legacy-datasets/banking77 (parquet mirror of PolyAI/banking77)\"\n",
    "\n",
    "train_raw = raw_datasets[\"train\"]\n",
    "test_raw = raw_datasets[\"test\"]\n",
    "label_names = train_raw.features[\"label\"].names\n",
    "num_labels = len(label_names)\n",
    "\n",
    "print(f\"Loaded from : {dataset_source}\")\n",
    "print(f\"Train examples : {len(train_raw)}\")\n",
    "print(f\"Test examples  : {len(test_raw)}\")\n",
    "print(f\"Classes        : {num_labels}\")\n",
    "print(f\"\\nSample row: {train_raw[0]}\")\n",
    "print(f\"First 10 intents: {label_names[:10]}\")\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "id": "ff9c585b",
   "metadata": {
    "execution": {
     "iopub.execute_input": "2026-08-12T16:11:18.824044Z",
     "iopub.status.busy": "2026-08-12T16:11:18.823920Z",
     "iopub.status.idle": "2026-08-12T16:11:20.635226Z",
     "shell.execute_reply": "2026-08-12T16:11:20.634838Z"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Tokenized train/test sets ready for the Teacher.\n"
     ]
    }
   ],
   "source": [
    "TEACHER_NAME = \"bert-base-uncased\"\n",
    "teacher_tokenizer = AutoTokenizer.from_pretrained(TEACHER_NAME)\n",
    "\n",
    "def teacher_tokenize(batch):\n",
    "    return teacher_tokenizer(batch[\"text\"], truncation=True, max_length=64)\n",
    "\n",
    "train_enc = train_raw.map(teacher_tokenize, batched=True)\n",
    "test_enc = test_raw.map(teacher_tokenize, batched=True)\n",
    "\n",
    "train_enc = train_enc.rename_column(\"label\", \"labels\")\n",
    "test_enc = test_enc.rename_column(\"label\", \"labels\")\n",
    "train_enc.set_format(type=\"torch\", columns=[\"input_ids\", \"attention_mask\", \"labels\"])\n",
    "test_enc.set_format(type=\"torch\", columns=[\"input_ids\", \"attention_mask\", \"labels\"])\n",
    "\n",
    "teacher_data_collator = DataCollatorWithPadding(tokenizer=teacher_tokenizer)\n",
    "print(\"Tokenized train/test sets ready for the Teacher.\")\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "id": "7792c6af",
   "metadata": {
    "execution": {
     "iopub.execute_input": "2026-08-12T16:11:20.636428Z",
     "iopub.status.busy": "2026-08-12T16:11:20.636352Z",
     "iopub.status.idle": "2026-08-12T16:17:32.367956Z",
     "shell.execute_reply": "2026-08-12T16:17:32.367509Z"
    }
   },
   "outputs": [
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "414f90d2277a489caf1c79d0b7c4b85e",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "Loading weights:   0%|          | 0/199 [00:00<?, ?it/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "[transformers] \u001b[1mBertForSequenceClassification LOAD REPORT\u001b[0m from: bert-base-uncased\n",
      "Key                                        | Status     | \n",
      "-------------------------------------------+------------+-\n",
      "cls.seq_relationship.weight                | UNEXPECTED | \n",
      "cls.predictions.bias                       | UNEXPECTED | \n",
      "cls.predictions.transform.dense.weight     | UNEXPECTED | \n",
      "cls.seq_relationship.bias                  | UNEXPECTED | \n",
      "cls.predictions.transform.LayerNorm.weight | UNEXPECTED | \n",
      "cls.predictions.transform.dense.bias       | UNEXPECTED | \n",
      "cls.predictions.transform.LayerNorm.bias   | UNEXPECTED | \n",
      "classifier.weight                          | MISSING    | \n",
      "classifier.bias                            | MISSING    | \n",
      "\n",
      "Notes:\n",
      "- UNEXPECTED:\tcan be ignored when loading from different task/architecture; not ok if you expect identical arch.\n",
      "- MISSING:\tthose params were newly initialized because missing from the checkpoint. Consider training on your downstream task.\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/opt/homebrew/Caskroom/miniforge/base/envs/agn_env/lib/python3.12/site-packages/torch/utils/data/dataloader.py:759: UserWarning: 'pin_memory' argument is set as true but not supported on MPS now, device pinned memory won't be used.\n",
      "  super().__init__(loader)\n"
     ]
    },
    {
     "data": {
      "text/html": [
       "\n",
       "    <div>\n",
       "      \n",
       "      <progress value='939' max='939' style='width:300px; height:20px; vertical-align: middle;'></progress>\n",
       "      [939/939 06:10, Epoch 3/3]\n",
       "    </div>\n",
       "    <table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       " <tr style=\"text-align: left;\">\n",
       "      <th>Epoch</th>\n",
       "      <th>Training Loss</th>\n",
       "      <th>Validation Loss</th>\n",
       "      <th>Accuracy</th>\n",
       "      <th>Macro F1</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <td>1</td>\n",
       "      <td>2.285605</td>\n",
       "      <td>1.966460</td>\n",
       "      <td>0.727273</td>\n",
       "      <td>0.694643</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>2</td>\n",
       "      <td>1.147723</td>\n",
       "      <td>1.047979</td>\n",
       "      <td>0.840909</td>\n",
       "      <td>0.823815</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>3</td>\n",
       "      <td>0.827817</td>\n",
       "      <td>0.838797</td>\n",
       "      <td>0.875649</td>\n",
       "      <td>0.867707</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table><p>"
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/opt/homebrew/Caskroom/miniforge/base/envs/agn_env/lib/python3.12/site-packages/torch/utils/data/dataloader.py:759: UserWarning: 'pin_memory' argument is set as true but not supported on MPS now, device pinned memory won't be used.\n",
      "  super().__init__(loader)\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/opt/homebrew/Caskroom/miniforge/base/envs/agn_env/lib/python3.12/site-packages/torch/utils/data/dataloader.py:759: UserWarning: 'pin_memory' argument is set as true but not supported on MPS now, device pinned memory won't be used.\n",
      "  super().__init__(loader)\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n",
      "Teacher fine-tuning took 6.2 minutes\n"
     ]
    }
   ],
   "source": [
    "teacher_model = BertForSequenceClassification.from_pretrained(TEACHER_NAME, num_labels=num_labels)\n",
    "\n",
    "def compute_metrics(eval_pred):\n",
    "    logits, labels = eval_pred\n",
    "    preds = np.argmax(logits, axis=-1)\n",
    "    return {\n",
    "        \"accuracy\": accuracy_score(labels, preds),\n",
    "        \"macro_f1\": f1_score(labels, preds, average=\"macro\"),\n",
    "    }\n",
    "\n",
    "teacher_training_args = TrainingArguments(\n",
    "    output_dir=\"./teacher_ckpt\",\n",
    "    num_train_epochs=3,\n",
    "    per_device_train_batch_size=32,\n",
    "    per_device_eval_batch_size=64,\n",
    "    learning_rate=3e-5,\n",
    "    weight_decay=0.01,\n",
    "    eval_strategy=\"epoch\",\n",
    "    save_strategy=\"no\",\n",
    "    logging_steps=50,\n",
    "    report_to=\"none\",\n",
    "    seed=SEED,\n",
    ")\n",
    "\n",
    "trainer = Trainer(\n",
    "    model=teacher_model,\n",
    "    args=teacher_training_args,\n",
    "    train_dataset=train_enc,\n",
    "    eval_dataset=test_enc,\n",
    "    data_collator=teacher_data_collator,\n",
    "    compute_metrics=compute_metrics,\n",
    ")\n",
    "\n",
    "teacher_train_start = time.time()\n",
    "trainer.train()\n",
    "teacher_train_seconds = time.time() - teacher_train_start\n",
    "print(f\"\\nTeacher fine-tuning took {teacher_train_seconds/60:.1f} minutes\")\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "id": "ac970e37",
   "metadata": {
    "execution": {
     "iopub.execute_input": "2026-08-12T16:17:32.369373Z",
     "iopub.status.busy": "2026-08-12T16:17:32.369284Z",
     "iopub.status.idle": "2026-08-12T16:17:37.620805Z",
     "shell.execute_reply": "2026-08-12T16:17:37.620383Z"
    }
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/opt/homebrew/Caskroom/miniforge/base/envs/agn_env/lib/python3.12/site-packages/torch/utils/data/dataloader.py:759: UserWarning: 'pin_memory' argument is set as true but not supported on MPS now, device pinned memory won't be used.\n",
      "  super().__init__(loader)\n"
     ]
    },
    {
     "data": {
      "text/html": [
       "\n",
       "    <div>\n",
       "      \n",
       "      <progress value='49' max='49' style='width:300px; height:20px; vertical-align: middle;'></progress>\n",
       "      [49/49 00:05]\n",
       "    </div>\n",
       "    "
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "text/html": [
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       " <tr style=\"text-align: left;\">\n",
       "      <th>Training Loss</th>\n",
       "      <th>Validation Loss</th>\n",
       "      <th>Epoch</th>\n",
       "      <th>Accuracy</th>\n",
       "      <th>Macro F1</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <td>0.827817</td>\n",
       "      <td>0.838797</td>\n",
       "      <td>3</td>\n",
       "      <td>0.875649</td>\n",
       "      <td>0.867707</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table><p>"
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Teacher eval metrics (test set): {'eval_loss': 0.8387965559959412, 'eval_accuracy': 0.8756493506493507, 'eval_macro_f1': 0.8677074681344383}\n"
     ]
    }
   ],
   "source": [
    "teacher_eval_metrics = trainer.evaluate()\n",
    "print(\"Teacher eval metrics (test set):\", teacher_eval_metrics)\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "id": "020eeca7",
   "metadata": {
    "execution": {
     "iopub.execute_input": "2026-08-12T16:17:37.622023Z",
     "iopub.status.busy": "2026-08-12T16:17:37.621953Z",
     "iopub.status.idle": "2026-08-12T16:17:59.886883Z",
     "shell.execute_reply": "2026-08-12T16:17:59.886409Z"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Cached teacher train logits: torch.Size([10003, 77])\n",
      "Cached teacher test logits : torch.Size([3080, 77])\n"
     ]
    }
   ],
   "source": [
    "# Cache soft labels: one no-grad forward pass over the FULL training set, in original (unshuffled)\n",
    "# order, so teacher_train_logits[i] corresponds exactly to train_raw[i] by index.\n",
    "teacher_model.eval()\n",
    "teacher_model.to(TRAIN_DEVICE)\n",
    "\n",
    "@torch.no_grad()\n",
    "def get_teacher_logits(dataset, batch_size=64):\n",
    "    loader = DataLoader(dataset, batch_size=batch_size, shuffle=False, collate_fn=teacher_data_collator)\n",
    "    all_logits = []\n",
    "    for batch in loader:\n",
    "        inputs = {k: v.to(TRAIN_DEVICE) for k, v in batch.items() if k != \"labels\"}\n",
    "        outputs = teacher_model(**inputs)\n",
    "        all_logits.append(outputs.logits.detach().cpu())\n",
    "    return torch.cat(all_logits, dim=0)\n",
    "\n",
    "teacher_train_logits = get_teacher_logits(train_enc)\n",
    "teacher_test_logits = get_teacher_logits(test_enc)\n",
    "\n",
    "print(f\"Cached teacher train logits: {teacher_train_logits.shape}\")\n",
    "print(f\"Cached teacher test logits : {teacher_test_logits.shape}\")\n",
    "assert teacher_train_logits.shape[0] == len(train_raw)\n"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "81336e0d",
   "metadata": {},
   "source": [
    "#### Why soft labels carry \"dark knowledge\"\n",
    "\n",
    "A one-hot hard label for the utterance *\"I am still waiting on my card?\"* says only: **the correct class is\n",
    "`card_arrival`, and every other one of the 77 classes is equally, absolutely wrong.** That is not true, and\n",
    "it is not what the Teacher actually believes. The Teacher's full softmax distribution over 77 classes might\n",
    "look like `card_arrival: 0.62, card_delivery_estimate: 0.21, lost_or_stolen_card: 0.05, ...` — it correctly\n",
    "picks `card_arrival`, but it also encodes *how confusable* the other intents are with it.\n",
    "\n",
    "This matters a great deal on a fine-grained, semantically overlapping taxonomy like banking77, which\n",
    "contains many near-duplicate intents (`card_arrival` vs. `card_delivery_estimate`, `declined_card_payment`\n",
    "vs. `declined_cash_withdrawal`, `top_up_failed` vs. `pending_top_up`, ...). Training only against a hard\n",
    "label throws away exactly the information that describes *why* those pairs are confusable — the relative\n",
    "geometry of the Teacher's learned decision boundary. Training against the full distribution instead:\n",
    "\n",
    "1. **Transfers inter-class similarity structure.** The relative magnitudes of the non-argmax probabilities\n",
    "   act as a learned \"confusion prior\" — pairs of intents the Teacher finds similar get correlated soft\n",
    "   targets across many training examples, which is a much richer training signal than 76 identical zeros.\n",
    "2. **Provides a smoother, higher-entropy target**, which acts as an implicit regularizer: the Student is not\n",
    "   forced to drive its logits to ±∞ to satisfy a one-hot target, so its learned representations generalize\n",
    "   better on held-out data, especially with the very limited parameter budget a compact Student has.\n",
    "3. **Effectively gives more supervision per example.** A hard label is `log2(77) ≈ 6.3` bits of information at\n",
    "   best (which class). A soft label is a full probability vector — many more effective bits — so a Student\n",
    "   with less capacity and less data than the Teacher can still recover much of the Teacher's decision\n",
    "   surface from the *same* training set.\n",
    "4. **Raising the temperature $T$ before softmax** (used in Task 4) further amplifies the small\n",
    "   probabilities on non-target classes, which is precisely where most of this structural information lives —\n",
    "   at $T{=}1$, those probabilities are so close to zero that gradients from them barely register.\n"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "5949b6af",
   "metadata": {},
   "source": [
    "### Task 2 — Student Tokenizer Alignment\n",
    "\n",
    "The Student will use its **own** compact WordPiece tokenizer (trained from scratch on the banking77 corpus,\n",
    "with a vocabulary roughly 10x smaller than BERT's), rather than reusing the Teacher's `bert-base-uncased`\n",
    "tokenizer. This section builds that tokenizer and demonstrates how it segments text differently from the\n",
    "Teacher's."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "id": "29359069",
   "metadata": {
    "execution": {
     "iopub.execute_input": "2026-08-12T16:17:59.888254Z",
     "iopub.status.busy": "2026-08-12T16:17:59.888169Z",
     "iopub.status.idle": "2026-08-12T16:18:00.004711Z",
     "shell.execute_reply": "2026-08-12T16:18:00.004335Z"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n",
      "\n",
      "\n",
      "Student vocab size : 3000\n",
      "Teacher vocab size  : 30522\n",
      "Vocab compression   : 10.2x smaller\n"
     ]
    }
   ],
   "source": [
    "student_tok_backend = HFTokenizer(WordPiece(unk_token=\"[UNK]\"))\n",
    "student_tok_backend.normalizer = BertNormalizer(lowercase=True)\n",
    "student_tok_backend.pre_tokenizer = Whitespace()\n",
    "\n",
    "STUDENT_VOCAB_SIZE_TARGET = 3000\n",
    "wp_trainer = WordPieceTrainer(\n",
    "    vocab_size=STUDENT_VOCAB_SIZE_TARGET,\n",
    "    special_tokens=[\"[PAD]\", \"[UNK]\", \"[CLS]\", \"[SEP]\"],\n",
    "    min_frequency=1,\n",
    ")\n",
    "\n",
    "corpus_texts = train_raw[\"text\"]\n",
    "student_tok_backend.train_from_iterator(corpus_texts, wp_trainer)\n",
    "\n",
    "student_tok_backend.post_processor = TemplateProcessing(\n",
    "    single=\"[CLS] $A [SEP]\",\n",
    "    special_tokens=[\n",
    "        (\"[CLS]\", student_tok_backend.token_to_id(\"[CLS]\")),\n",
    "        (\"[SEP]\", student_tok_backend.token_to_id(\"[SEP]\")),\n",
    "    ],\n",
    ")\n",
    "\n",
    "STUDENT_VOCAB_SIZE = student_tok_backend.get_vocab_size()\n",
    "STUDENT_PAD_ID = student_tok_backend.token_to_id(\"[PAD]\")\n",
    "STUDENT_MAX_LEN = 32\n",
    "\n",
    "print(f\"Student vocab size : {STUDENT_VOCAB_SIZE}\")\n",
    "print(f\"Teacher vocab size  : {teacher_tokenizer.vocab_size}\")\n",
    "print(f\"Vocab compression   : {teacher_tokenizer.vocab_size / STUDENT_VOCAB_SIZE:.1f}x smaller\")\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "id": "ae1e33d1",
   "metadata": {
    "execution": {
     "iopub.execute_input": "2026-08-12T16:18:00.006047Z",
     "iopub.status.busy": "2026-08-12T16:18:00.005956Z",
     "iopub.status.idle": "2026-08-12T16:18:00.015538Z",
     "shell.execute_reply": "2026-08-12T16:18:00.015188Z"
    }
   },
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>text</th>\n",
       "      <th>teacher_n_tokens</th>\n",
       "      <th>teacher_tokens</th>\n",
       "      <th>student_n_tokens</th>\n",
       "      <th>student_tokens</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>I am still waiting on my card?</td>\n",
       "      <td>8</td>\n",
       "      <td>i am still waiting on my card ?</td>\n",
       "      <td>10</td>\n",
       "      <td>[CLS] i am still waiting on my card ? [SEP]</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>What can I do if my card still hasn't arrived ...</td>\n",
       "      <td>16</td>\n",
       "      <td>what can i do if my card still hasn ' t arrive...</td>\n",
       "      <td>18</td>\n",
       "      <td>[CLS] what can i do if my card still hasn ' t ...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>I have been waiting over a week. Is the card s...</td>\n",
       "      <td>14</td>\n",
       "      <td>i have been waiting over a week . is the card ...</td>\n",
       "      <td>16</td>\n",
       "      <td>[CLS] i have been waiting over a week . is the...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>Can I track my card while it is in the process...</td>\n",
       "      <td>14</td>\n",
       "      <td>can i track my card while it is in the process...</td>\n",
       "      <td>16</td>\n",
       "      <td>[CLS] can i track my card while it is in the p...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>How do I know if I will get my card, or if it ...</td>\n",
       "      <td>17</td>\n",
       "      <td>how do i know if i will get my card , or if it...</td>\n",
       "      <td>19</td>\n",
       "      <td>[CLS] how do i know if i will get my card , or...</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "                                                text  teacher_n_tokens  \\\n",
       "0                     I am still waiting on my card?                 8   \n",
       "1  What can I do if my card still hasn't arrived ...                16   \n",
       "2  I have been waiting over a week. Is the card s...                14   \n",
       "3  Can I track my card while it is in the process...                14   \n",
       "4  How do I know if I will get my card, or if it ...                17   \n",
       "\n",
       "                                      teacher_tokens  student_n_tokens  \\\n",
       "0                    i am still waiting on my card ?                10   \n",
       "1  what can i do if my card still hasn ' t arrive...                18   \n",
       "2  i have been waiting over a week . is the card ...                16   \n",
       "3  can i track my card while it is in the process...                16   \n",
       "4  how do i know if i will get my card , or if it...                19   \n",
       "\n",
       "                                      student_tokens  \n",
       "0        [CLS] i am still waiting on my card ? [SEP]  \n",
       "1  [CLS] what can i do if my card still hasn ' t ...  \n",
       "2  [CLS] i have been waiting over a week . is the...  \n",
       "3  [CLS] can i track my card while it is in the p...  \n",
       "4  [CLS] how do i know if i will get my card , or...  "
      ]
     },
     "execution_count": 9,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "sample_texts = [train_raw[i][\"text\"] for i in [0, 1, 2, 3, 4]]\n",
    "rows = []\n",
    "for t in sample_texts:\n",
    "    teacher_toks = teacher_tokenizer.tokenize(t)\n",
    "    student_toks = student_tok_backend.encode(t).tokens\n",
    "    rows.append({\n",
    "        \"text\": t,\n",
    "        \"teacher_n_tokens\": len(teacher_toks),\n",
    "        \"teacher_tokens\": \" \".join(teacher_toks),\n",
    "        \"student_n_tokens\": len(student_toks),\n",
    "        \"student_tokens\": \" \".join(student_toks),\n",
    "    })\n",
    "\n",
    "tokenization_comparison = pd.DataFrame(rows)\n",
    "tokenization_comparison\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "id": "cf59e7c5",
   "metadata": {
    "execution": {
     "iopub.execute_input": "2026-08-12T16:18:00.016762Z",
     "iopub.status.busy": "2026-08-12T16:18:00.016697Z",
     "iopub.status.idle": "2026-08-12T16:18:00.020674Z",
     "shell.execute_reply": "2026-08-12T16:18:00.020310Z"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Example padded student input_ids shape: torch.Size([2, 32])\n",
      "tensor([[  2,  37, 212, 292, 638, 158, 101, 117,  28,   3,   0,   0,   0,   0,\n",
      "           0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,\n",
      "           0,   0,   0,   0],\n",
      "        [  2, 145, 121,  37, 124, 273, 101, 117, 292, 415,   9,  48, 756, 533,\n",
      "          18, 829,  28,   3,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,\n",
      "           0,   0,   0,   0]])\n"
     ]
    }
   ],
   "source": [
    "def student_encode_batch(texts, max_len=STUDENT_MAX_LEN, pad_id=STUDENT_PAD_ID):\n",
    "    encs = student_tok_backend.encode_batch(list(texts))\n",
    "    input_ids, attn = [], []\n",
    "    for e in encs:\n",
    "        ids = e.ids[:max_len]\n",
    "        pad_len = max_len - len(ids)\n",
    "        attention = [1] * len(ids) + [0] * pad_len\n",
    "        ids = ids + [pad_id] * pad_len\n",
    "        input_ids.append(ids)\n",
    "        attn.append(attention)\n",
    "    return torch.tensor(input_ids, dtype=torch.long), torch.tensor(attn, dtype=torch.long)\n",
    "\n",
    "_demo_ids, _demo_attn = student_encode_batch(sample_texts[:2])\n",
    "print(\"Example padded student input_ids shape:\", _demo_ids.shape)\n",
    "print(_demo_ids)\n"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "7bb271bf",
   "metadata": {},
   "source": [
    "#### Handling the vocabulary mismatch between Teacher and Student\n",
    "\n",
    "A naive version of logit distillation — as used in **token-level** distillation for tasks like\n",
    "sequence-to-sequence generation or token classification — requires the Teacher's and Student's output\n",
    "sequences to line up position-by-position, which breaks immediately if the two models tokenize the same\n",
    "text into different numbers of tokens (which they always will here: the Teacher's ~30k-token vocabulary\n",
    "segments text far more coarsely than the Student's ~3k-token vocabulary, as the comparison table above\n",
    "shows — the Student consistently needs *more* subword pieces for the same utterance).\n",
    "\n",
    "**Why this is not actually a hard problem for us:** distillation here is **sequence classification**, not\n",
    "sequence generation. The Teacher does not produce a per-token output that would need to line up with the\n",
    "Student's per-token output — it produces exactly **one 77-dimensional probability vector per example**,\n",
    "regardless of how many tokens that example was split into internally. So the only alignment that matters is\n",
    "at the **example (row) level**, not the token level:\n",
    "\n",
    "- **Strategy used:** compute Teacher logits once per raw-text example (Task 1), indexed by the example's\n",
    "  position in the (unshuffled) training set. Independently tokenize the *same* raw text with the Student's\n",
    "  own tokenizer for the Student's forward pass. Join the two by index — `teacher_train_logits[i]` always\n",
    "  corresponds to `train_raw[i]`, no matter how differently `train_raw[i][\"text\"]` was tokenized by each\n",
    "  side. This is implemented directly in `BankingStudentDataset` below.\n",
    "- **Residual risk, and why it's acceptable here:** a much smaller vocabulary means more aggressive subword\n",
    "  splitting and a higher `[UNK]` rate, which *can* lose lexical signal the Teacher had access to via its\n",
    "  richer vocabulary. We mitigate this by training the Student tokenizer directly on in-domain banking77\n",
    "  text (rather than a generic corpus), so the ~3k tokens it does have are the ones that matter most for this\n",
    "  task's vocabulary (e.g. \"card\", \"transfer\", \"pin\", \"exchange\" are highly likely to survive intact as\n",
    "  whole-word tokens instead of being fragmented).\n"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "4fcceb36",
   "metadata": {},
   "source": [
    "<br>\n",
    "<hr>\n",
    "<h1>Module 2: Distillation Architecture & Training</h1>\n",
    "<hr>\n",
    "<br>"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "65c1c6f9",
   "metadata": {},
   "source": [
    "### Task 3 — Compact Student Transformer Construction\n",
    "\n",
    "A small, hand-built encoder-only Transformer: learned token + positional embeddings, 4 Transformer encoder\n",
    "layers (hidden size 256, 4 attention heads, feed-forward size 512), mean-pooling over non-padding tokens,\n",
    "and a linear classification head to 77 classes. Built directly from `nn.Module` / `nn.TransformerEncoderLayer`\n",
    "primitives rather than repurposing a pretrained architecture, and sized to run comfortably on CPU."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "id": "7b1e4e6d",
   "metadata": {
    "execution": {
     "iopub.execute_input": "2026-08-12T16:18:00.021914Z",
     "iopub.status.busy": "2026-08-12T16:18:00.021831Z",
     "iopub.status.idle": "2026-08-12T16:18:00.036918Z",
     "shell.execute_reply": "2026-08-12T16:18:00.036579Z"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Student parameter breakdown\n",
      "----------------------------------------\n",
      "  Token + position embeddings : 776,192\n",
      "  Transformer encoder (4 layers): 2,108,416\n",
      "  Classification head          : 19,789\n",
      "  TOTAL                         : 2,904,397\n",
      "\n",
      "Teacher (bert-base-uncased) TOTAL: 109,541,453\n",
      "Compression ratio: 37.7x fewer parameters\n"
     ]
    }
   ],
   "source": [
    "class CompactStudentTransformer(nn.Module):\n",
    "    def __init__(self, vocab_size, num_labels, hidden_size=256, num_layers=4, num_heads=4,\n",
    "                 ffn_size=512, max_len=32, dropout=0.1, pad_id=0):\n",
    "        super().__init__()\n",
    "        self.pad_id = pad_id\n",
    "        self.token_embedding = nn.Embedding(vocab_size, hidden_size, padding_idx=pad_id)\n",
    "        self.position_embedding = nn.Embedding(max_len, hidden_size)\n",
    "        encoder_layer = nn.TransformerEncoderLayer(\n",
    "            d_model=hidden_size,\n",
    "            nhead=num_heads,\n",
    "            dim_feedforward=ffn_size,\n",
    "            dropout=dropout,\n",
    "            activation=\"gelu\",\n",
    "            batch_first=True,\n",
    "        )\n",
    "        self.encoder = nn.TransformerEncoder(encoder_layer, num_layers=num_layers)\n",
    "        self.dropout = nn.Dropout(dropout)\n",
    "        self.classifier = nn.Linear(hidden_size, num_labels)\n",
    "\n",
    "    def forward(self, input_ids, attention_mask):\n",
    "        seq_len = input_ids.size(1)\n",
    "        positions = torch.arange(seq_len, device=input_ids.device).unsqueeze(0)\n",
    "        x = self.token_embedding(input_ids) + self.position_embedding(positions)\n",
    "        pad_mask = attention_mask == 0  # True where padded -> ignored by attention\n",
    "        x = self.encoder(x, src_key_padding_mask=pad_mask)\n",
    "        mask = attention_mask.unsqueeze(-1).float()\n",
    "        pooled = (x * mask).sum(dim=1) / mask.sum(dim=1).clamp(min=1e-6)\n",
    "        return self.classifier(self.dropout(pooled))\n",
    "\n",
    "\n",
    "# inspect.getsource() cannot recover source for classes defined inside a notebook cell executed by\n",
    "# nbconvert (no backing file for linecache to read), so we keep an explicit copy of this class's source\n",
    "# alongside it for the subprocess-isolated benchmarking worker script in Task 6.\n",
    "STUDENT_CLASS_SOURCE = '''class CompactStudentTransformer(nn.Module):\n",
    "    def __init__(self, vocab_size, num_labels, hidden_size=256, num_layers=4, num_heads=4,\n",
    "                 ffn_size=512, max_len=32, dropout=0.1, pad_id=0):\n",
    "        super().__init__()\n",
    "        self.pad_id = pad_id\n",
    "        self.token_embedding = nn.Embedding(vocab_size, hidden_size, padding_idx=pad_id)\n",
    "        self.position_embedding = nn.Embedding(max_len, hidden_size)\n",
    "        encoder_layer = nn.TransformerEncoderLayer(\n",
    "            d_model=hidden_size,\n",
    "            nhead=num_heads,\n",
    "            dim_feedforward=ffn_size,\n",
    "            dropout=dropout,\n",
    "            activation=\"gelu\",\n",
    "            batch_first=True,\n",
    "        )\n",
    "        self.encoder = nn.TransformerEncoder(encoder_layer, num_layers=num_layers)\n",
    "        self.dropout = nn.Dropout(dropout)\n",
    "        self.classifier = nn.Linear(hidden_size, num_labels)\n",
    "\n",
    "    def forward(self, input_ids, attention_mask):\n",
    "        seq_len = input_ids.size(1)\n",
    "        positions = torch.arange(seq_len, device=input_ids.device).unsqueeze(0)\n",
    "        x = self.token_embedding(input_ids) + self.position_embedding(positions)\n",
    "        pad_mask = attention_mask == 0  # True where padded -> ignored by attention\n",
    "        x = self.encoder(x, src_key_padding_mask=pad_mask)\n",
    "        mask = attention_mask.unsqueeze(-1).float()\n",
    "        pooled = (x * mask).sum(dim=1) / mask.sum(dim=1).clamp(min=1e-6)\n",
    "        return self.classifier(self.dropout(pooled))\n",
    "'''\n",
    "\n",
    "\n",
    "STUDENT_CONFIG = dict(\n",
    "    vocab_size=STUDENT_VOCAB_SIZE,\n",
    "    num_labels=num_labels,\n",
    "    hidden_size=256,\n",
    "    num_layers=4,\n",
    "    num_heads=4,\n",
    "    ffn_size=512,\n",
    "    max_len=STUDENT_MAX_LEN,\n",
    "    dropout=0.1,\n",
    "    pad_id=STUDENT_PAD_ID,\n",
    ")\n",
    "\n",
    "def build_student():\n",
    "    torch.manual_seed(SEED)\n",
    "    return CompactStudentTransformer(**STUDENT_CONFIG)\n",
    "\n",
    "_probe = build_student()\n",
    "\n",
    "def count_params(module):\n",
    "    return sum(p.numel() for p in module.parameters())\n",
    "\n",
    "total_student_params = count_params(_probe)\n",
    "embedding_params = count_params(_probe.token_embedding) + count_params(_probe.position_embedding)\n",
    "encoder_params = count_params(_probe.encoder)\n",
    "classifier_params = count_params(_probe.classifier)\n",
    "teacher_params = count_params(teacher_model)\n",
    "\n",
    "print(\"Student parameter breakdown\")\n",
    "print(\"-\" * 40)\n",
    "print(f\"  Token + position embeddings : {embedding_params:,}\")\n",
    "print(f\"  Transformer encoder (4 layers): {encoder_params:,}\")\n",
    "print(f\"  Classification head          : {classifier_params:,}\")\n",
    "print(f\"  TOTAL                         : {total_student_params:,}\")\n",
    "print()\n",
    "print(f\"Teacher (bert-base-uncased) TOTAL: {teacher_params:,}\")\n",
    "print(f\"Compression ratio: {teacher_params / total_student_params:,.1f}x fewer parameters\")\n",
    "del _probe\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "id": "90ab9db2",
   "metadata": {
    "execution": {
     "iopub.execute_input": "2026-08-12T16:18:00.037989Z",
     "iopub.status.busy": "2026-08-12T16:18:00.037917Z",
     "iopub.status.idle": "2026-08-12T16:18:00.040227Z",
     "shell.execute_reply": "2026-08-12T16:18:00.039854Z"
    }
   },
   "outputs": [
    {
     "data": {
      "text/markdown": [
       "\n",
       "#### Interpretation\n",
       "\n",
       "The compact Student has **2,904,397 parameters** against the Teacher's\n",
       "**109,541,453** — a **37.7x** reduction. Roughly\n",
       "**27%** of the Student's budget sits in its embedding table alone,\n",
       "which is the direct payoff of Task 2's small, domain-specific vocabulary (3000 tokens vs.\n",
       "BERT's 30522): most of a Transformer's parameter count for short-sequence\n",
       "classification tasks scales with `vocab_size x hidden_size`, so shrinking the vocabulary is one of the single\n",
       "highest-leverage compression decisions available, independent of how many encoder layers are kept.\n"
      ],
      "text/plain": [
       "<IPython.core.display.Markdown object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "display(Markdown(f'''\n",
    "#### Interpretation\n",
    "\n",
    "The compact Student has **{total_student_params:,} parameters** against the Teacher's\n",
    "**{teacher_params:,}** — a **{teacher_params/total_student_params:,.1f}x** reduction. Roughly\n",
    "**{embedding_params/total_student_params:.0%}** of the Student's budget sits in its embedding table alone,\n",
    "which is the direct payoff of Task 2's small, domain-specific vocabulary ({STUDENT_VOCAB_SIZE} tokens vs.\n",
    "BERT's {teacher_tokenizer.vocab_size}): most of a Transformer's parameter count for short-sequence\n",
    "classification tasks scales with `vocab_size x hidden_size`, so shrinking the vocabulary is one of the single\n",
    "highest-leverage compression decisions available, independent of how many encoder layers are kept.\n",
    "'''))\n"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "21c36067",
   "metadata": {},
   "source": [
    "### Task 4 — Distillation Loss Function\n",
    "\n",
    "$$\\text{Loss} = \\alpha \\cdot T^2 \\cdot \\text{KL}\\big(P_{\\text{student}}^T \\,\\|\\, P_{\\text{teacher}}^T\\big) + (1-\\alpha)\\cdot \\text{CE}(y_{\\text{student}}, y_{\\text{true}})$$\n",
    "\n",
    "with $T = 4.0$ and $\\alpha = 0.7$. The $T^2$ scaling (Hinton et al., 2015) compensates for the fact that\n",
    "raising the temperature shrinks the magnitude of the gradients coming from the soft-label term by roughly\n",
    "$1/T^2$ relative to the hard-label term, so without it the KD loss would be under-weighted relative to\n",
    "$\\alpha$ once a large $T$ is introduced."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "id": "1c6ab457",
   "metadata": {
    "execution": {
     "iopub.execute_input": "2026-08-12T16:18:00.041260Z",
     "iopub.status.busy": "2026-08-12T16:18:00.041193Z",
     "iopub.status.idle": "2026-08-12T16:18:00.400864Z",
     "shell.execute_reply": "2026-08-12T16:18:00.400547Z"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Distillation train set: 10003 examples (with cached teacher logits)\n",
      "Baseline train set    : 10003 examples (hard labels only)\n"
     ]
    }
   ],
   "source": [
    "class DistillationLoss(nn.Module):\n",
    "    def __init__(self, temperature=4.0, alpha=0.7):\n",
    "        super().__init__()\n",
    "        self.T = temperature\n",
    "        self.alpha = alpha\n",
    "        self.kl = nn.KLDivLoss(reduction=\"batchmean\")\n",
    "        self.ce = nn.CrossEntropyLoss()\n",
    "\n",
    "    def forward(self, student_logits, teacher_logits, true_labels):\n",
    "        student_log_probs_T = F.log_softmax(student_logits / self.T, dim=-1)\n",
    "        teacher_probs_T = F.softmax(teacher_logits / self.T, dim=-1)\n",
    "        kd_loss = self.kl(student_log_probs_T, teacher_probs_T) * (self.T ** 2)\n",
    "        ce_loss = self.ce(student_logits, true_labels)\n",
    "        total = self.alpha * kd_loss + (1 - self.alpha) * ce_loss\n",
    "        return total, kd_loss.detach(), ce_loss.detach()\n",
    "\n",
    "\n",
    "class BankingStudentDataset(Dataset):\n",
    "    # Joins raw text (re-tokenized with the STUDENT tokenizer) to cached TEACHER logits by row index.\n",
    "    def __init__(self, texts, labels, teacher_logits=None):\n",
    "        self.input_ids, self.attention_mask = student_encode_batch(texts)\n",
    "        self.labels = torch.tensor(labels, dtype=torch.long)\n",
    "        self.teacher_logits = teacher_logits\n",
    "\n",
    "    def __len__(self):\n",
    "        return len(self.labels)\n",
    "\n",
    "    def __getitem__(self, idx):\n",
    "        item = {\n",
    "            \"input_ids\": self.input_ids[idx],\n",
    "            \"attention_mask\": self.attention_mask[idx],\n",
    "            \"labels\": self.labels[idx],\n",
    "        }\n",
    "        if self.teacher_logits is not None:\n",
    "            item[\"teacher_logits\"] = self.teacher_logits[idx]\n",
    "        return item\n",
    "\n",
    "\n",
    "train_texts, train_labels = train_raw[\"text\"], train_raw[\"label\"]\n",
    "test_texts, test_labels = test_raw[\"text\"], test_raw[\"label\"]\n",
    "\n",
    "distill_train_dataset = BankingStudentDataset(train_texts, train_labels, teacher_logits=teacher_train_logits)\n",
    "baseline_train_dataset = BankingStudentDataset(train_texts, train_labels, teacher_logits=None)\n",
    "student_test_dataset = BankingStudentDataset(test_texts, test_labels, teacher_logits=None)\n",
    "\n",
    "print(f\"Distillation train set: {len(distill_train_dataset)} examples (with cached teacher logits)\")\n",
    "print(f\"Baseline train set    : {len(baseline_train_dataset)} examples (hard labels only)\")\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "id": "e57ed00e",
   "metadata": {
    "execution": {
     "iopub.execute_input": "2026-08-12T16:18:00.402097Z",
     "iopub.status.busy": "2026-08-12T16:18:00.402022Z",
     "iopub.status.idle": "2026-08-12T16:18:00.405453Z",
     "shell.execute_reply": "2026-08-12T16:18:00.405075Z"
    }
   },
   "outputs": [],
   "source": [
    "def train_student(model, dataset, device, epochs=8, batch_size=32, lr=3e-4,\n",
    "                   distill=False, temperature=4.0, alpha=0.7, log_prefix=\"student\"):\n",
    "    model.to(device)\n",
    "    loader = DataLoader(dataset, batch_size=batch_size, shuffle=True)\n",
    "    optimizer = torch.optim.AdamW(model.parameters(), lr=lr)\n",
    "    ce_loss_fn = nn.CrossEntropyLoss()\n",
    "    distill_loss_fn = DistillationLoss(temperature=temperature, alpha=alpha) if distill else None\n",
    "\n",
    "    history = {\"total\": [], \"kd\": [], \"ce\": []}\n",
    "    model.train()\n",
    "    for epoch in range(epochs):\n",
    "        totals, kds, ces = [], [], []\n",
    "        for batch in loader:\n",
    "            input_ids = batch[\"input_ids\"].to(device)\n",
    "            attention_mask = batch[\"attention_mask\"].to(device)\n",
    "            labels = batch[\"labels\"].to(device)\n",
    "\n",
    "            optimizer.zero_grad()\n",
    "            logits = model(input_ids, attention_mask)\n",
    "\n",
    "            if distill:\n",
    "                teacher_logits = batch[\"teacher_logits\"].to(device)\n",
    "                loss, kd, ce = distill_loss_fn(logits, teacher_logits, labels)\n",
    "                kds.append(kd.item())\n",
    "                ces.append(ce.item())\n",
    "            else:\n",
    "                loss = ce_loss_fn(logits, labels)\n",
    "\n",
    "            loss.backward()\n",
    "            optimizer.step()\n",
    "            totals.append(loss.item())\n",
    "\n",
    "        avg_total = float(np.mean(totals))\n",
    "        history[\"total\"].append(avg_total)\n",
    "        if distill:\n",
    "            history[\"kd\"].append(float(np.mean(kds)))\n",
    "            history[\"ce\"].append(float(np.mean(ces)))\n",
    "            print(f\"[{log_prefix}] epoch {epoch+1}/{epochs} - loss {avg_total:.4f} \"\n",
    "                  f\"(kd {history['kd'][-1]:.4f}, ce {history['ce'][-1]:.4f})\")\n",
    "        else:\n",
    "            print(f\"[{log_prefix}] epoch {epoch+1}/{epochs} - loss {avg_total:.4f}\")\n",
    "    return history\n",
    "\n",
    "STUDENT_EPOCHS = 8\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "id": "88932833",
   "metadata": {
    "execution": {
     "iopub.execute_input": "2026-08-12T16:18:00.406485Z",
     "iopub.status.busy": "2026-08-12T16:18:00.406414Z",
     "iopub.status.idle": "2026-08-12T16:20:11.219388Z",
     "shell.execute_reply": "2026-08-12T16:20:11.218892Z"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[distilled] epoch 1/8 - loss 1.1419 (kd 0.5564, ce 2.5081)\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[distilled] epoch 2/8 - loss 0.5591 (kd 0.3059, ce 1.1498)\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[distilled] epoch 3/8 - loss 0.3874 (kd 0.2167, ce 0.7857)\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[distilled] epoch 4/8 - loss 0.3012 (kd 0.1740, ce 0.5979)\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[distilled] epoch 5/8 - loss 0.2480 (kd 0.1476, ce 0.4821)\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[distilled] epoch 6/8 - loss 0.2110 (kd 0.1316, ce 0.3963)\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[distilled] epoch 7/8 - loss 0.1841 (kd 0.1184, ce 0.3374)\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[distilled] epoch 8/8 - loss 0.1658 (kd 0.1098, ce 0.2965)\n",
      "\n",
      "Distilled student training took 130.8s on cpu\n"
     ]
    }
   ],
   "source": [
    "torch.manual_seed(SEED)\n",
    "distilled_student = build_student()\n",
    "\n",
    "distill_start = time.time()\n",
    "distill_history = train_student(\n",
    "    distilled_student, distill_train_dataset, STUDENT_DEVICE,\n",
    "    epochs=STUDENT_EPOCHS, batch_size=32, lr=3e-4,\n",
    "    distill=True, temperature=4.0, alpha=0.7, log_prefix=\"distilled\",\n",
    ")\n",
    "distill_seconds = time.time() - distill_start\n",
    "print(f\"\\nDistilled student training took {distill_seconds:.1f}s on {STUDENT_DEVICE}\")\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "id": "2add1584",
   "metadata": {
    "execution": {
     "iopub.execute_input": "2026-08-12T16:20:11.220535Z",
     "iopub.status.busy": "2026-08-12T16:20:11.220454Z",
     "iopub.status.idle": "2026-08-12T16:20:11.332390Z",
     "shell.execute_reply": "2026-08-12T16:20:11.331982Z"
    }
   },
   "outputs": [
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<Figure size 1100x400 with 2 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "fig, axes = plt.subplots(1, 2, figsize=(11, 4))\n",
    "axes[0].plot(distill_history[\"total\"], marker=\"o\", label=\"total loss\")\n",
    "axes[0].set_title(\"Distilled student: total loss\")\n",
    "axes[0].set_xlabel(\"epoch\"); axes[0].set_ylabel(\"loss\"); axes[0].legend()\n",
    "\n",
    "axes[1].plot(distill_history[\"kd\"], marker=\"o\", label=\"KD term (KL, T-scaled)\")\n",
    "axes[1].plot(distill_history[\"ce\"], marker=\"s\", label=\"CE term (hard labels)\")\n",
    "axes[1].set_title(\"Distilled student: loss components\")\n",
    "axes[1].set_xlabel(\"epoch\"); axes[1].set_ylabel(\"loss\"); axes[1].legend()\n",
    "plt.tight_layout()\n",
    "plt.show()\n"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "80a983e5",
   "metadata": {},
   "source": [
    "**Interpretation:** the CE component (against ground-truth hard labels) typically drops faster and\n",
    "further than the KD component, because a 3000-token-vocabulary, 4-layer Student can quickly memorize the\n",
    "*correct class* for a small, well-separated training set, while matching the Teacher's *full smoothed\n",
    "distribution* over 77 classes at $T{=}4$ is a strictly harder target — the KD term keeps providing a\n",
    "non-trivial gradient signal well after the CE term has largely converged, which is exactly the regime where\n",
    "distillation is doing useful work beyond what hard labels alone would teach."
   ]
  },
  {
   "cell_type": "markdown",
   "id": "d12346db",
   "metadata": {},
   "source": [
    "<br>\n",
    "<hr>\n",
    "<h1>Module 3: Comparative Analysis & Benchmarking</h1>\n",
    "<hr>\n",
    "<br>"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "c03183d5",
   "metadata": {},
   "source": [
    "### Task 5 — Accuracy vs. Compression Evaluation\n",
    "\n",
    "We now train an **identical-architecture baseline Student** from scratch using plain cross-entropy on\n",
    "ground-truth labels only (no Teacher signal at all), so any accuracy difference between it and the\n",
    "distilled Student isolates the effect of distillation itself, holding architecture, tokenizer, optimizer,\n",
    "and epoch budget fixed."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "id": "94314539",
   "metadata": {
    "execution": {
     "iopub.execute_input": "2026-08-12T16:20:11.333646Z",
     "iopub.status.busy": "2026-08-12T16:20:11.333568Z",
     "iopub.status.idle": "2026-08-12T16:22:18.942562Z",
     "shell.execute_reply": "2026-08-12T16:22:18.942204Z"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[baseline] epoch 1/8 - loss 2.4675\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[baseline] epoch 2/8 - loss 0.9531\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[baseline] epoch 3/8 - loss 0.5728\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[baseline] epoch 4/8 - loss 0.3868\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[baseline] epoch 5/8 - loss 0.2510\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[baseline] epoch 6/8 - loss 0.1739\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[baseline] epoch 7/8 - loss 0.1214\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[baseline] epoch 8/8 - loss 0.1027\n",
      "\n",
      "Baseline student training took 127.6s on cpu\n"
     ]
    }
   ],
   "source": [
    "torch.manual_seed(SEED)\n",
    "baseline_student = build_student()\n",
    "\n",
    "baseline_start = time.time()\n",
    "baseline_history = train_student(\n",
    "    baseline_student, baseline_train_dataset, STUDENT_DEVICE,\n",
    "    epochs=STUDENT_EPOCHS, batch_size=32, lr=3e-4,\n",
    "    distill=False, log_prefix=\"baseline\",\n",
    ")\n",
    "baseline_seconds = time.time() - baseline_start\n",
    "print(f\"\\nBaseline student training took {baseline_seconds:.1f}s on {STUDENT_DEVICE}\")\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "id": "9a353539",
   "metadata": {
    "execution": {
     "iopub.execute_input": "2026-08-12T16:22:18.943947Z",
     "iopub.status.busy": "2026-08-12T16:22:18.943837Z",
     "iopub.status.idle": "2026-08-12T16:22:18.991016Z",
     "shell.execute_reply": "2026-08-12T16:22:18.990689Z"
    }
   },
   "outputs": [
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<Figure size 600x400 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plt.figure(figsize=(6, 4))\n",
    "plt.plot(baseline_history[\"total\"], marker=\"o\", label=\"Baseline student (CE only)\")\n",
    "plt.plot(distill_history[\"total\"], marker=\"s\", label=\"Distilled student (KD + CE)\")\n",
    "plt.title(\"Training loss: baseline vs. distilled student\")\n",
    "plt.xlabel(\"epoch\"); plt.ylabel(\"loss (not directly comparable in scale)\"); plt.legend()\n",
    "plt.tight_layout()\n",
    "plt.show()\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "id": "3d7bbfc0",
   "metadata": {
    "execution": {
     "iopub.execute_input": "2026-08-12T16:22:18.992264Z",
     "iopub.status.busy": "2026-08-12T16:22:18.992183Z",
     "iopub.status.idle": "2026-08-12T16:22:27.493008Z",
     "shell.execute_reply": "2026-08-12T16:22:27.492595Z"
    }
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/opt/homebrew/Caskroom/miniforge/base/envs/agn_env/lib/python3.12/site-packages/torch/utils/data/dataloader.py:759: UserWarning: 'pin_memory' argument is set as true but not supported on MPS now, device pinned memory won't be used.\n",
      "  super().__init__(loader)\n"
     ]
    },
    {
     "data": {
      "text/html": [],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/opt/homebrew/Caskroom/miniforge/base/envs/agn_env/lib/python3.12/site-packages/torch/nn/modules/transformer.py:529: UserWarning: The PyTorch API of nested tensors is in prototype stage and will change in the near future. We recommend specifying layout=torch.jagged when constructing a nested tensor, as this layout receives active development, has better operator coverage, and works with torch.compile. (Triggered internally at /Users/runner/work/pytorch/pytorch/aten/src/ATen/NestedTensorImpl.cpp:181.)\n",
      "  output = torch._nested_tensor_from_mask(\n"
     ]
    },
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>Model</th>\n",
       "      <th>Accuracy</th>\n",
       "      <th>Macro F1</th>\n",
       "      <th>Weighted F1</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>Teacher (bert-base-uncased)</td>\n",
       "      <td>0.8756</td>\n",
       "      <td>0.8677</td>\n",
       "      <td>0.8677</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>Student — WITHOUT distillation</td>\n",
       "      <td>0.8347</td>\n",
       "      <td>0.8355</td>\n",
       "      <td>0.8355</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>Student — WITH distillation</td>\n",
       "      <td>0.8744</td>\n",
       "      <td>0.8743</td>\n",
       "      <td>0.8743</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "                            Model  Accuracy  Macro F1  Weighted F1\n",
       "0     Teacher (bert-base-uncased)    0.8756    0.8677       0.8677\n",
       "1  Student — WITHOUT distillation    0.8347    0.8355       0.8355\n",
       "2     Student — WITH distillation    0.8744    0.8743       0.8743"
      ]
     },
     "execution_count": 19,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "@torch.no_grad()\n",
    "def evaluate_student(model, dataset, device, batch_size=64):\n",
    "    model.eval()\n",
    "    model.to(device)\n",
    "    loader = DataLoader(dataset, batch_size=batch_size, shuffle=False)\n",
    "    all_preds, all_labels = [], []\n",
    "    for batch in loader:\n",
    "        input_ids = batch[\"input_ids\"].to(device)\n",
    "        attention_mask = batch[\"attention_mask\"].to(device)\n",
    "        logits = model(input_ids, attention_mask)\n",
    "        all_preds.extend(torch.argmax(logits, dim=-1).cpu().numpy().tolist())\n",
    "        all_labels.extend(batch[\"labels\"].numpy().tolist())\n",
    "    return np.array(all_preds), np.array(all_labels)\n",
    "\n",
    "\n",
    "teacher_pred_output = trainer.predict(test_enc)\n",
    "teacher_test_preds = np.argmax(teacher_pred_output.predictions, axis=-1)\n",
    "teacher_test_labels = np.array(test_raw[\"label\"])\n",
    "\n",
    "baseline_preds, baseline_labels = evaluate_student(baseline_student, student_test_dataset, STUDENT_DEVICE)\n",
    "distilled_preds, distilled_labels = evaluate_student(distilled_student, student_test_dataset, STUDENT_DEVICE)\n",
    "\n",
    "def summarize(name, preds, labels):\n",
    "    return {\n",
    "        \"Model\": name,\n",
    "        \"Accuracy\": accuracy_score(labels, preds),\n",
    "        \"Macro F1\": f1_score(labels, preds, average=\"macro\"),\n",
    "        \"Weighted F1\": f1_score(labels, preds, average=\"weighted\"),\n",
    "    }\n",
    "\n",
    "results_df = pd.DataFrame([\n",
    "    summarize(\"Teacher (bert-base-uncased)\", teacher_test_preds, teacher_test_labels),\n",
    "    summarize(\"Student — WITHOUT distillation\", baseline_preds, baseline_labels),\n",
    "    summarize(\"Student — WITH distillation\", distilled_preds, distilled_labels),\n",
    "])\n",
    "results_df[[\"Accuracy\", \"Macro F1\", \"Weighted F1\"]] = results_df[[\"Accuracy\", \"Macro F1\", \"Weighted F1\"]].round(4)\n",
    "results_df\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "id": "60d0cb4c",
   "metadata": {
    "execution": {
     "iopub.execute_input": "2026-08-12T16:22:27.494377Z",
     "iopub.status.busy": "2026-08-12T16:22:27.494274Z",
     "iopub.status.idle": "2026-08-12T16:22:27.498192Z",
     "shell.execute_reply": "2026-08-12T16:22:27.497818Z"
    }
   },
   "outputs": [
    {
     "data": {
      "text/markdown": [
       "\n",
       "#### Comparison table\n",
       "\n",
       "| Model | Accuracy | Macro F1 | Weighted F1 |\n",
       "|---|---|---|---|\n",
       "| Teacher (bert-base-uncased) | 0.8756 | 0.8677 | 0.8677 |\n",
       "| Student — WITHOUT distillation | 0.8347 | 0.8355 | 0.8355 |\n",
       "| Student — WITH distillation | 0.8744 | 0.8743 | 0.8743 |\n",
       "\n",
       "#### Interpretation\n",
       "\n",
       "- The undistilled Student, trained only on hard labels with a **38x**\n",
       "  smaller architecture and a **8**-epoch budget over the same 10003 examples,\n",
       "  reaches **83.5%** accuracy — a substantial gap below the Teacher's **87.6%**,\n",
       "  as expected given how much capacity was removed.\n",
       "- Adding the Teacher's soft labels (identical architecture, identical data, identical epoch budget — only the loss function differs) raises the Student to **87.4%** accuracy, closing **97%** of the accuracy gap between the undistilled Student and the Teacher.\n",
       "- The distilled Student retains **99.9%** of the Teacher's accuracy at a fraction of its parameter count — this is the central empirical claim of knowledge distillation: dark knowledge in the soft labels lets a small model recover much more of a large model's decision surface than the same small model could learn from hard labels alone.\n"
      ],
      "text/plain": [
       "<IPython.core.display.Markdown object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "def to_markdown_table(df):\n",
    "    header = \"| \" + \" | \".join(df.columns) + \" |\"\n",
    "    sep = \"|\" + \"|\".join([\"---\"] * len(df.columns)) + \"|\"\n",
    "    rows = [\"| \" + \" | \".join(str(v) for v in row) + \" |\" for row in df.values]\n",
    "    return \"\\n\".join([header, sep] + rows)\n",
    "\n",
    "teacher_acc = results_df.loc[0, \"Accuracy\"]\n",
    "baseline_acc = results_df.loc[1, \"Accuracy\"]\n",
    "distilled_acc = results_df.loc[2, \"Accuracy\"]\n",
    "recovery = (distilled_acc - baseline_acc) / max(teacher_acc - baseline_acc, 1e-9) * 100\n",
    "retention = distilled_acc / teacher_acc * 100\n",
    "\n",
    "if distilled_acc >= teacher_acc:\n",
    "    gap_line = (\n",
    "        f\"- Adding the Teacher's soft labels (identical architecture, identical data, identical epoch budget — \"\n",
    "        f\"only the loss function differs) raises the Student to **{distilled_acc:.1%}** accuracy, which \"\n",
    "        f\"**fully closes** the gap to the Teacher's **{teacher_acc:.1%}** and slightly surpasses it.\"\n",
    "    )\n",
    "    retention_line = (\n",
    "        f\"- Remarkably, the distilled Student's **{distilled_acc:.1%}** accuracy matches — and here, \"\n",
    "        f\"slightly exceeds — the Teacher's **{teacher_acc:.1%}**, at **{teacher_params/total_student_params:,.0f}x** \"\n",
    "        f\"fewer parameters. This is a stronger-than-typical (though not unheard of) outcome: the Teacher was \"\n",
    "        f\"only fine-tuned for a few epochs, so its own decision boundary still carries some noise, and the \"\n",
    "        f\"test set is a few thousand examples, so a couple of points either way is within normal variance. \"\n",
    "        f\"The soft-label targets act as a strong regularizer that helps the tiny Student generalize at least \"\n",
    "        f\"as well on this held-out set — that does **not** mean the Student has absorbed *all* of the \"\n",
    "        f\"Teacher's knowledge, only that on this test split and this metric, distillation fully closed the gap.\"\n",
    "    )\n",
    "else:\n",
    "    gap_line = (\n",
    "        f\"- Adding the Teacher's soft labels (identical architecture, identical data, identical epoch budget — \"\n",
    "        f\"only the loss function differs) raises the Student to **{distilled_acc:.1%}** accuracy, closing \"\n",
    "        f\"**{recovery:.0f}%** of the accuracy gap between the undistilled Student and the Teacher.\"\n",
    "    )\n",
    "    retention_line = (\n",
    "        f\"- The distilled Student retains **{retention:.1f}%** of the Teacher's accuracy at a fraction of its \"\n",
    "        f\"parameter count — this is the central empirical claim of knowledge distillation: dark knowledge in \"\n",
    "        f\"the soft labels lets a small model recover much more of a large model's decision surface than the \"\n",
    "        f\"same small model could learn from hard labels alone.\"\n",
    "    )\n",
    "\n",
    "display(Markdown(f'''\n",
    "#### Comparison table\n",
    "\n",
    "{to_markdown_table(results_df)}\n",
    "\n",
    "#### Interpretation\n",
    "\n",
    "- The undistilled Student, trained only on hard labels with a **{teacher_params/total_student_params:,.0f}x**\n",
    "  smaller architecture and a **{STUDENT_EPOCHS}**-epoch budget over the same {len(train_raw)} examples,\n",
    "  reaches **{baseline_acc:.1%}** accuracy — a substantial gap below the Teacher's **{teacher_acc:.1%}**,\n",
    "  as expected given how much capacity was removed.\n",
    "{gap_line}\n",
    "{retention_line}\n",
    "'''))\n"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "0b92c230",
   "metadata": {},
   "source": [
    "### Task 6 — Deployment Metrics Analysis\n",
    "\n",
    "We measure three deployment-relevant metrics for the Teacher vs. the (distilled) Student:\n",
    "\n",
    "1. **Disk size** — serialized `state_dict` size on disk.\n",
    "2. **CPU inference latency** — mean wall-clock time per single-example (`batch_size=1`) forward pass on CPU.\n",
    "3. **Peak process RAM** — measured in an **isolated subprocess per model** (via `resource.getrusage`), so\n",
    "   the Teacher's ~440MB footprint doesn't contaminate the Student's measurement just because both happen to\n",
    "   be loaded in the same notebook kernel."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "id": "9bf9f05e",
   "metadata": {
    "execution": {
     "iopub.execute_input": "2026-08-12T16:22:27.499319Z",
     "iopub.status.busy": "2026-08-12T16:22:27.499235Z",
     "iopub.status.idle": "2026-08-12T16:22:28.131984Z",
     "shell.execute_reply": "2026-08-12T16:22:28.131585Z"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Teacher disk size : 417.9 MB\n",
      "Student disk size : 11.1 MB\n",
      "Size reduction    : 37.7x smaller\n"
     ]
    }
   ],
   "source": [
    "def get_model_disk_size_mb(state_dict):\n",
    "    with tempfile.NamedTemporaryFile(suffix=\".pt\") as f:\n",
    "        torch.save(state_dict, f.name)\n",
    "        f.flush()\n",
    "        size_bytes = os.path.getsize(f.name)\n",
    "    return size_bytes / (1024 ** 2)\n",
    "\n",
    "teacher_size_mb = get_model_disk_size_mb(teacher_model.state_dict())\n",
    "student_size_mb = get_model_disk_size_mb(distilled_student.state_dict())\n",
    "\n",
    "print(f\"Teacher disk size : {teacher_size_mb:,.1f} MB\")\n",
    "print(f\"Student disk size : {student_size_mb:,.1f} MB\")\n",
    "print(f\"Size reduction    : {teacher_size_mb/student_size_mb:,.1f}x smaller\")\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "id": "d85d32cf",
   "metadata": {
    "execution": {
     "iopub.execute_input": "2026-08-12T16:22:28.133141Z",
     "iopub.status.busy": "2026-08-12T16:22:28.133069Z",
     "iopub.status.idle": "2026-08-12T16:22:29.806548Z",
     "shell.execute_reply": "2026-08-12T16:22:29.806070Z"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Teacher CPU latency : 27.76 +/- 3.07 ms/query\n",
      "Student CPU latency : 1.18 +/- 0.09 ms/query\n",
      "Speedup             : 23.5x faster\n"
     ]
    }
   ],
   "source": [
    "@torch.no_grad()\n",
    "def measure_teacher_cpu_latency(model, tokenizer, texts, n_warmup=5, n_runs=50):\n",
    "    model.to(CPU_DEVICE)\n",
    "    model.eval()\n",
    "    encs = [tokenizer(t, return_tensors=\"pt\", truncation=True, max_length=64) for t in texts]\n",
    "    for i in range(n_warmup):\n",
    "        model(**encs[i % len(encs)])\n",
    "    times = []\n",
    "    for i in range(n_runs):\n",
    "        enc = encs[i % len(encs)]\n",
    "        start = time.perf_counter()\n",
    "        model(**enc)\n",
    "        times.append((time.perf_counter() - start) * 1000)\n",
    "    return float(np.mean(times)), float(np.std(times))\n",
    "\n",
    "@torch.no_grad()\n",
    "def measure_student_cpu_latency(model, texts, n_warmup=5, n_runs=50):\n",
    "    model.to(CPU_DEVICE)\n",
    "    model.eval()\n",
    "    ids, attn = student_encode_batch(texts)\n",
    "    for i in range(n_warmup):\n",
    "        idx = i % len(texts)\n",
    "        model(ids[idx].unsqueeze(0), attn[idx].unsqueeze(0))\n",
    "    times = []\n",
    "    for i in range(n_runs):\n",
    "        idx = i % len(texts)\n",
    "        start = time.perf_counter()\n",
    "        model(ids[idx].unsqueeze(0), attn[idx].unsqueeze(0))\n",
    "        times.append((time.perf_counter() - start) * 1000)\n",
    "    return float(np.mean(times)), float(np.std(times))\n",
    "\n",
    "latency_sample_texts = test_texts[:20]\n",
    "teacher_latency_mean, teacher_latency_std = measure_teacher_cpu_latency(teacher_model, teacher_tokenizer, latency_sample_texts)\n",
    "student_latency_mean, student_latency_std = measure_student_cpu_latency(distilled_student, latency_sample_texts)\n",
    "\n",
    "print(f\"Teacher CPU latency : {teacher_latency_mean:.2f} +/- {teacher_latency_std:.2f} ms/query\")\n",
    "print(f\"Student CPU latency : {student_latency_mean:.2f} +/- {student_latency_std:.2f} ms/query\")\n",
    "print(f\"Speedup             : {teacher_latency_mean/student_latency_mean:,.1f}x faster\")\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "id": "fedbc12e",
   "metadata": {
    "execution": {
     "iopub.execute_input": "2026-08-12T16:22:29.807836Z",
     "iopub.status.busy": "2026-08-12T16:22:29.807755Z",
     "iopub.status.idle": "2026-08-12T16:22:38.087063Z",
     "shell.execute_reply": "2026-08-12T16:22:38.086596Z"
    }
   },
   "outputs": [
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "647e0e6626fe4b2d94cd4ded2e446145",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "Writing model shards:   0%|          | 0/1 [00:00<?, ?it/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Teacher peak RAM (isolated process) : 893.8 MB\n",
      "Student peak RAM (isolated process) : 237.0 MB\n",
      "RAM reduction                        : 3.8x smaller\n"
     ]
    }
   ],
   "source": [
    "# Peak RAM, measured per model in an isolated subprocess so each number reflects ONLY that model's\n",
    "# footprint (loading both Teacher and Student into one long-lived kernel would make ru_maxrss monotonically\n",
    "# dominated by whichever model was loaded first).\n",
    "teacher_worker_src = (\n",
    "    \"import sys, resource, torch\\n\"\n",
    "    \"from transformers import BertForSequenceClassification, AutoTokenizer\\n\"\n",
    "    \"model_dir = sys.argv[1]\\n\"\n",
    "    \"model = BertForSequenceClassification.from_pretrained(model_dir)\\n\"\n",
    "    \"tokenizer = AutoTokenizer.from_pretrained(model_dir)\\n\"\n",
    "    \"model.eval()\\n\"\n",
    "    \"texts = ['I am still waiting on my card?'] * 30\\n\"\n",
    "    \"with torch.no_grad():\\n\"\n",
    "    \"    for t in texts:\\n\"\n",
    "    \"        enc = tokenizer(t, return_tensors='pt', truncation=True, max_length=64)\\n\"\n",
    "    \"        model(**enc)\\n\"\n",
    "    \"print(resource.getrusage(resource.RUSAGE_SELF).ru_maxrss)\\n\"\n",
    ")\n",
    "\n",
    "student_class_src = STUDENT_CLASS_SOURCE\n",
    "student_worker_src = (\n",
    "    \"import sys, json, resource, torch\\n\"\n",
    "    \"import torch.nn as nn\\n\"\n",
    "    \"from tokenizers import Tokenizer as HFTokenizer\\n\\n\"\n",
    "    + student_class_src + \"\\n\\n\"\n",
    "    \"model_dir = sys.argv[1]\\n\"\n",
    "    \"cfg = json.load(open(model_dir + '/config.json'))\\n\"\n",
    "    \"model = CompactStudentTransformer(**cfg)\\n\"\n",
    "    \"model.load_state_dict(torch.load(model_dir + '/weights.pt', map_location='cpu'))\\n\"\n",
    "    \"model.eval()\\n\"\n",
    "    \"tok = HFTokenizer.from_file(model_dir + '/tokenizer.json')\\n\"\n",
    "    \"texts = ['I am still waiting on my card?'] * 30\\n\"\n",
    "    \"with torch.no_grad():\\n\"\n",
    "    \"    for t in texts:\\n\"\n",
    "    \"        ids = tok.encode(t).ids[:cfg['max_len']]\\n\"\n",
    "    \"        pad_len = cfg['max_len'] - len(ids)\\n\"\n",
    "    \"        attn = [1]*len(ids) + [0]*pad_len\\n\"\n",
    "    \"        ids = ids + [cfg['pad_id']]*pad_len\\n\"\n",
    "    \"        input_ids = torch.tensor([ids])\\n\"\n",
    "    \"        attention_mask = torch.tensor([attn])\\n\"\n",
    "    \"        model(input_ids, attention_mask)\\n\"\n",
    "    \"print(resource.getrusage(resource.RUSAGE_SELF).ru_maxrss)\\n\"\n",
    ")\n",
    "\n",
    "with tempfile.TemporaryDirectory() as tmp_root:\n",
    "    teacher_dir = os.path.join(tmp_root, \"teacher\")\n",
    "    student_dir = os.path.join(tmp_root, \"student\")\n",
    "    os.makedirs(teacher_dir, exist_ok=True)\n",
    "    os.makedirs(student_dir, exist_ok=True)\n",
    "\n",
    "    teacher_model.save_pretrained(teacher_dir)\n",
    "    teacher_tokenizer.save_pretrained(teacher_dir)\n",
    "\n",
    "    torch.save(distilled_student.state_dict(), os.path.join(student_dir, \"weights.pt\"))\n",
    "    with open(os.path.join(student_dir, \"config.json\"), \"w\") as f:\n",
    "        json.dump(STUDENT_CONFIG, f)\n",
    "    student_tok_backend.save(os.path.join(student_dir, \"tokenizer.json\"))\n",
    "\n",
    "    teacher_worker_path = os.path.join(tmp_root, \"teacher_worker.py\")\n",
    "    student_worker_path = os.path.join(tmp_root, \"student_worker.py\")\n",
    "    with open(teacher_worker_path, \"w\") as f:\n",
    "        f.write(teacher_worker_src)\n",
    "    with open(student_worker_path, \"w\") as f:\n",
    "        f.write(student_worker_src)\n",
    "\n",
    "    teacher_ram_out = subprocess.run(\n",
    "        [sys.executable, teacher_worker_path, teacher_dir],\n",
    "        capture_output=True, text=True, check=True,\n",
    "    )\n",
    "    student_ram_out = subprocess.run(\n",
    "        [sys.executable, student_worker_path, student_dir],\n",
    "        capture_output=True, text=True, check=True,\n",
    "    )\n",
    "\n",
    "_bytes_per_unit = 1  # macOS ru_maxrss is already in bytes\n",
    "teacher_peak_ram_mb = int(teacher_ram_out.stdout.strip()) * _bytes_per_unit / (1024 ** 2)\n",
    "student_peak_ram_mb = int(student_ram_out.stdout.strip()) * _bytes_per_unit / (1024 ** 2)\n",
    "\n",
    "print(f\"Teacher peak RAM (isolated process) : {teacher_peak_ram_mb:,.1f} MB\")\n",
    "print(f\"Student peak RAM (isolated process) : {student_peak_ram_mb:,.1f} MB\")\n",
    "print(f\"RAM reduction                        : {teacher_peak_ram_mb/student_peak_ram_mb:,.1f}x smaller\")\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "id": "ad4e83d9",
   "metadata": {
    "execution": {
     "iopub.execute_input": "2026-08-12T16:22:38.088375Z",
     "iopub.status.busy": "2026-08-12T16:22:38.088277Z",
     "iopub.status.idle": "2026-08-12T16:22:38.092993Z",
     "shell.execute_reply": "2026-08-12T16:22:38.092631Z"
    }
   },
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>Metric</th>\n",
       "      <th>Teacher</th>\n",
       "      <th>Student</th>\n",
       "      <th>Reduction</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>Disk size (MB)</td>\n",
       "      <td>417.90</td>\n",
       "      <td>11.10</td>\n",
       "      <td>37.7x</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>CPU latency (ms/query)</td>\n",
       "      <td>27.76</td>\n",
       "      <td>1.18</td>\n",
       "      <td>23.5x</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>Peak RAM (MB)</td>\n",
       "      <td>893.80</td>\n",
       "      <td>237.00</td>\n",
       "      <td>3.8x</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "                   Metric  Teacher  Student Reduction\n",
       "0          Disk size (MB)   417.90    11.10     37.7x\n",
       "1  CPU latency (ms/query)    27.76     1.18     23.5x\n",
       "2           Peak RAM (MB)   893.80   237.00      3.8x"
      ]
     },
     "execution_count": 24,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "deployment_df = pd.DataFrame([\n",
    "    {\"Metric\": \"Disk size (MB)\", \"Teacher\": round(teacher_size_mb, 1), \"Student\": round(student_size_mb, 1),\n",
    "     \"Reduction\": f\"{teacher_size_mb/student_size_mb:.1f}x\"},\n",
    "    {\"Metric\": \"CPU latency (ms/query)\", \"Teacher\": round(teacher_latency_mean, 2), \"Student\": round(student_latency_mean, 2),\n",
    "     \"Reduction\": f\"{teacher_latency_mean/student_latency_mean:.1f}x\"},\n",
    "    {\"Metric\": \"Peak RAM (MB)\", \"Teacher\": round(teacher_peak_ram_mb, 1), \"Student\": round(student_peak_ram_mb, 1),\n",
    "     \"Reduction\": f\"{teacher_peak_ram_mb/student_peak_ram_mb:.1f}x\"},\n",
    "])\n",
    "deployment_df\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "id": "e130ff8c",
   "metadata": {
    "execution": {
     "iopub.execute_input": "2026-08-12T16:22:38.094229Z",
     "iopub.status.busy": "2026-08-12T16:22:38.094165Z",
     "iopub.status.idle": "2026-08-12T16:22:38.097467Z",
     "shell.execute_reply": "2026-08-12T16:22:38.097095Z"
    }
   },
   "outputs": [
    {
     "data": {
      "text/markdown": [
       "\n",
       "#### Deployment-readiness analysis\n",
       "\n",
       "| Metric | Teacher | Student | Reduction |\n",
       "|---|---|---|---|\n",
       "| Disk size (MB) | 417.9 | 11.1 | 37.7x |\n",
       "| CPU latency (ms/query) | 27.76 | 1.18 | 23.5x |\n",
       "| Peak RAM (MB) | 893.8 | 237.0 | 3.8x |\n",
       "\n",
       "The distilled Student is **37.7x smaller on disk**, **23.5x faster**\n",
       "per CPU query, and uses **3.8x less peak RAM** than the Teacher, while retaining **99.9%** of its accuracy on the exact same 77-way classification task.\n",
       "\n",
       "**Is this deployment-ready for edge/mobile?**\n",
       "\n",
       "- **Size and RAM** — at ~11MB on disk and ~237MB of peak RAM, the\n",
       "  Student comfortably fits within the memory budgets of edge devices and mobile apps, where a\n",
       "  418MB+ BERT-base checkpoint is frequently a non-starter (app-store bundle-size limits,\n",
       "  low-RAM Android devices, on-device model caches).\n",
       "- **Latency** — 1.2ms/query on CPU is well within the range needed for a responsive,\n",
       "  synchronous UI interaction (e.g. intent routing as a user types), whereas the Teacher's\n",
       "  27.8ms/query, multiplied across a request queue on a resource-constrained device,\n",
       "  would noticeably degrade perceived responsiveness.\n",
       "- **Accuracy trade-off** — the real cost is the **0.1 percentage points** of relative accuracy given up. Whether that is acceptable depends entirely on the product: for a first-pass intent router that falls back to a human agent or a larger cloud model on low confidence, this trade is usually a clear win — the compute/latency/cost savings at massive query volume outweigh a modest accuracy gap. For a fully autonomous decision with no fallback (e.g. auto-approving a refund), the remaining gap to the Teacher may still be too large to deploy the Student *alone*.\n",
       "- **Practical recommendation** — deploy the distilled Student as the default path, and route low-confidence\n",
       "  predictions (small margin between the top-2 softmax probabilities) to the Teacher or a human reviewer.\n",
       "  This captures most of the size/latency/RAM benefits demonstrated above while bounding the accuracy risk\n",
       "  to only the genuinely ambiguous cases — which is exactly the scenario dark-knowledge distillation is\n",
       "  suited for, since the Student was trained to mimic the Teacher's *confidence structure*, not just its\n",
       "  argmax.\n"
      ],
      "text/plain": [
       "<IPython.core.display.Markdown object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "accuracy_retained_pct = distilled_acc / teacher_acc * 100\n",
    "size_reduction_x = teacher_size_mb / student_size_mb\n",
    "latency_reduction_x = teacher_latency_mean / student_latency_mean\n",
    "ram_reduction_x = teacher_peak_ram_mb / student_peak_ram_mb\n",
    "\n",
    "if distilled_acc >= teacher_acc:\n",
    "    accuracy_summary_line = (\n",
    "        f\"while matching (here, slightly exceeding) its accuracy on the exact same 77-way classification task.\"\n",
    "    )\n",
    "    accuracy_tradeoff_bullet = (\n",
    "        f\"- **Accuracy trade-off** — in this run there isn't one: the distilled Student's accuracy is at \"\n",
    "        f\"least as high as the Teacher's, so the size/latency/RAM wins below come essentially for free on \"\n",
    "        f\"this test set. That is a favorable outcome, not a guarantee — it reflects a lightly-fine-tuned \"\n",
    "        f\"Teacher and a modest-size test set as much as it reflects the Student's quality, so a production \"\n",
    "        f\"rollout should still monitor accuracy on live traffic rather than assuming this margin holds \"\n",
    "        f\"indefinitely as data drifts.\"\n",
    "    )\n",
    "else:\n",
    "    accuracy_summary_line = (\n",
    "        f\"while retaining **{accuracy_retained_pct:.1f}%** of its accuracy on the exact same 77-way \"\n",
    "        f\"classification task.\"\n",
    "    )\n",
    "    accuracy_tradeoff_bullet = (\n",
    "        f\"- **Accuracy trade-off** — the real cost is the **{100-accuracy_retained_pct:.1f} percentage \"\n",
    "        f\"points** of relative accuracy given up. Whether that is acceptable depends entirely on the \"\n",
    "        f\"product: for a first-pass intent router that falls back to a human agent or a larger cloud model \"\n",
    "        f\"on low confidence, this trade is usually a clear win — the compute/latency/cost savings at massive \"\n",
    "        f\"query volume outweigh a modest accuracy gap. For a fully autonomous decision with no fallback \"\n",
    "        f\"(e.g. auto-approving a refund), the remaining gap to the Teacher may still be too large to deploy \"\n",
    "        f\"the Student *alone*.\"\n",
    "    )\n",
    "\n",
    "display(Markdown(f'''\n",
    "#### Deployment-readiness analysis\n",
    "\n",
    "{to_markdown_table(deployment_df)}\n",
    "\n",
    "The distilled Student is **{size_reduction_x:.1f}x smaller on disk**, **{latency_reduction_x:.1f}x faster**\n",
    "per CPU query, and uses **{ram_reduction_x:.1f}x less peak RAM** than the Teacher, {accuracy_summary_line}\n",
    "\n",
    "**Is this deployment-ready for edge/mobile?**\n",
    "\n",
    "- **Size and RAM** — at ~{student_size_mb:.0f}MB on disk and ~{student_peak_ram_mb:.0f}MB of peak RAM, the\n",
    "  Student comfortably fits within the memory budgets of edge devices and mobile apps, where a\n",
    "  {teacher_size_mb:.0f}MB+ BERT-base checkpoint is frequently a non-starter (app-store bundle-size limits,\n",
    "  low-RAM Android devices, on-device model caches).\n",
    "- **Latency** — {student_latency_mean:.1f}ms/query on CPU is well within the range needed for a responsive,\n",
    "  synchronous UI interaction (e.g. intent routing as a user types), whereas the Teacher's\n",
    "  {teacher_latency_mean:.1f}ms/query, multiplied across a request queue on a resource-constrained device,\n",
    "  would noticeably degrade perceived responsiveness.\n",
    "{accuracy_tradeoff_bullet}\n",
    "- **Practical recommendation** — deploy the distilled Student as the default path, and route low-confidence\n",
    "  predictions (small margin between the top-2 softmax probabilities) to the Teacher or a human reviewer.\n",
    "  This captures most of the size/latency/RAM benefits demonstrated above while bounding the accuracy risk\n",
    "  to only the genuinely ambiguous cases — which is exactly the scenario dark-knowledge distillation is\n",
    "  suited for, since the Student was trained to mimic the Teacher's *confidence structure*, not just its\n",
    "  argmax.\n",
    "'''))\n"
   ]
  }
 ],
 "metadata": {
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.12.8"
  },
  "widgets": {
   "application/vnd.jupyter.widget-state+json": {
    "state": {
     "088fb781006d43ada251ba3f5c985b46": {
      "model_module": "@jupyter-widgets/controls",
      "model_module_version": "2.0.0",
      "model_name": "HTMLStyleModel",
      "state": {
       "_model_module": "@jupyter-widgets/controls",
       "_model_module_version": "2.0.0",
       "_model_name": "HTMLStyleModel",
       "_view_count": null,
       "_view_module": "@jupyter-widgets/base",
       "_view_module_version": "2.0.0",
       "_view_name": "StyleView",
       "background": null,
       "description_width": "",
       "font_size": null,
       "text_color": null
      }
     },
     "18f7003fa73b4f83a4cb60efa1363936": {
      "model_module": "@jupyter-widgets/base",
      "model_module_version": "2.0.0",
      "model_name": "LayoutModel",
      "state": {
       "_model_module": "@jupyter-widgets/base",
       "_model_module_version": "2.0.0",
       "_model_name": "LayoutModel",
       "_view_count": null,
       "_view_module": "@jupyter-widgets/base",
       "_view_module_version": "2.0.0",
       "_view_name": "LayoutView",
       "align_content": null,
       "align_items": null,
       "align_self": null,
       "border_bottom": null,
       "border_left": null,
       "border_right": null,
       "border_top": null,
       "bottom": null,
       "display": null,
       "flex": null,
       "flex_flow": null,
       "grid_area": null,
       "grid_auto_columns": null,
       "grid_auto_flow": null,
       "grid_auto_rows": null,
       "grid_column": null,
       "grid_gap": null,
       "grid_row": null,
       "grid_template_areas": null,
       "grid_template_columns": null,
       "grid_template_rows": null,
       "height": null,
       "justify_content": null,
       "justify_items": null,
       "left": null,
       "margin": null,
       "max_height": null,
       "max_width": null,
       "min_height": null,
       "min_width": null,
       "object_fit": null,
       "object_position": null,
       "order": null,
       "overflow": null,
       "padding": null,
       "right": null,
       "top": null,
       "visibility": null,
       "width": null
      }
     },
     "2db8f5a7e9fa4b6e85a21dd7cb3ef687": {
      "model_module": "@jupyter-widgets/controls",
      "model_module_version": "2.0.0",
      "model_name": "HTMLModel",
      "state": {
       "_dom_classes": [],
       "_model_module": "@jupyter-widgets/controls",
       "_model_module_version": "2.0.0",
       "_model_name": "HTMLModel",
       "_view_count": null,
       "_view_module": "@jupyter-widgets/controls",
       "_view_module_version": "2.0.0",
       "_view_name": "HTMLView",
       "description": "",
       "description_allow_html": false,
       "layout": "IPY_MODEL_971915584dad4bffac849dda80f071b4",
       "placeholder": "​",
       "style": "IPY_MODEL_820cd91f7c314e1bae0469f922759721",
       "tabbable": null,
       "tooltip": null,
       "value": " 1/1 [00:00&lt;00:00, 15.26it/s]"
      }
     },
     "414f90d2277a489caf1c79d0b7c4b85e": {
      "model_module": "@jupyter-widgets/controls",
      "model_module_version": "2.0.0",
      "model_name": "HBoxModel",
      "state": {
       "_dom_classes": [],
       "_model_module": "@jupyter-widgets/controls",
       "_model_module_version": "2.0.0",
       "_model_name": "HBoxModel",
       "_view_count": null,
       "_view_module": "@jupyter-widgets/controls",
       "_view_module_version": "2.0.0",
       "_view_name": "HBoxView",
       "box_style": "",
       "children": [
        "IPY_MODEL_4596d05480454e6c9dca466f4e944382",
        "IPY_MODEL_4b8c4a079fa8480389a9ff82d79db71d",
        "IPY_MODEL_b3e7de65e5224472aac841f802e01a75"
       ],
       "layout": "IPY_MODEL_d755eff3c6384fc9be70547c60cc8320",
       "tabbable": null,
       "tooltip": null
      }
     },
     "4596d05480454e6c9dca466f4e944382": {
      "model_module": "@jupyter-widgets/controls",
      "model_module_version": "2.0.0",
      "model_name": "HTMLModel",
      "state": {
       "_dom_classes": [],
       "_model_module": "@jupyter-widgets/controls",
       "_model_module_version": "2.0.0",
       "_model_name": "HTMLModel",
       "_view_count": null,
       "_view_module": "@jupyter-widgets/controls",
       "_view_module_version": "2.0.0",
       "_view_name": "HTMLView",
       "description": "",
       "description_allow_html": false,
       "layout": "IPY_MODEL_18f7003fa73b4f83a4cb60efa1363936",
       "placeholder": "​",
       "style": "IPY_MODEL_088fb781006d43ada251ba3f5c985b46",
       "tabbable": null,
       "tooltip": null,
       "value": "Loading weights: 100%"
      }
     },
     "4b8c4a079fa8480389a9ff82d79db71d": {
      "model_module": "@jupyter-widgets/controls",
      "model_module_version": "2.0.0",
      "model_name": "FloatProgressModel",
      "state": {
       "_dom_classes": [],
       "_model_module": "@jupyter-widgets/controls",
       "_model_module_version": "2.0.0",
       "_model_name": "FloatProgressModel",
       "_view_count": null,
       "_view_module": "@jupyter-widgets/controls",
       "_view_module_version": "2.0.0",
       "_view_name": "ProgressView",
       "bar_style": "success",
       "description": "",
       "description_allow_html": false,
       "layout": "IPY_MODEL_ac4c9fc0c1da4a0b9b59bab752e70a2d",
       "max": 199,
       "min": 0,
       "orientation": "horizontal",
       "style": "IPY_MODEL_ae4f118d7d1041a7926ba4a6953439c9",
       "tabbable": null,
       "tooltip": null,
       "value": 199
      }
     },
     "63598b522ec749a387a3823e29fd6175": {
      "model_module": "@jupyter-widgets/controls",
      "model_module_version": "2.0.0",
      "model_name": "HTMLStyleModel",
      "state": {
       "_model_module": "@jupyter-widgets/controls",
       "_model_module_version": "2.0.0",
       "_model_name": "HTMLStyleModel",
       "_view_count": null,
       "_view_module": "@jupyter-widgets/base",
       "_view_module_version": "2.0.0",
       "_view_name": "StyleView",
       "background": null,
       "description_width": "",
       "font_size": null,
       "text_color": null
      }
     },
     "647e0e6626fe4b2d94cd4ded2e446145": {
      "model_module": "@jupyter-widgets/controls",
      "model_module_version": "2.0.0",
      "model_name": "HBoxModel",
      "state": {
       "_dom_classes": [],
       "_model_module": "@jupyter-widgets/controls",
       "_model_module_version": "2.0.0",
       "_model_name": "HBoxModel",
       "_view_count": null,
       "_view_module": "@jupyter-widgets/controls",
       "_view_module_version": "2.0.0",
       "_view_name": "HBoxView",
       "box_style": "",
       "children": [
        "IPY_MODEL_b816c7889b8a496cb46ff661ad8c5437",
        "IPY_MODEL_d4fbd03314474323b00ea93ea907de7e",
        "IPY_MODEL_2db8f5a7e9fa4b6e85a21dd7cb3ef687"
       ],
       "layout": "IPY_MODEL_dc3eb58c103f4c648d70063c9b973430",
       "tabbable": null,
       "tooltip": null
      }
     },
     "820cd91f7c314e1bae0469f922759721": {
      "model_module": "@jupyter-widgets/controls",
      "model_module_version": "2.0.0",
      "model_name": "HTMLStyleModel",
      "state": {
       "_model_module": "@jupyter-widgets/controls",
       "_model_module_version": "2.0.0",
       "_model_name": "HTMLStyleModel",
       "_view_count": null,
       "_view_module": "@jupyter-widgets/base",
       "_view_module_version": "2.0.0",
       "_view_name": "StyleView",
       "background": null,
       "description_width": "",
       "font_size": null,
       "text_color": null
      }
     },
     "971915584dad4bffac849dda80f071b4": {
      "model_module": "@jupyter-widgets/base",
      "model_module_version": "2.0.0",
      "model_name": "LayoutModel",
      "state": {
       "_model_module": "@jupyter-widgets/base",
       "_model_module_version": "2.0.0",
       "_model_name": "LayoutModel",
       "_view_count": null,
       "_view_module": "@jupyter-widgets/base",
       "_view_module_version": "2.0.0",
       "_view_name": "LayoutView",
       "align_content": null,
       "align_items": null,
       "align_self": null,
       "border_bottom": null,
       "border_left": null,
       "border_right": null,
       "border_top": null,
       "bottom": null,
       "display": null,
       "flex": null,
       "flex_flow": null,
       "grid_area": null,
       "grid_auto_columns": null,
       "grid_auto_flow": null,
       "grid_auto_rows": null,
       "grid_column": null,
       "grid_gap": null,
       "grid_row": null,
       "grid_template_areas": null,
       "grid_template_columns": null,
       "grid_template_rows": null,
       "height": null,
       "justify_content": null,
       "justify_items": null,
       "left": null,
       "margin": null,
       "max_height": null,
       "max_width": null,
       "min_height": null,
       "min_width": null,
       "object_fit": null,
       "object_position": null,
       "order": null,
       "overflow": null,
       "padding": null,
       "right": null,
       "top": null,
       "visibility": null,
       "width": null
      }
     },
     "a2f785a5bc9248908f4c9d10c1374be4": {
      "model_module": "@jupyter-widgets/base",
      "model_module_version": "2.0.0",
      "model_name": "LayoutModel",
      "state": {
       "_model_module": "@jupyter-widgets/base",
       "_model_module_version": "2.0.0",
       "_model_name": "LayoutModel",
       "_view_count": null,
       "_view_module": "@jupyter-widgets/base",
       "_view_module_version": "2.0.0",
       "_view_name": "LayoutView",
       "align_content": null,
       "align_items": null,
       "align_self": null,
       "border_bottom": null,
       "border_left": null,
       "border_right": null,
       "border_top": null,
       "bottom": null,
       "display": null,
       "flex": null,
       "flex_flow": null,
       "grid_area": null,
       "grid_auto_columns": null,
       "grid_auto_flow": null,
       "grid_auto_rows": null,
       "grid_column": null,
       "grid_gap": null,
       "grid_row": null,
       "grid_template_areas": null,
       "grid_template_columns": null,
       "grid_template_rows": null,
       "height": null,
       "justify_content": null,
       "justify_items": null,
       "left": null,
       "margin": null,
       "max_height": null,
       "max_width": null,
       "min_height": null,
       "min_width": null,
       "object_fit": null,
       "object_position": null,
       "order": null,
       "overflow": null,
       "padding": null,
       "right": null,
       "top": null,
       "visibility": null,
       "width": null
      }
     },
     "aa7c638e69b647c2941f619b51c783b0": {
      "model_module": "@jupyter-widgets/controls",
      "model_module_version": "2.0.0",
      "model_name": "HTMLStyleModel",
      "state": {
       "_model_module": "@jupyter-widgets/controls",
       "_model_module_version": "2.0.0",
       "_model_name": "HTMLStyleModel",
       "_view_count": null,
       "_view_module": "@jupyter-widgets/base",
       "_view_module_version": "2.0.0",
       "_view_name": "StyleView",
       "background": null,
       "description_width": "",
       "font_size": null,
       "text_color": null
      }
     },
     "ac4c9fc0c1da4a0b9b59bab752e70a2d": {
      "model_module": "@jupyter-widgets/base",
      "model_module_version": "2.0.0",
      "model_name": "LayoutModel",
      "state": {
       "_model_module": "@jupyter-widgets/base",
       "_model_module_version": "2.0.0",
       "_model_name": "LayoutModel",
       "_view_count": null,
       "_view_module": "@jupyter-widgets/base",
       "_view_module_version": "2.0.0",
       "_view_name": "LayoutView",
       "align_content": null,
       "align_items": null,
       "align_self": null,
       "border_bottom": null,
       "border_left": null,
       "border_right": null,
       "border_top": null,
       "bottom": null,
       "display": null,
       "flex": null,
       "flex_flow": null,
       "grid_area": null,
       "grid_auto_columns": null,
       "grid_auto_flow": null,
       "grid_auto_rows": null,
       "grid_column": null,
       "grid_gap": null,
       "grid_row": null,
       "grid_template_areas": null,
       "grid_template_columns": null,
       "grid_template_rows": null,
       "height": null,
       "justify_content": null,
       "justify_items": null,
       "left": null,
       "margin": null,
       "max_height": null,
       "max_width": null,
       "min_height": null,
       "min_width": null,
       "object_fit": null,
       "object_position": null,
       "order": null,
       "overflow": null,
       "padding": null,
       "right": null,
       "top": null,
       "visibility": null,
       "width": null
      }
     },
     "ae4f118d7d1041a7926ba4a6953439c9": {
      "model_module": "@jupyter-widgets/controls",
      "model_module_version": "2.0.0",
      "model_name": "ProgressStyleModel",
      "state": {
       "_model_module": "@jupyter-widgets/controls",
       "_model_module_version": "2.0.0",
       "_model_name": "ProgressStyleModel",
       "_view_count": null,
       "_view_module": "@jupyter-widgets/base",
       "_view_module_version": "2.0.0",
       "_view_name": "StyleView",
       "bar_color": null,
       "description_width": ""
      }
     },
     "b3e7de65e5224472aac841f802e01a75": {
      "model_module": "@jupyter-widgets/controls",
      "model_module_version": "2.0.0",
      "model_name": "HTMLModel",
      "state": {
       "_dom_classes": [],
       "_model_module": "@jupyter-widgets/controls",
       "_model_module_version": "2.0.0",
       "_model_name": "HTMLModel",
       "_view_count": null,
       "_view_module": "@jupyter-widgets/controls",
       "_view_module_version": "2.0.0",
       "_view_name": "HTMLView",
       "description": "",
       "description_allow_html": false,
       "layout": "IPY_MODEL_a2f785a5bc9248908f4c9d10c1374be4",
       "placeholder": "​",
       "style": "IPY_MODEL_63598b522ec749a387a3823e29fd6175",
       "tabbable": null,
       "tooltip": null,
       "value": " 199/199 [00:00&lt;00:00, 6547.74it/s]"
      }
     },
     "b816c7889b8a496cb46ff661ad8c5437": {
      "model_module": "@jupyter-widgets/controls",
      "model_module_version": "2.0.0",
      "model_name": "HTMLModel",
      "state": {
       "_dom_classes": [],
       "_model_module": "@jupyter-widgets/controls",
       "_model_module_version": "2.0.0",
       "_model_name": "HTMLModel",
       "_view_count": null,
       "_view_module": "@jupyter-widgets/controls",
       "_view_module_version": "2.0.0",
       "_view_name": "HTMLView",
       "description": "",
       "description_allow_html": false,
       "layout": "IPY_MODEL_d6e30f0f8a8b4a2d904b43980a735d40",
       "placeholder": "​",
       "style": "IPY_MODEL_aa7c638e69b647c2941f619b51c783b0",
       "tabbable": null,
       "tooltip": null,
       "value": "Writing model shards: 100%"
      }
     },
     "ca58ade0af764adda244d915d2a214a8": {
      "model_module": "@jupyter-widgets/controls",
      "model_module_version": "2.0.0",
      "model_name": "ProgressStyleModel",
      "state": {
       "_model_module": "@jupyter-widgets/controls",
       "_model_module_version": "2.0.0",
       "_model_name": "ProgressStyleModel",
       "_view_count": null,
       "_view_module": "@jupyter-widgets/base",
       "_view_module_version": "2.0.0",
       "_view_name": "StyleView",
       "bar_color": null,
       "description_width": ""
      }
     },
     "d4fbd03314474323b00ea93ea907de7e": {
      "model_module": "@jupyter-widgets/controls",
      "model_module_version": "2.0.0",
      "model_name": "FloatProgressModel",
      "state": {
       "_dom_classes": [],
       "_model_module": "@jupyter-widgets/controls",
       "_model_module_version": "2.0.0",
       "_model_name": "FloatProgressModel",
       "_view_count": null,
       "_view_module": "@jupyter-widgets/controls",
       "_view_module_version": "2.0.0",
       "_view_name": "ProgressView",
       "bar_style": "success",
       "description": "",
       "description_allow_html": false,
       "layout": "IPY_MODEL_f3f06f2f47a34c9c997d7816a3b0ff02",
       "max": 1,
       "min": 0,
       "orientation": "horizontal",
       "style": "IPY_MODEL_ca58ade0af764adda244d915d2a214a8",
       "tabbable": null,
       "tooltip": null,
       "value": 1
      }
     },
     "d6e30f0f8a8b4a2d904b43980a735d40": {
      "model_module": "@jupyter-widgets/base",
      "model_module_version": "2.0.0",
      "model_name": "LayoutModel",
      "state": {
       "_model_module": "@jupyter-widgets/base",
       "_model_module_version": "2.0.0",
       "_model_name": "LayoutModel",
       "_view_count": null,
       "_view_module": "@jupyter-widgets/base",
       "_view_module_version": "2.0.0",
       "_view_name": "LayoutView",
       "align_content": null,
       "align_items": null,
       "align_self": null,
       "border_bottom": null,
       "border_left": null,
       "border_right": null,
       "border_top": null,
       "bottom": null,
       "display": null,
       "flex": null,
       "flex_flow": null,
       "grid_area": null,
       "grid_auto_columns": null,
       "grid_auto_flow": null,
       "grid_auto_rows": null,
       "grid_column": null,
       "grid_gap": null,
       "grid_row": null,
       "grid_template_areas": null,
       "grid_template_columns": null,
       "grid_template_rows": null,
       "height": null,
       "justify_content": null,
       "justify_items": null,
       "left": null,
       "margin": null,
       "max_height": null,
       "max_width": null,
       "min_height": null,
       "min_width": null,
       "object_fit": null,
       "object_position": null,
       "order": null,
       "overflow": null,
       "padding": null,
       "right": null,
       "top": null,
       "visibility": null,
       "width": null
      }
     },
     "d755eff3c6384fc9be70547c60cc8320": {
      "model_module": "@jupyter-widgets/base",
      "model_module_version": "2.0.0",
      "model_name": "LayoutModel",
      "state": {
       "_model_module": "@jupyter-widgets/base",
       "_model_module_version": "2.0.0",
       "_model_name": "LayoutModel",
       "_view_count": null,
       "_view_module": "@jupyter-widgets/base",
       "_view_module_version": "2.0.0",
       "_view_name": "LayoutView",
       "align_content": null,
       "align_items": null,
       "align_self": null,
       "border_bottom": null,
       "border_left": null,
       "border_right": null,
       "border_top": null,
       "bottom": null,
       "display": null,
       "flex": null,
       "flex_flow": null,
       "grid_area": null,
       "grid_auto_columns": null,
       "grid_auto_flow": null,
       "grid_auto_rows": null,
       "grid_column": null,
       "grid_gap": null,
       "grid_row": null,
       "grid_template_areas": null,
       "grid_template_columns": null,
       "grid_template_rows": null,
       "height": null,
       "justify_content": null,
       "justify_items": null,
       "left": null,
       "margin": null,
       "max_height": null,
       "max_width": null,
       "min_height": null,
       "min_width": null,
       "object_fit": null,
       "object_position": null,
       "order": null,
       "overflow": null,
       "padding": null,
       "right": null,
       "top": null,
       "visibility": null,
       "width": null
      }
     },
     "dc3eb58c103f4c648d70063c9b973430": {
      "model_module": "@jupyter-widgets/base",
      "model_module_version": "2.0.0",
      "model_name": "LayoutModel",
      "state": {
       "_model_module": "@jupyter-widgets/base",
       "_model_module_version": "2.0.0",
       "_model_name": "LayoutModel",
       "_view_count": null,
       "_view_module": "@jupyter-widgets/base",
       "_view_module_version": "2.0.0",
       "_view_name": "LayoutView",
       "align_content": null,
       "align_items": null,
       "align_self": null,
       "border_bottom": null,
       "border_left": null,
       "border_right": null,
       "border_top": null,
       "bottom": null,
       "display": null,
       "flex": null,
       "flex_flow": null,
       "grid_area": null,
       "grid_auto_columns": null,
       "grid_auto_flow": null,
       "grid_auto_rows": null,
       "grid_column": null,
       "grid_gap": null,
       "grid_row": null,
       "grid_template_areas": null,
       "grid_template_columns": null,
       "grid_template_rows": null,
       "height": null,
       "justify_content": null,
       "justify_items": null,
       "left": null,
       "margin": null,
       "max_height": null,
       "max_width": null,
       "min_height": null,
       "min_width": null,
       "object_fit": null,
       "object_position": null,
       "order": null,
       "overflow": null,
       "padding": null,
       "right": null,
       "top": null,
       "visibility": null,
       "width": null
      }
     },
     "f3f06f2f47a34c9c997d7816a3b0ff02": {
      "model_module": "@jupyter-widgets/base",
      "model_module_version": "2.0.0",
      "model_name": "LayoutModel",
      "state": {
       "_model_module": "@jupyter-widgets/base",
       "_model_module_version": "2.0.0",
       "_model_name": "LayoutModel",
       "_view_count": null,
       "_view_module": "@jupyter-widgets/base",
       "_view_module_version": "2.0.0",
       "_view_name": "LayoutView",
       "align_content": null,
       "align_items": null,
       "align_self": null,
       "border_bottom": null,
       "border_left": null,
       "border_right": null,
       "border_top": null,
       "bottom": null,
       "display": null,
       "flex": null,
       "flex_flow": null,
       "grid_area": null,
       "grid_auto_columns": null,
       "grid_auto_flow": null,
       "grid_auto_rows": null,
       "grid_column": null,
       "grid_gap": null,
       "grid_row": null,
       "grid_template_areas": null,
       "grid_template_columns": null,
       "grid_template_rows": null,
       "height": null,
       "justify_content": null,
       "justify_items": null,
       "left": null,
       "margin": null,
       "max_height": null,
       "max_width": null,
       "min_height": null,
       "min_width": null,
       "object_fit": null,
       "object_position": null,
       "order": null,
       "overflow": null,
       "padding": null,
       "right": null,
       "top": null,
       "visibility": null,
       "width": null
      }
     }
    },
    "version_major": 2,
    "version_minor": 0
   }
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}