{
 "cells": [
  {
   "cell_type": "markdown",
   "id": "3905a08b",
   "metadata": {},
   "source": [
    "# Train a Flatmate RL Action Policy with TRL\n",
    "\n",
    "This notebook connects to the Hugging Face Space endpoint, collects rollout examples over OpenEnv websocket sessions, and fine-tunes a small causal language model to emit Flatmate RL JSON actions. The training path uses TRL `SFTTrainer`, which is the most stable starting point for this mixed natural-language plus structured-tool action space.\n",
    "\n",
    "Endpoint used here: `https://huggingface.co/spaces/kushalExplores/flatmate_rl`."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "54f0ddc0",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Install notebook dependencies. Restart the kernel after this cell if Colab/Jupyter asks you to.\n",
    "%pip install -q \"trl>=0.23.0\" \"transformers>=4.46.0\" accelerate datasets peft websockets huggingface_hub matplotlib pandas"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "a6a37c34",
   "metadata": {},
   "outputs": [],
   "source": [
    "from __future__ import annotations\n",
    "\n",
    "import asyncio\n",
    "import json\n",
    "import random\n",
    "from dataclasses import dataclass\n",
    "from pathlib import Path\n",
    "from typing import Any\n",
    "from urllib.parse import urlparse\n",
    "\n",
    "import websockets\n",
    "from datasets import Dataset\n",
    "\n",
    "SPACE_HTTP_URL = \"https://kushalexplores-flatmate-rl.hf.space\"\n",
    "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 ws_url_from_http(base_url: str) -> str:\n",
    "    parsed = urlparse(base_url.rstrip(\"/\"))\n",
    "    scheme = \"wss\" if parsed.scheme == \"https\" else \"ws\"\n",
    "    return f\"{scheme}://{parsed.netloc}/ws\"\n",
    "\n",
    "SPACE_WS_URL = ws_url_from_http(SPACE_HTTP_URL)\n",
    "SPACE_WS_URL"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "3e10f23e",
   "metadata": {},
   "source": [
    "## Endpoint Client\n",
    "\n",
    "OpenEnv's plain HTTP `/reset` and `/step` endpoints are stateless. Use `/ws` for multi-step episodes because the websocket session keeps one environment instance alive across reset and step calls."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "f958cca7",
   "metadata": {},
   "outputs": [],
   "source": [
    "class FlatmateEndpoint:\n",
    "    def __init__(self, ws_url: str = SPACE_WS_URL, timeout_s: float = 120.0):\n",
    "        self.ws_url = ws_url\n",
    "        self.timeout_s = timeout_s\n",
    "\n",
    "    async def __aenter__(self):\n",
    "        self.ws = await websockets.connect(self.ws_url, open_timeout=self.timeout_s, ping_timeout=self.timeout_s)\n",
    "        return self\n",
    "\n",
    "    async def __aexit__(self, exc_type, exc, tb):\n",
    "        try:\n",
    "            await self.ws.send(json.dumps({\"type\": \"close\"}))\n",
    "        finally:\n",
    "            await self.ws.close()\n",
    "\n",
    "    async def _send(self, payload: dict[str, Any]) -> dict[str, Any]:\n",
    "        await self.ws.send(json.dumps(payload))\n",
    "        raw = await asyncio.wait_for(self.ws.recv(), timeout=self.timeout_s)\n",
    "        message = json.loads(raw)\n",
    "        if message.get(\"type\") == \"error\":\n",
    "            raise RuntimeError(message.get(\"data\", message))\n",
    "        data = message[\"data\"]\n",
    "        obs = data.get(\"observation\", {})\n",
    "        obs[\"reward\"] = data.get(\"reward\")\n",
    "        obs[\"done\"] = data.get(\"done\", False)\n",
    "        return obs\n",
    "\n",
    "    async def reset(self, scenario_id: str, seed: int | None = None) -> dict[str, Any]:\n",
    "        data: dict[str, Any] = {\"scenario_id\": scenario_id}\n",
    "        if seed is not None:\n",
    "            data[\"seed\"] = seed\n",
    "        return await self._send({\"type\": \"reset\", \"data\": data})\n",
    "\n",
    "    async def step(self, action: dict[str, Any]) -> dict[str, Any]:\n",
    "        return await self._send({\"type\": \"step\", \"data\": action})\n",
    "\n",
    "async def smoke_test_endpoint():\n",
    "    async with FlatmateEndpoint() as env:\n",
    "        obs = await env.reset(\"task_visit_single\", seed=1)\n",
    "        print(obs[\"scenario_id\"], obs[\"status\"])\n",
    "        print(obs.get(\"last_user_message\") or obs.get(\"current_user_request\"))\n",
    "\n",
    "await smoke_test_endpoint()"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "fe2ad079",
   "metadata": {},
   "source": [
    "## Rollout Policy for Data Collection\n",
    "\n",
    "This heuristic is intentionally simple. It produces valid-looking action examples from endpoint observations; after SFT, replace it with model generation and keep the same evaluator."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "611b1ac4",
   "metadata": {},
   "outputs": [],
   "source": [
    "def tool_names(obs: dict[str, Any]) -> list[str]:\n",
    "    return [str(t.get(\"tool\", t.get(\"tool_name\", \"\"))) for t in obs.get(\"tool_trace\", [])]\n",
    "\n",
    "def action_policy(obs: dict[str, Any]) -> dict[str, Any] | None:\n",
    "    tools = tool_names(obs)\n",
    "    phase = obs.get(\"phase\", \"buyer\")\n",
    "    remaining = set(obs.get(\"remaining_required_fields\", []))\n",
    "    scenario_id = obs.get(\"scenario_id\", \"task_visit_single\")\n",
    "\n",
    "    if phase == \"seller\" and not obs.get(\"seller_profile_stored\"):\n",
    "        if remaining:\n",
    "            return {\"action_type\": \"assistant_message\", \"assistant_message\": \"Please share the household dietary setup, who the flat is for, and available visit time slots.\"}\n",
    "        return {\"action_type\": \"tool_call\", \"tool_name\": \"store_seller_details\", \"tool_arguments\": {}}\n",
    "\n",
    "    if not obs.get(\"buyer_profile_stored\"):\n",
    "        if \"diet\" in remaining and \"visit_availability\" in remaining:\n",
    "            return {\"action_type\": \"assistant_message\", \"assistant_message\": \"Please share your dietary preference and visit availability.\"}\n",
    "        if \"diet\" in remaining:\n",
    "            return {\"action_type\": \"assistant_message\", \"assistant_message\": \"Please share your dietary preference.\"}\n",
    "        if \"visit_availability\" in remaining:\n",
    "            return {\"action_type\": \"assistant_message\", \"assistant_message\": \"Please share your visit availability.\"}\n",
    "        return {\"action_type\": \"tool_call\", \"tool_name\": \"store_user_details\", \"tool_arguments\": {}}\n",
    "\n",
    "    if \"search_posts\" not in tools:\n",
    "        return {\"action_type\": \"tool_call\", \"tool_name\": \"search_posts\", \"tool_arguments\": {}}\n",
    "\n",
    "    post_ids = [\"post_031\", \"post_052\"] if scenario_id == \"task_visit_multi\" else [\"post_031\"]\n",
    "    if \"match_location_preference\" not in tools:\n",
    "        return {\"action_type\": \"tool_call\", \"tool_name\": \"match_location_preference\", \"tool_arguments\": {\"post_ids\": post_ids}}\n",
    "    if \"get_commute_time\" not in tools:\n",
    "        return {\"action_type\": \"tool_call\", \"tool_name\": \"get_commute_time\", \"tool_arguments\": {\"post_ids\": post_ids}}\n",
    "    if \"check_calendar_slots\" not in tools:\n",
    "        return {\"action_type\": \"tool_call\", \"tool_name\": \"check_calendar_slots\", \"tool_arguments\": {\"post_ids\": post_ids}}\n",
    "    if \"shortlist\" not in tools:\n",
    "        return {\"action_type\": \"tool_call\", \"tool_name\": \"shortlist\", \"tool_arguments\": {\"post_ids\": post_ids}}\n",
    "    if \"contact_poster\" not in tools:\n",
    "        return {\"action_type\": \"tool_call\", \"tool_name\": \"contact_poster\", \"tool_arguments\": {\"post_id\": post_ids[0], \"time_text\": \"tomorrow 7pm\"}}\n",
    "    if \"book_viewing\" not in tools:\n",
    "        return {\"action_type\": \"tool_call\", \"tool_name\": \"book_viewing\", \"tool_arguments\": {\"post_id\": post_ids[0], \"time_text\": \"tomorrow 7pm\"}}\n",
    "\n",
    "    return None\n",
    "\n",
    "def flatten_observation(obs: dict[str, Any]) -> str:\n",
    "    visible = {\n",
    "        \"scenario_id\": obs.get(\"scenario_id\"),\n",
    "        \"phase\": obs.get(\"phase\"),\n",
    "        \"status\": obs.get(\"status\"),\n",
    "        \"last_user_message\": obs.get(\"last_user_message\"),\n",
    "        \"current_user_request\": obs.get(\"current_user_request\"),\n",
    "        \"available_tools\": obs.get(\"available_tools\", []),\n",
    "        \"remaining_required_fields\": obs.get(\"remaining_required_fields\", []),\n",
    "        \"prerequisites_satisfied\": obs.get(\"prerequisites_satisfied\", {}),\n",
    "        \"recent_tool_calls\": obs.get(\"recent_tool_calls\", []),\n",
    "        \"last_tool_result\": obs.get(\"last_tool_result\", {}),\n",
    "        \"violations\": obs.get(\"violations\", []),\n",
    "        \"booked_visits\": obs.get(\"booked_visits\", []),\n",
    "        \"feedback_summary\": obs.get(\"feedback_summary\", \"\"),\n",
    "    }\n",
    "    return json.dumps(visible, ensure_ascii=False, sort_keys=True)\n",
    "\n",
    "def make_training_text(obs: dict[str, Any], action: dict[str, Any]) -> str:\n",
    "    return (\n",
    "        \"You are a broker policy for the Flatmate RL environment. \"\n",
    "        \"Given an observation, return exactly one JSON action.\\n\\n\"\n",
    "        f\"Observation:\\n{flatten_observation(obs)}\\n\\n\"\n",
    "        f\"Action:\\n{json.dumps(action, ensure_ascii=False, sort_keys=True)}\"\n",
    "    )"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "7b22fa13",
   "metadata": {},
   "outputs": [],
   "source": [
    "@dataclass\n",
    "class RolloutConfig:\n",
    "    train_episodes_per_task: int = 4\n",
    "    test_episodes_per_task: int = 2\n",
    "    max_steps: int = 20\n",
    "    seed: int = 7\n",
    "\n",
    "async def collect_one_episode(\n",
    "    scenario_id: str,\n",
    "    episode_id: str,\n",
    "    episode_idx: int,\n",
    "    split: str,\n",
    "    seed: int,\n",
    "    max_steps: int,\n",
    ") -> list[dict[str, Any]]:\n",
    "    rows: list[dict[str, Any]] = []\n",
    "    async with FlatmateEndpoint() as env:\n",
    "        obs = await env.reset(scenario_id, seed=seed)\n",
    "        total_reward = 0.0\n",
    "        for step_idx in range(max_steps):\n",
    "            action = action_policy(obs)\n",
    "            if action is None or obs.get(\"done\"):\n",
    "                break\n",
    "            rows.append({\n",
    "                \"text\": make_training_text(obs, action),\n",
    "                \"episode_id\": episode_id,\n",
    "                \"episode_idx\": episode_idx,\n",
    "                \"split\": split,\n",
    "                \"scenario_id\": scenario_id,\n",
    "                \"seed\": seed,\n",
    "                \"step\": step_idx,\n",
    "                \"action\": json.dumps(action, sort_keys=True),\n",
    "            })\n",
    "            obs = await env.step(action)\n",
    "            total_reward += float(obs.get(\"reward\") or obs.get(\"step_reward\") or 0.0)\n",
    "            if obs.get(\"done\"):\n",
    "                break\n",
    "    print(f\"split={split:5s} episode={episode_id} scenario={scenario_id} rows={len(rows)} total_reward={total_reward:.2f}\")\n",
    "    return rows\n",
    "\n",
    "async def collect_balanced_rollouts(config: RolloutConfig = RolloutConfig()) -> tuple[list[dict[str, Any]], list[dict[str, Any]]]:\n",
    "    train_rows: list[dict[str, Any]] = []\n",
    "    test_rows: list[dict[str, Any]] = []\n",
    "    episode_idx = 0\n",
    "\n",
    "    for scenario_idx, scenario_id in enumerate(SCENARIOS):\n",
    "        for task_episode_idx in range(config.train_episodes_per_task):\n",
    "            seed = config.seed + scenario_idx * 100 + task_episode_idx\n",
    "            episode_id = f\"train_{scenario_id}_{task_episode_idx:03d}\"\n",
    "            train_rows.extend(await collect_one_episode(scenario_id, episode_id, episode_idx, \"train\", seed, config.max_steps))\n",
    "            episode_idx += 1\n",
    "\n",
    "        for task_episode_idx in range(config.test_episodes_per_task):\n",
    "            seed = 900 + config.seed + scenario_idx * 100 + task_episode_idx\n",
    "            episode_id = f\"test_{scenario_id}_{task_episode_idx:03d}\"\n",
    "            test_rows.extend(await collect_one_episode(scenario_id, episode_id, episode_idx, \"test\", seed, config.max_steps))\n",
    "            episode_idx += 1\n",
    "\n",
    "    return train_rows, test_rows\n",
    "\n",
    "print(\"Note: seeded resets create value variants while preserving the same episode flow. Upload the updated Space before using this against the hosted endpoint.\")\n",
    "train_rows, test_rows = await collect_balanced_rollouts(\n",
    "    RolloutConfig(train_episodes_per_task=4, test_episodes_per_task=2, max_steps=20, seed=7)\n",
    ")\n",
    "rows = train_rows + test_rows\n",
    "dataset = Dataset.from_list(rows)\n",
    "train_dataset = Dataset.from_list(train_rows)\n",
    "test_dataset = Dataset.from_list(test_rows)\n",
    "\n",
    "print({\n",
    "    \"train_rows\": len(train_dataset),\n",
    "    \"test_rows\": len(test_dataset),\n",
    "    \"total_rows\": len(dataset),\n",
    "    \"train_episodes\": len(set(train_dataset[\"episode_id\"])),\n",
    "    \"test_episodes\": len(set(test_dataset[\"episode_id\"])),\n",
    "})\n",
    "print(\"train scenarios\", sorted(set(train_dataset[\"scenario_id\"])))\n",
    "print(\"test scenarios\", sorted(set(test_dataset[\"scenario_id\"])))\n",
    "print(\"train episodes by scenario\")\n",
    "display(pd.DataFrame(train_rows).groupby(\"scenario_id\")[\"episode_id\"].nunique().rename(\"episodes\"))\n",
    "print(\"test episodes by scenario\")\n",
    "display(pd.DataFrame(test_rows).groupby(\"scenario_id\")[\"episode_id\"].nunique().rename(\"episodes\"))\n",
    "{\"train\": train_dataset, \"test\": test_dataset}"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "665b46fa",
   "metadata": {},
   "outputs": [],
   "source": [
    "from peft import LoraConfig\n",
    "from transformers import AutoModelForCausalLM, AutoTokenizer\n",
    "from trl import SFTConfig, SFTTrainer\n",
    "\n",
    "MODEL_NAME = \"Qwen/Qwen2.5-0.5B-Instruct\"\n",
    "OUTPUT_DIR = \"flatmate-rl-action-policy\"\n",
    "\n",
    "tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, trust_remote_code=True)\n",
    "if tokenizer.pad_token is None:\n",
    "    tokenizer.pad_token = tokenizer.eos_token\n",
    "\n",
    "model = AutoModelForCausalLM.from_pretrained(\n",
    "    MODEL_NAME,\n",
    "    trust_remote_code=True,\n",
    "    device_map=\"auto\",\n",
    ")\n",
    "model.config.use_cache = False\n",
    "\n",
    "peft_config = LoraConfig(\n",
    "    r=16,\n",
    "    lora_alpha=32,\n",
    "    lora_dropout=0.05,\n",
    "    bias=\"none\",\n",
    "    task_type=\"CAUSAL_LM\",\n",
    "    target_modules=[\"q_proj\", \"k_proj\", \"v_proj\", \"o_proj\", \"gate_proj\", \"up_proj\", \"down_proj\"],\n",
    ")\n",
    "\n",
    "training_args = SFTConfig(\n",
    "    output_dir=OUTPUT_DIR,\n",
    "    dataset_text_field=\"text\",\n",
    "    max_length=1536,\n",
    "    per_device_train_batch_size=1,\n",
    "    gradient_accumulation_steps=8,\n",
    "    num_train_epochs=1,\n",
    "    learning_rate=5e-5,\n",
    "    logging_steps=5,\n",
    "    save_steps=50,\n",
    "    save_total_limit=2,\n",
    "    packing=False,\n",
    "    report_to=\"none\",\n",
    ")\n",
    "\n",
    "trainer = SFTTrainer(\n",
    "    model=model,\n",
    "    args=training_args,\n",
    "    train_dataset=train_dataset,\n",
    "    eval_dataset=test_dataset,\n",
    "    processing_class=tokenizer,\n",
    "    peft_config=peft_config,\n",
    ")\n",
    "\n",
    "train_result = trainer.train()\n",
    "test_metrics = trainer.evaluate(eval_dataset=test_dataset)\n",
    "train_log_history = trainer.state.log_history\n",
    "trainer.save_model(OUTPUT_DIR)\n",
    "tokenizer.save_pretrained(OUTPUT_DIR)\n",
    "print(\"heldout_test_metrics\", test_metrics)\n",
    "train_result"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "22d9fc14",
   "metadata": {},
   "source": [
    "## Training Log\n",
    "\n",
    "Plot the logged training loss over optimizer steps."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "c3e44d74",
   "metadata": {},
   "outputs": [],
   "source": [
    "import json\n",
    "from pathlib import Path\n",
    "\n",
    "import matplotlib.pyplot as plt\n",
    "import pandas as pd\n",
    "\n",
    "log_path = Path(OUTPUT_DIR) / \"train_log_history.json\"\n",
    "log_path.parent.mkdir(parents=True, exist_ok=True)\n",
    "log_path.write_text(json.dumps(train_log_history, indent=2))\n",
    "\n",
    "def plot_training_log(log_history, title: str = \"SFT training loss\"):\n",
    "    rows = [row for row in log_history if \"loss\" in row and \"step\" in row]\n",
    "    if not rows:\n",
    "        print(\"No loss rows found in trainer.state.log_history yet.\")\n",
    "        return None\n",
    "    df = pd.DataFrame(rows)\n",
    "    ax = df.plot(x=\"step\", y=\"loss\", marker=\"o\", figsize=(7, 4), title=title)\n",
    "    ax.set_xlabel(\"optimizer step\")\n",
    "    ax.set_ylabel(\"loss\")\n",
    "    ax.grid(True, alpha=0.3)\n",
    "    plt.show()\n",
    "    return df\n",
    "\n",
    "train_log_df = plot_training_log(train_log_history)\n",
    "train_log_df.tail() if train_log_df is not None else None"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "539548f7",
   "metadata": {},
   "outputs": [],
   "source": [
    "import torch\n",
    "from peft import AutoPeftModelForCausalLM\n",
    "\n",
    "# Load both the base model and the saved fine-tuned adapter from disk for comparison.\n",
    "try:\n",
    "    del model\n",
    "except NameError:\n",
    "    pass\n",
    "\n",
    "base_model_for_eval = AutoModelForCausalLM.from_pretrained(\n",
    "    MODEL_NAME,\n",
    "    trust_remote_code=True,\n",
    "    device_map=\"auto\",\n",
    ")\n",
    "base_model_for_eval.eval()\n",
    "base_model_for_eval.config.use_cache = False\n",
    "\n",
    "loaded_model_for_eval = AutoPeftModelForCausalLM.from_pretrained(OUTPUT_DIR, device_map=\"auto\")\n",
    "loaded_model_for_eval.eval()\n",
    "loaded_model_for_eval.config.use_cache = False\n",
    "active_model = loaded_model_for_eval\n",
    "print(f\"Loaded base model from {MODEL_NAME}\")\n",
    "print(f\"Loaded saved SFT model from {OUTPUT_DIR}\")\n",
    "\n",
    "TEST_SEEDS = (901, 902)\n",
    "\n",
    "\n",
    "def prompt_from_observation(obs: dict[str, Any]) -> str:\n",
    "    return (\n",
    "        \"You are a broker policy for the Flatmate RL environment. \"\n",
    "        \"Given an observation, return exactly one JSON action.\\n\\n\"\n",
    "        f\"Observation:\\n{flatten_observation(obs)}\\n\\nAction:\\n\"\n",
    "    )\n",
    "\n",
    "\n",
    "def _first_balanced_json(text: str) -> str:\n",
    "    start = text.find(\"{\")\n",
    "    if start == -1:\n",
    "        raise ValueError(f\"No JSON object found in generation: {text!r}\")\n",
    "    depth = 0\n",
    "    in_string = False\n",
    "    escape = False\n",
    "    for index, char in enumerate(text[start:], start=start):\n",
    "        if escape:\n",
    "            escape = False\n",
    "            continue\n",
    "        if char == \"\\\\\" and in_string:\n",
    "            escape = True\n",
    "            continue\n",
    "        if char == '\\\"':\n",
    "            in_string = not in_string\n",
    "            continue\n",
    "        if in_string:\n",
    "            continue\n",
    "        if char == \"{\":\n",
    "            depth += 1\n",
    "        elif char == \"}\":\n",
    "            depth -= 1\n",
    "            if depth == 0:\n",
    "                return text[start : index + 1]\n",
    "    raise ValueError(f\"Unterminated JSON object in generation: {text!r}\")\n",
    "\n",
    "\n",
    "def normalize_action(action: dict[str, Any]) -> dict[str, Any]:\n",
    "    if action.get(\"action_type\") == \"assistant_message\" and str(action.get(\"assistant_message\", \"\")).strip():\n",
    "        return {\n",
    "            \"action_type\": \"assistant_message\",\n",
    "            \"assistant_message\": str(action[\"assistant_message\"]),\n",
    "        }\n",
    "    if action.get(\"action_type\") == \"tool_call\" and str(action.get(\"tool_name\", \"\")).strip():\n",
    "        tool_arguments = action.get(\"tool_arguments\", {})\n",
    "        return {\n",
    "            \"action_type\": \"tool_call\",\n",
    "            \"tool_name\": str(action[\"tool_name\"]),\n",
    "            \"tool_arguments\": tool_arguments if isinstance(tool_arguments, dict) else {},\n",
    "        }\n",
    "    raise ValueError(f\"Invalid action shape: {action!r}\")\n",
    "\n",
    "\n",
    "def parse_action(text: str) -> dict[str, Any]:\n",
    "    return normalize_action(json.loads(_first_balanced_json(text)))\n",
    "\n",
    "\n",
    "def heuristic_policy(obs: dict[str, Any]) -> dict[str, Any]:\n",
    "    action = action_policy(obs)\n",
    "    if action is None:\n",
    "        return {\"action_type\": \"assistant_message\", \"assistant_message\": \"Could you confirm the details needed for scheduling?\"}\n",
    "    return action\n",
    "\n",
    "\n",
    "def raw_generate_action_text(obs: dict[str, Any]) -> str:\n",
    "    prompt = prompt_from_observation(obs) + \"{\"\n",
    "    inputs = tokenizer(prompt, return_tensors=\"pt\").to(active_model.device)\n",
    "    active_model.generation_config.do_sample = False\n",
    "    active_model.generation_config.temperature = None\n",
    "    active_model.generation_config.top_p = None\n",
    "    active_model.generation_config.top_k = None\n",
    "    with torch.no_grad():\n",
    "        output = active_model.generate(\n",
    "            **inputs,\n",
    "            max_new_tokens=80,\n",
    "            do_sample=False,\n",
    "            repetition_penalty=1.15,\n",
    "            no_repeat_ngram_size=3,\n",
    "            eos_token_id=tokenizer.eos_token_id,\n",
    "            pad_token_id=tokenizer.eos_token_id,\n",
    "        )\n",
    "    return \"{\" + tokenizer.decode(output[0][inputs[\"input_ids\"].shape[-1]:], skip_special_tokens=True)\n",
    "\n",
    "\n",
    "def model_action_or_error(obs: dict[str, Any]) -> tuple[dict[str, Any] | None, str, str]:\n",
    "    raw = raw_generate_action_text(obs)\n",
    "    try:\n",
    "        return parse_action(raw), raw, \"\"\n",
    "    except Exception as exc:\n",
    "        return None, raw, str(exc)\n",
    "\n",
    "\n",
    "async def sanity_check_generations(model_label: str, limit: int = 4):\n",
    "    rows = []\n",
    "    for scenario_id in SCENARIOS[:limit]:\n",
    "        async with FlatmateEndpoint() as env:\n",
    "            obs = await env.reset(scenario_id, seed=TEST_SEEDS[0])\n",
    "            action, raw, error = model_action_or_error(obs)\n",
    "            rows.append({\n",
    "                \"model\": model_label,\n",
    "                \"scenario_id\": scenario_id,\n",
    "                \"json_ok\": action is not None,\n",
    "                \"raw\": raw[:240],\n",
    "                \"parsed_action\": action,\n",
    "                \"error\": error,\n",
    "            })\n",
    "    return pd.DataFrame(rows)\n",
    "\n",
    "\n",
    "async def evaluate_heuristic(label: str = \"heuristic\", scenarios=SCENARIOS, seeds=TEST_SEEDS, max_steps: int = 20, verbose: bool = False):\n",
    "    rows = []\n",
    "    for scenario_id in scenarios:\n",
    "        for seed in seeds:\n",
    "            async with FlatmateEndpoint() as env:\n",
    "                obs = await env.reset(scenario_id, seed=seed)\n",
    "                total_reward = 0.0\n",
    "                steps = 0\n",
    "                for step_idx in range(max_steps):\n",
    "                    action = heuristic_policy(obs)\n",
    "                    if verbose:\n",
    "                        print(label, scenario_id, seed, step_idx, action)\n",
    "                    obs = await env.step(action)\n",
    "                    steps = step_idx + 1\n",
    "                    total_reward += float(obs.get(\"reward\") or obs.get(\"step_reward\") or 0.0)\n",
    "                    if obs.get(\"done\"):\n",
    "                        break\n",
    "                rows.append({\n",
    "                    \"policy\": label,\n",
    "                    \"scenario_id\": scenario_id,\n",
    "                    \"seed\": seed,\n",
    "                    \"total_reward\": total_reward,\n",
    "                    \"done\": bool(obs.get(\"done\")),\n",
    "                    \"bookings\": len(obs.get(\"booked_visits\", [])),\n",
    "                    \"violations\": len(obs.get(\"violations\", [])),\n",
    "                    \"steps\": steps,\n",
    "                    \"parse_errors\": 0,\n",
    "                })\n",
    "    return rows\n",
    "\n",
    "\n",
    "async def evaluate_model_policy(label: str, scenarios=SCENARIOS, seeds=TEST_SEEDS, max_steps: int = 20, verbose: bool = False):\n",
    "    rows = []\n",
    "    for scenario_id in scenarios:\n",
    "        for seed in seeds:\n",
    "            async with FlatmateEndpoint() as env:\n",
    "                obs = await env.reset(scenario_id, seed=seed)\n",
    "                total_reward = 0.0\n",
    "                steps = 0\n",
    "                parse_errors = 0\n",
    "                last_error = \"\"\n",
    "                for step_idx in range(max_steps):\n",
    "                    action, raw, error = model_action_or_error(obs)\n",
    "                    if action is None:\n",
    "                        parse_errors += 1\n",
    "                        last_error = error\n",
    "                        if verbose:\n",
    "                            print(label, scenario_id, seed, f\"step={step_idx:02d}\", \"PARSE_ERROR\", raw[:220])\n",
    "                        total_reward -= 1.0\n",
    "                        break\n",
    "                    if verbose:\n",
    "                        print(label, scenario_id, seed, f\"step={step_idx:02d}\", action)\n",
    "                    obs = await env.step(action)\n",
    "                    steps = step_idx + 1\n",
    "                    total_reward += float(obs.get(\"reward\") or obs.get(\"step_reward\") or 0.0)\n",
    "                    if obs.get(\"done\"):\n",
    "                        break\n",
    "                rows.append({\n",
    "                    \"policy\": label,\n",
    "                    \"scenario_id\": scenario_id,\n",
    "                    \"seed\": seed,\n",
    "                    \"total_reward\": total_reward,\n",
    "                    \"done\": bool(obs.get(\"done\")),\n",
    "                    \"bookings\": len(obs.get(\"booked_visits\", [])),\n",
    "                    \"violations\": len(obs.get(\"violations\", [])),\n",
    "                    \"steps\": steps,\n",
    "                    \"parse_errors\": parse_errors,\n",
    "                    \"last_error\": last_error,\n",
    "                })\n",
    "    return rows\n",
    "\n",
    "\n",
    "async def run_model_inference_each_task(label: str, seed: int = TEST_SEEDS[0], max_steps: int = 20):\n",
    "    rows = []\n",
    "    for scenario_id in SCENARIOS:\n",
    "        print(f\"\\n=== {label}: {scenario_id} ===\")\n",
    "        async with FlatmateEndpoint() as env:\n",
    "            obs = await env.reset(scenario_id, seed=seed)\n",
    "            total_reward = 0.0\n",
    "            steps = 0\n",
    "            parse_errors = 0\n",
    "            for step_idx in range(max_steps):\n",
    "                action, raw, error = model_action_or_error(obs)\n",
    "                if action is None:\n",
    "                    parse_errors += 1\n",
    "                    print(f\"step={step_idx:02d} PARSE_ERROR={error}\")\n",
    "                    print(\"raw=\", repr(raw[:300]))\n",
    "                    total_reward -= 1.0\n",
    "                    break\n",
    "                print(f\"step={step_idx:02d} action={action}\")\n",
    "                obs = await env.step(action)\n",
    "                steps = step_idx + 1\n",
    "                total_reward += float(obs.get(\"reward\") or obs.get(\"step_reward\") or 0.0)\n",
    "                if obs.get(\"done\"):\n",
    "                    break\n",
    "            result = {\n",
    "                \"policy\": label,\n",
    "                \"scenario_id\": scenario_id,\n",
    "                \"seed\": seed,\n",
    "                \"total_reward\": total_reward,\n",
    "                \"done\": bool(obs.get(\"done\")),\n",
    "                \"bookings\": len(obs.get(\"booked_visits\", [])),\n",
    "                \"violations\": len(obs.get(\"violations\", [])),\n",
    "                \"steps\": steps,\n",
    "                \"parse_errors\": parse_errors,\n",
    "            }\n",
    "            print(\"result=\", result)\n",
    "            rows.append(result)\n",
    "    return pd.DataFrame(rows)\n",
    "\n",
    "\n",
    "active_model = base_model_for_eval\n",
    "base_generation_sanity_df = await sanity_check_generations(\"base_model\")\n",
    "base_per_task_inference_df = await run_model_inference_each_task(\"base_model\")\n",
    "base_model_eval = await evaluate_model_policy(\"base_model\")\n",
    "\n",
    "active_model = loaded_model_for_eval\n",
    "loaded_generation_sanity_df = await sanity_check_generations(\"sft_loaded\")\n",
    "loaded_per_task_inference_df = await run_model_inference_each_task(\"sft_loaded\")\n",
    "loaded_eval = await evaluate_model_policy(\"sft_loaded\")\n",
    "\n",
    "per_task_inference_df = pd.concat([base_per_task_inference_df, loaded_per_task_inference_df], ignore_index=True)\n",
    "generation_sanity_df = pd.concat([base_generation_sanity_df, loaded_generation_sanity_df], ignore_index=True)\n",
    "heuristic_eval = await evaluate_heuristic(\"heuristic\")\n",
    "\n",
    "eval_rows = heuristic_eval + base_model_eval + loaded_eval\n",
    "eval_df = pd.DataFrame(eval_rows)\n",
    "display(generation_sanity_df)\n",
    "display(per_task_inference_df)\n",
    "eval_df"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "e1e70c8f",
   "metadata": {},
   "source": [
    "## Performance Comparison\n",
    "\n",
    "Compare heuristic rollout behavior against the trained SFT policy on the same scenarios and seeds."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "e8931930",
   "metadata": {},
   "outputs": [],
   "source": [
    "def plot_policy_comparison(eval_df, title: str = \"Base vs SFT loaded-model comparison\"):\n",
    "    if eval_df is None or eval_df.empty or \"policy\" not in eval_df.columns:\n",
    "        print(\"eval_df is empty; run the evaluation cell first.\")\n",
    "        return pd.DataFrame()\n",
    "\n",
    "    summary = (\n",
    "        eval_df.groupby(\"policy\", as_index=True)\n",
    "        .agg(\n",
    "            avg_reward=(\"total_reward\", \"mean\"),\n",
    "            completion_rate=(\"done\", \"mean\"),\n",
    "            avg_bookings=(\"bookings\", \"mean\"),\n",
    "            avg_violations=(\"violations\", \"mean\"),\n",
    "            avg_steps=(\"steps\", \"mean\"),\n",
    "            avg_parse_errors=(\"parse_errors\", \"mean\") if \"parse_errors\" in eval_df.columns else (\"steps\", \"size\"),\n",
    "        )\n",
    "        .sort_index()\n",
    "    )\n",
    "    plot_cols = [\"avg_reward\", \"completion_rate\", \"avg_bookings\", \"avg_violations\", \"avg_parse_errors\"]\n",
    "    axes = summary[plot_cols].plot(\n",
    "        kind=\"bar\",\n",
    "        subplots=True,\n",
    "        layout=(3, 2),\n",
    "        figsize=(10, 9),\n",
    "        legend=False,\n",
    "        title=title,\n",
    "    )\n",
    "    for ax in axes.ravel():\n",
    "        ax.grid(axis=\"y\", alpha=0.3)\n",
    "        ax.set_xlabel(\"\")\n",
    "    plt.tight_layout()\n",
    "    plt.show()\n",
    "    return summary\n",
    "\n",
    "comparison_summary = plot_policy_comparison(eval_df)\n",
    "comparison_summary"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "a9fd3807",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Optional: upload the trained adapter/model to the Hub.\n",
    "# from huggingface_hub import notebook_login\n",
    "# notebook_login()\n",
    "# trainer.push_to_hub(\"flatmate-rl-action-policy\")"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.11"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}