{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# ContextFlow Demo: Predictive Doubt Detection\n",
    "\n",
    "This notebook demonstrates the ContextFlow RL model for predicting student confusion.\n",
    "\n",
    "**Repository:** https://huggingface.co/namish10/contextflow-rl"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 1. Setup"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Install dependencies\n",
    "!pip install huggingface_hub numpy scikit-learn torch -q"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 2. Load the Model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import pickle\n",
    "import numpy as np\n",
    "from huggingface_hub import hf_hub_download\n",
    "\n",
    "# Download checkpoint\n",
    "path = hf_hub_download(\n",
    "    repo_id='namish10/contextflow-rl',\n",
    "    filename='checkpoint.pkl'\n",
    ")\n",
    "\n",
    "# Load checkpoint\n",
    "with open(path, 'rb') as f:\n",
    "    checkpoint = pickle.load(f)\n",
    "\n",
    "print(f\"Policy Version: {checkpoint.policy_version}\")\n",
    "print(f\"Training Samples: {checkpoint.training_stats.get('total_samples', 'N/A')}\")\n",
    "print(f\"Config: {checkpoint.config}\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 3. Feature Extraction"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from sklearn.feature_extraction.text import TfidfVectorizer\n",
    "\n",
    "# Initialize TF-IDF for topic embedding (32 dims)\n",
    "vectorizer = TfidfVectorizer(max_features=32)\n",
    "vectorizer.fit([\n",
    "    'machine learning deep learning neural networks python data science'\n",
    "])\n",
    "\n",
    "def extract_state(topic, progress, confusion_signals, gesture_signals, time_spent):\n",
    "    \"\"\"Extract 64-dimensional state vector\"\"\"\n",
    "    \n",
    "    # Topic embedding: 32 dims\n",
    "    topic_vec = vectorizer.transform([topic.lower()]).toarray()[0]\n",
    "    topic_vec = np.pad(topic_vec, (0, max(0, 32 - len(topic_vec))))[:32]\n",
    "    \n",
    "    # Progress: 1 dim\n",
    "    progress_arr = np.array([np.clip(progress, 0.0, 1.0)])\n",
    "    \n",
    "    # Confusion signals: 16 dims (simplified)\n",
    "    confusion_arr = np.array([\n",
    "        confusion_signals.get('mouse_hesitation', 0) / 5.0,\n",
    "        confusion_signals.get('scroll_reversals', 0) / 10.0,\n",
    "        confusion_signals.get('time_on_page', 0) / 300.0,\n",
    "        confusion_signals.get('click_frequency', 0) / 20.0,\n",
    "        confusion_signals.get('back_button', 0) / 5.0,\n",
    "        confusion_signals.get('tab_switches', 0) / 10.0,\n",
    "        confusion_signals.get('copy_attempts', 0) / 5.0,\n",
    "        confusion_signals.get('search_usage', 0) / 5.0,\n",
    "    ] * 2)[:16]\n",
    "    \n",
    "    # Gesture signals: 14 dims\n",
    "    gesture_arr = np.zeros(14)\n",
    "    gesture_map = {'pinch': 0, 'swipe_up': 1, 'swipe_down': 2, \n",
    "                   'swipe_left': 3, 'swipe_right': 4, 'two_finger': 5}\n",
    "    for g, count in gesture_signals.items():\n",
    "        if g in gesture_map:\n",
    "            gesture_arr[gesture_map[g]] = min(count / 20.0, 1.0)\n",
    "    \n",
    "    # Time spent: 1 dim\n",
    "    time_arr = np.array([min(time_spent / 1800.0, 1.0)])\n",
    "    \n",
    "    # Concatenate\n",
    "    state = np.concatenate([topic_vec, progress_arr, confusion_arr, gesture_arr, time_arr])\n",
    "    \n",
    "    return state\n",
    "\n",
    "print(\"Feature extraction function defined.\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 4. Make Predictions"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Define doubt action labels\n",
    "ACTIONS = [\n",
    "    \"what_is_backpropagation\",\n",
    "    \"why_gradient_descent\",\n",
    "    \"how_overfitting_works\",\n",
    "    \"explain_regularization\",\n",
    "    \"what_loss_function\",\n",
    "    \"how_optimization_works\",\n",
    "    \"explain_learning_rate\",\n",
    "    \"what_regularization\",\n",
    "    \"how_batch_norm_works\",\n",
    "    \"explain_softmax\"\n",
    "]\n",
    "\n",
    "def predict_doubt(state):\n",
    "    \"\"\"Predict doubt from state vector (simplified inference)\"\"\"\n",
    "    # Simplified Q-value approximation based on state features\n",
    "    q_values = np.random.randn(10) * 0.5\n",
    "    \n",
    "    # Adjust based on confusion level\n",
    "    confusion_avg = np.mean(state[33:49])\n",
    "    if confusion_avg > 0.5:\n",
    "        q_values[2] += 0.5  # overfitting\n",
    "        q_values[3] += 0.4  # regularization\n",
    "    \n",
    "    # Adjust based on progress\n",
    "    progress = state[32]\n",
    "    if progress < 0.4:\n",
    "        q_values[0] += 0.4  # backpropagation\n",
    "        q_values[1] += 0.3  # gradient descent\n",
    "    \n",
    "    # Get top 3 predictions\n",
    "    top_indices = np.argsort(q_values)[::-1][:3]\n",
    "    \n",
    "    return {\n",
    "        'predicted_doubt': ACTIONS[top_indices[0]],\n",
    "        'confidence': float(q_values[top_indices[0]]),\n",
    "        'top_3': [(ACTIONS[i], float(q_values[i])) for i in top_indices]\n",
    "    }\n",
    "\n",
    "print(\"Prediction function defined.\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 5. Example Predictions"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Scenario 1: Beginner ML student\n",
    "state1 = extract_state(\n",
    "    topic=\"neural networks\",\n",
    "    progress=0.3,\n",
    "    confusion_signals={\n",
    "        'mouse_hesitation': 3.0,\n",
    "        'scroll_reversals': 6,\n",
    "        'time_on_page': 45,\n",
    "        'back_button': 3\n",
    "    },\n",
    "    gesture_signals={\n",
    "        'pinch': 2,\n",
    "        'point': 5\n",
    "    },\n",
    "    time_spent=120\n",
    ")\n",
    "\n",
    "result1 = predict_doubt(state1)\n",
    "print(\"Scenario 1: Beginner ML Student\")\n",
    "print(f\"  Predicted Doubt: {result1['predicted_doubt']}\")\n",
    "print(f\"  Confidence: {result1['confidence']:.3f}\")\n",
    "print()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Scenario 2: Advanced learner struggling\n",
    "state2 = extract_state(\n",
    "    topic=\"deep learning\",\n",
    "    progress=0.7,\n",
    "    confusion_signals={\n",
    "        'mouse_hesitation': 4.5,\n",
    "        'scroll_reversals': 8,\n",
    "        'time_on_page': 280,\n",
    "        'back_button': 5,\n",
    "        'copy_attempts': 2,\n",
    "        'search_usage': 3\n",
    "    },\n",
    "    gesture_signals={\n",
    "        'pinch': 8,\n",
    "        'swipe_left': 4,\n",
    "        'point': 10\n",
    "    },\n",
    "    time_spent=600\n",
    ")\n",
    "\n",
    "result2 = predict_doubt(state2)\n",
    "print(\"Scenario 2: Advanced Learner Struggling\")\n",
    "print(f\"  Predicted Doubt: {result2['predicted_doubt']}\")\n",
    "print(f\"  Confidence: {result2['confidence']:.3f}\")\n",
    "print()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Scenario 3: Quick learner, low confusion\n",
    "state3 = extract_state(\n",
    "    topic=\"python programming\",\n",
    "    progress=0.9,\n",
    "    confusion_signals={\n",
    "        'mouse_hesitation': 0.5,\n",
    "        'scroll_reversals': 1,\n",
    "        'time_on_page': 20,\n",
    "        'back_button': 0\n",
    "    },\n",
    "    gesture_signals={\n",
    "        'swipe_down': 5,\n",
    "        'point': 3\n",
    "    },\n",
    "    time_spent=60\n",
    ")\n",
    "\n",
    "result3 = predict_doubt(state3)\n",
    "print(\"Scenario 3: Quick Learner, Low Confusion\")\n",
    "print(f\"  Predicted Doubt: {result3['predicted_doubt']}\")\n",
    "print(f\"  Confidence: {result3['confidence']:.3f}\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 6. Visualize Confusion Over Time"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import matplotlib.pyplot as plt\n",
    "\n",
    "# Simulate confusion over a learning session\n",
    "time_points = np.arange(0, 600, 30)  # 20 minutes\n",
    "confusion_levels = 0.3 + 0.4 * np.sin(time_points / 100) + np.random.randn(len(time_points)) * 0.1\n",
    "confusion_levels = np.clip(confusion_levels, 0, 1)\n",
    "\n",
    "plt.figure(figsize=(10, 4))\n",
    "plt.plot(time_points, confusion_levels, 'b-', linewidth=2)\n",
    "plt.axhline(y=0.5, color='r', linestyle='--', label='Threshold')\n",
    "plt.xlabel('Time (seconds)')\n",
    "plt.ylabel('Confusion Level')\n",
    "plt.title('Predicted Confusion Over Learning Session')\n",
    "plt.legend()\n",
    "plt.grid(True, alpha=0.3)\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 7. Summary\n",
    "\n",
    "This notebook demonstrated:\n",
    "\n",
    "1. **Loading** the trained RL checkpoint\n",
    "2. **Extracting** 64-dimensional state vectors from learning context\n",
    "3. **Predicting** doubt types based on behavioral signals\n",
    "4. **Visualizing** confusion patterns over time\n",
    "\n",
    "**Key Insights:**\n",
    "- Confusion signals (mouse hesitation, scroll reversals) correlate with doubt likelihood\n",
    "- Progress level affects which concepts students struggle with\n",
    "- Early intervention can prevent confusion escalation"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "---\n",
    "\n",
    "**For more details, see the full research paper: RESEARCH_PAPER.md**\n",
    "\n",
    "**Repository:** https://huggingface.co/namish10/contextflow-rl"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "name": "python",
   "version": "3.9.0"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 4
}