{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# 🐛 CodeDebugger — GRPO Training on Google Colab\n",
    "\n",
    "**Meta + Scalar OpenEnv Hackathon 2026**\n",
    "\n",
    "This notebook trains an LLM to fix buggy Python code using GRPO (Group Relative Policy Optimization) via TRL + Unsloth.\n",
    "\n",
    "Requirements: **T4 GPU** (free tier) or better."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Step 1: Install dependencies\n",
    "!pip install -q groq unsloth trl peft transformers datasets accelerate\n",
    "!pip install -q fastapi uvicorn streamlit pyyaml"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Step 2: Clone the repo\n",
    "!git clone https://github.com/YOUR_USERNAME/codedebugger.git\n",
    "%cd codedebugger"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Step 3: Set API key\n",
    "import os\n",
    "os.environ['GROQ_API_KEY'] = 'your-groq-api-key-here'  # Replace with your key"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Step 4: Verify environment works\n",
    "from codedebugger.env.codedebugger_env import CodeDebuggerEnv\n",
    "from codedebugger.data.bug_dataset import get_problem\n",
    "\n",
    "env = CodeDebuggerEnv(max_steps=3)\n",
    "obs = env.reset(problem_id='bug_001')\n",
    "print('Observation keys:', list(obs.keys()))\n",
    "print('Buggy code:', obs['buggy_code'])\n",
    "print('Error:', obs['error_message'])\n",
    "print()\n",
    "print(env.render())"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Step 5: Test with the correct fix\n",
    "problem = get_problem('bug_001')\n",
    "obs, reward, done, info = env.step(problem['correct_code'])\n",
    "print(f'Reward: {reward:.4f}')\n",
    "print(f'Done: {done}')\n",
    "print(f'Pass rate: {info[\"step_record\"][\"pass_rate\"]:.0%}')\n",
    "print(f'Reward breakdown: {info[\"step_record\"][\"reward_breakdown\"]}')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Step 6: Run baselines\n",
    "from codedebugger.training.run_baseline import run_no_change_baseline, run_correct_code_baseline\n",
    "\n",
    "no_change = run_no_change_baseline()\n",
    "oracle = run_correct_code_baseline()\n",
    "\n",
    "print(f\"\\nNo-change solve rate: {no_change['solve_rate']:.0%}\")\n",
    "print(f\"Oracle solve rate: {oracle['solve_rate']:.0%}\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Step 7: Train with GRPO\n",
    "from codedebugger.training.train_grpo import train, GRPOConfig\n",
    "\n",
    "config = GRPOConfig(\n",
    "    model_name='unsloth/Llama-3.2-3B-Instruct',\n",
    "    num_epochs=1,\n",
    "    batch_size=2,\n",
    "    group_size=4,\n",
    "    max_steps=100,\n",
    "    learning_rate=5e-6,\n",
    ")\n",
    "\n",
    "train(config)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Step 8: Run LLM fixer baseline (requires GROQ_API_KEY)\n",
    "from codedebugger.orchestrator import Orchestrator\n",
    "\n",
    "orch = Orchestrator(max_steps=3, max_episodes=5)\n",
    "results = orch.run()"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "name": "python",
   "version": "3.11.0"
  },
  "accelerator": "GPU",
  "gpuClass": "standard"
 },
 "nbformat": 4,
 "nbformat_minor": 4
}