{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Telecom Intent-to-Config Pipeline\n",
    "\n",
    "Fine-tune Qwen2.5-7B on your TMF921 intent dataset using QLoRA on Kaggle T4x2.\n",
    "\n",
    "## Step 1: Install Dependencies"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "source": [
    "!pip install -q transformers trl peft accelerate bitsandbytes datasets liger-kernel sentence-transformers huggingface-hub\n",
    "!pip install -q --upgrade transformers"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Step 2: Login to Hugging Face\n",
    "\n",
    "Get your token from https://huggingface.co/settings/tokens (needs `write` access)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "source": [
    "from huggingface_hub import notebook_login\n",
    "notebook_login()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Step 3: Download Scripts from Hub"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "source": [
    "!wget -q https://huggingface.co/nraptisss/telecom-intent-pipeline/resolve/main/train.py\n",
    "!wget -q https://huggingface.co/nraptisss/telecom-intent-pipeline/resolve/main/inference.py\n",
    "!wget -q https://huggingface.co/nraptisss/telecom-intent-pipeline/resolve/main/merge_and_push.py\n",
    "!wget -q https://huggingface.co/nraptisss/telecom-intent-pipeline/resolve/main/benchmark.py\n",
    "!ls -la"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Step 4: Run Training\n",
    "\n",
    "This takes ~2-3 hours on Kaggle T4x2 for 3 epochs on 30K samples.\n",
    "\n",
    "**Edit `train.py` first** if you want to change dataset, model, or hyperparameters."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "source": [
    "!python train.py"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Step 5: Test Inference"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "source": [
    "!python inference.py --intent \"Deploy a low-latency URLLC slice for autonomous drones in the harbor zone with 1ms latency and 99.999% reliability\""
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Step 6: Merge & Push to Hub"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "source": [
    "!python merge_and_push.py"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Step 7: Benchmark on Test Set"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "source": [
    "!python benchmark.py --max_samples 100 --output benchmark_results.json"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## View Results"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "source": [
    "import json\n",
    "with open('benchmark_results.json', 'r') as f:\n",
    "    data = json.load(f)\n",
    "\n",
    "print(f\"JSON Valid Rate: {data['summary']['json_valid_rate']:.1%}\")\n",
    "print(f\"Schema Compliance: {data['summary']['avg_schema_compliance']:.1%}\")\n",
    "if data['summary'].get('semantic_similarity_avg'):\n",
    "    print(f\"Semantic Similarity: {data['summary']['semantic_similarity_avg']:.3f}\")\n",
    "\n",
    "for layer, s in data['summary']['per_layer'].items():\n",
    "    print(f\"  {layer:20s} valid={s['valid_rate']:.1%} compliance={s['avg_compliance']:.1%}\")"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "name": "python",
   "version": "3.10.0"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 4
}