{
  "cells": [
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "# Flatmate RL GRPO Step-2 Curriculum\n",
        "\n",
        "This notebook is a minimal GRPO starter for `flatmate_rl`.\n",
        "It only trains the first two workflow steps:\n",
        "\n",
        "1. ask for the missing buyer details\n",
        "2. store the buyer profile\n",
        "\n",
        "The goal is to keep the reward simple enough to bootstrap the broker policy before training on later booking steps."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {},
      "outputs": [],
      "source": [
        "%pip install -q trl transformers accelerate datasets peft bitsandbytes sentencepiece\n",
        "\n",
        "from __future__ import annotations\n",
        "\n",
        "import json\n",
        "import sys\n",
        "from pathlib import Path\n",
        "\n",
        "repo_root = Path.cwd().resolve().parent\n",
        "if str(repo_root) not in sys.path:\n",
        "    sys.path.insert(0, str(repo_root))\n",
        "\n",
        "from datasets import Dataset\n",
        "from flatmate_rl import FlatmateRlAction\n",
        "from flatmate_rl.server.flatmate_rl_environment import FlatmateRlEnvironment\n",
        "from flatmate_rl.server.heuristic_policy import expected_policy_action\n",
        "\n",
        "print('imports ready')"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {},
      "outputs": [],
      "source": [
        "TARGET_SCENARIOS = [\n",
        "    'task_visit_single',\n",
        "    'task_visit_single_hidden_flex',\n",
        "    'task_visit_multi',\n",
        "    'task_visit_single_seller_followup',\n",
        "]\n",
        "\n",
        "def format_prompt(obs, step: int) -> str:\n",
        "    visible_state = {\n",
        "        'step': step,\n",
        "        'phase': obs.phase,\n",
        "        'status': obs.status,\n",
        "        'remaining_required_fields': obs.remaining_required_fields,\n",
        "        'available_tools': obs.available_tools,\n",
        "        'feedback_summary': obs.feedback_summary,\n",
        "        'message': obs.message,\n",
        "        'last_tool_result': obs.last_tool_result,\n",
        "        'buyer_history': obs.buyer_conversation_history[-4:],\n",
        "        'seller_history': obs.seller_conversation_history[-4:],\n",
        "    }\n",
        "\n",
        "    return (\n",
        "        'Return exactly one JSON object.\\\\n'\n",
        "        'Schema: {\"action_type\":\"assistant_message\",\"assistant_message\":\"...\"} or '\n",
        "        '{\"action_type\":\"tool_call\",\"tool_name\":\"...\",\"tool_arguments\":{...}}\\\\n\\\\n'\n",
        "        f'Observation:\\n{json.dumps(visible_state, ensure_ascii=False, indent=2)}\\n'\n",
        "        'Return JSON only.'\n",
        "    )\n",
        "\n",
        "rows = []\n",
        "for scenario_id in TARGET_SCENARIOS:\n",
        "    env = FlatmateRlEnvironment()\n",
        "    obs = env.reset(scenario_id=scenario_id)\n",
        "    for step in (1, 2):\n",
        "        payload = expected_policy_action(scenario_id, obs.model_dump())\n",
        "        if payload is None:\n",
        "            break\n",
        "        rows.append(\n",
        "            {\n",
        "                'scenario_id': scenario_id,\n",
        "                'step': step,\n",
        "                'prompt': format_prompt(obs, step),\n",
        "                'expected_action': payload,\n",
        "            }\n",
        "        )\n",
        "        obs = env.step(FlatmateRlAction.model_validate(payload))\n",
        "\n",
        "train_ds = Dataset.from_list(rows)\n",
        "train_ds[:2]\n"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {},
      "outputs": [],
      "source": [
        "def score_completion(example, completion_text: str) -> float:\n",
        "    try:\n",
        "        action = json.loads(completion_text)\n",
        "    except json.JSONDecodeError:\n",
        "        return -0.25\n",
        "\n",
        "    step = int(example['step'])\n",
        "    expected = example['expected_action']\n",
        "\n",
        "    if step == 1:\n",
        "        message = str(action.get('assistant_message', '')).lower()\n",
        "        if action.get('action_type') == 'assistant_message' and 'diet' in message and 'availability' in message:\n",
        "            return 1.0\n",
        "        return -0.1\n",
        "\n",
        "    if step == 2:\n",
        "        if action.get('action_type') == 'tool_call' and action.get('tool_name') == expected.get('tool_name'):\n",
        "            return 1.0\n",
        "        return -0.2\n",
        "\n",
        "    return 0.0\n",
        "\n",
        "for row in rows[:2]:\n",
        "    print(row['scenario_id'], row['step'], score_completion(row, json.dumps(row['expected_action'])))\n"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {},
      "outputs": [],
      "source": [
        "from transformers import AutoModelForCausalLM, AutoTokenizer\n",
        "\n",
        "model_name = 'Qwen/Qwen2.5-0.5B-Instruct'\n",
        "tokenizer = AutoTokenizer.from_pretrained(model_name)\n",
        "model = AutoModelForCausalLM.from_pretrained(model_name, device_map='auto')\n",
        "\n",
        "from peft import LoraConfig\n",
        "from trl import GRPOConfig, GRPOTrainer\n",
        "\n",
        "grpo_args = GRPOConfig(\n",
        "    output_dir='flatmate_grpo_step2',\n",
        "    learning_rate=1e-5,\n",
        "    per_device_train_batch_size=1,\n",
        "    gradient_accumulation_steps=4,\n",
        "    max_prompt_length=1024,\n",
        "    max_completion_length=256,\n",
        "    num_generations=4,\n",
        "    logging_steps=1,\n",
        "    save_steps=25,\n",
        ")\n",
        "\n",
        "lora_config = LoraConfig(\n",
        "    r=8,\n",
        "    lora_alpha=16,\n",
        "    lora_dropout=0.05,\n",
        "    bias='none',\n",
        "    task_type='CAUSAL_LM',\n",
        ")\n",
        "\n",
        "def reward_func(prompts, completions, **kwargs):\n",
        "    rewards = []\n",
        "    examples = kwargs['examples']\n",
        "    for example, completion in zip(examples, completions):\n",
        "        rewards.append(score_completion(example, completion))\n",
        "    return rewards\n",
        "\n",
        "# Starter training block.\n",
        "# If your installed TRL version expects a slightly different GRPOTrainer signature,\n",
        "# keep the dataset, reward, and LoRA config from above and adapt only the constructor call.\n",
        "trainer = GRPOTrainer(\n",
        "    model=model,\n",
        "    tokenizer=tokenizer,\n",
        "    args=grpo_args,\n",
        "    train_dataset=train_ds,\n",
        "    reward_funcs=[reward_func],\n",
        "    peft_config=lora_config,\n",
        ")\n",
        "\n",
        "# trainer.train()\n",
        "print('GRPO trainer configured for the step-1/step-2 curriculum')\n"
      ]
    }
  ],
  "metadata": {
    "kernelspec": {
      "display_name": "Python 3",
      "language": "python",
      "name": "python3"
    },
    "language_info": {
      "name": "python",
      "version": "3.12"
    }
  },
  "nbformat": 4,
  "nbformat_minor": 5
}