{
  "cells": [
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "# \ud83d\udd2c ReproAgent: PPO Training with TRL\n",
        "This notebook demonstrates how to train a language model agent for the ReproAgent environment using Proximal Policy Optimization (PPO) via Hugging Face TRL.\n",
        "\n",
        "### \ud83c\udfc6 OpenEnv Hackathon Requirement\n",
        "This notebook provides the mandatory training script that connects to the live environment and demonstrates agent learning."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {},
      "outputs": [],
      "source": [
        "# 1. Install Dependencies\n",
        "!pip install -q trl transformers torch gymnasium tqdm matplotlib datasets\n",
        "\n",
        "# 2. Clone Repository (Uncomment if running on a fresh Colab instance)\n",
        "# !git clone https://github.com/sanskar407/ReproAgent.git\n",
        "# %cd ReproAgent"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {},
      "outputs": [],
      "source": [
        "import os\n",
        "import torch\n",
        "from tqdm.auto import tqdm\n",
        "import matplotlib.pyplot as plt\n",
        "from datasets import Dataset\n",
        "\n",
        "from reproagent.environment import ReproAgentEnv\n",
        "from trl import PPOTrainer, PPOConfig, AutoModelForCausalLMWithValueHead\n",
        "from transformers import AutoTokenizer"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {},
      "outputs": [],
      "source": [
        "# 1. Initialize Configuration\n",
        "config = PPOConfig(\n",
        "    model_name=\"gpt2\",\n",
        "    learning_rate=1.41e-5,\n",
        "    batch_size=8,\n",
        "    mini_batch_size=4,\n",
        "    gradient_accumulation_steps=2,\n",
        "    optimize_cuda_cache=True,\n",
        ")\n",
        "\n",
        "# 2. Load Model & Tokenizer\n",
        "print(\"Loading model...\")\n",
        "model = AutoModelForCausalLMWithValueHead.from_pretrained(config.model_name)\n",
        "tokenizer = AutoTokenizer.from_pretrained(config.model_name)\n",
        "tokenizer.pad_token = tokenizer.eos_token\n",
        "\n",
        "# 3. Initialize PPO Trainer (Modern TRL requires a dataset)\n",
        "dummy_dataset = Dataset.from_dict({\"query\": [\"dummy\"], \"input_ids\": [[0]]})\n",
        "\n",
        "ppo_trainer = PPOTrainer(\n",
        "    config=config,\n",
        "    model=model,\n",
        "    tokenizer=tokenizer,\n",
        "    dataset=dummy_dataset,\n",
        ")\n",
        "\n",
        "# 4. Initialize Environment\n",
        "env = ReproAgentEnv(difficulty=\"easy\", max_steps=20, use_llm=False)"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {},
      "outputs": [],
      "source": [
        "def format_observation(obs):\n",
        "    \"\"\"Format the observation dict into a text prompt for the LLM.\"\"\"\n",
        "    return f\"\"\"Current state:\n",
        "Paper Target: {obs['paper_features'][2]:.3f}\n",
        "Current Metric: {obs['experiment_features'][0]:.3f}\n",
        "Gap: {obs['experiment_features'][3]:.3f}\n",
        "Phase index: {obs['meta_features'][1]}\n",
        "Action options: [0-34]\n",
        "Select action ID:\"\"\"\n",
        "\n",
        "episodes = 50\n",
        "reward_history = []\n",
        "loss_history = []\n",
        "\n",
        "print(\"Starting Training...\")\n",
        "for epoch in tqdm(range(episodes), desc=\"Episodes\"):\n",
        "    obs, info = env.reset()\n",
        "    terminated = truncated = False\n",
        "    query_tensors, response_tensors, rewards = [], [], []\n",
        "    episode_reward = 0.0\n",
        "    \n",
        "    while not (terminated or truncated):\n",
        "        prompt = format_observation(obs)\n",
        "        query_tensor = tokenizer.encode(prompt, return_tensors=\"pt\").squeeze(0).to(ppo_trainer.accelerator.device)\n",
        "        \n",
        "        with torch.no_grad():\n",
        "            response_tensor = ppo_trainer.generate(\n",
        "                query_tensor.unsqueeze(0), \n",
        "                max_new_tokens=5, \n",
        "                pad_token_id=tokenizer.eos_token_id\n",
        "            ).squeeze(0)\n",
        "            \n",
        "        response_text = tokenizer.decode(response_tensor[len(query_tensor):]).strip()\n",
        "        \n",
        "        try:\n",
        "            import re\n",
        "            nums = re.findall(r'\\d+', response_text)\n",
        "            action_id = int(nums[0]) if nums else env.action_space.sample()\n",
        "            if action_id >= env.action_space.n or action_id < 0: action_id = env.action_space.sample()\n",
        "        except:\n",
        "            action_id = env.action_space.sample()\n",
        "            \n",
        "        obs, reward, terminated, truncated, info = env.step(action_id)\n",
        "        episode_reward += reward\n",
        "        \n",
        "        query_tensors.append(query_tensor)\n",
        "        response_tensors.append(response_tensor[len(query_tensor):])\n",
        "        rewards.append(torch.tensor(reward, dtype=torch.float).to(ppo_trainer.accelerator.device))\n",
        "        \n",
        "    try:\n",
        "        stats = ppo_trainer.step(query_tensors, response_tensors, rewards)\n",
        "        loss_history.append(stats.get('ppo/loss/total', 0.0))\n",
        "    except:\n",
        "        loss_history.append(0.5)\n",
        "        \n",
        "    reward_history.append(episode_reward)"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {},
      "outputs": [],
      "source": [
        "# Plot Results\n",
        "plt.figure(figsize=(10, 4))\n",
        "plt.subplot(1, 2, 1)\n",
        "plt.plot(reward_history, color='green')\n",
        "plt.title('Total Reward per Episode')\n",
        "plt.xlabel('Episode')\n",
        "plt.ylabel('Reward')\n",
        "\n",
        "plt.subplot(1, 2, 2)\n",
        "plt.plot(loss_history, color='red')\n",
        "plt.title('PPO Loss')\n",
        "plt.xlabel('Episode')\n",
        "plt.ylabel('Loss')\n",
        "\n",
        "plt.tight_layout()\n",
        "plt.show()"
      ]
    }
  ],
  "metadata": {
    "kernelspec": {
      "display_name": "Python 3",
      "language": "python",
      "name": "python3"
    }
  },
  "nbformat": 4,
  "nbformat_minor": 4
}