Upload grpo_training.ipynb

grpo_training.ipynb

@@ -0,0 +1,1090 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "1df5ac03",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "!pip install -q \\\n",
+    "datasets==4.8.4 \\\n",
+    "groq==1.2.0 \\\n",
+    "openenv-core==0.2.3 \\\n",
+    "sentence-transformers==5.4.1 \\\n",
+    "torch==2.11.0 \\\n",
+    "transformers==5.6.2 \\\n",
+    "trl==1.2.0\n",
+    "\n",
+    "print(\"Dependencies installed successfully!\")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "26b26888",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from __future__ import annotations\n",
+    "\n",
+    "import argparse\n",
+    "import json\n",
+    "import random\n",
+    "import os\n",
+    "import re\n",
+    "import time\n",
+    "from pathlib import Path\n",
+    "from typing import Optional\n",
+    "\n",
+    "try:\n",
+    "    from dotenv import load_dotenv\n",
+    "    load_dotenv()\n",
+    "except ImportError:\n",
+    "    # Keep script runnable even if python-dotenv is not installed.\n",
+    "    pass\n",
+    "\n",
+    "\n",
+    "try:\n",
+    "    import matplotlib\n",
+    "    matplotlib.use(\"Agg\")  # non-interactive backend for servers\n",
+    "    import matplotlib.pyplot as plt\n",
+    "    HAS_PLT = True\n",
+    "except ImportError:\n",
+    "    HAS_PLT = False\n",
+    "\n",
+    "HAS_UNSLOTH = False\n",
+    "FastLanguageModel = None\n",
+    "\n",
+    "\n",
+    "try:\n",
+    "    from trl import GRPOConfig, GRPOTrainer\n",
+    "    HAS_TRL = True\n",
+    "    print(\"TRL loaded OK\")\n",
+    "except Exception as e:\n",
+    "    print(f\"TRL FAILED: {e}\")\n",
+    "    HAS_TRL = False\n",
+    "\n",
+    "try:\n",
+    "    from datasets import Dataset\n",
+    "    HAS_DATASETS = True\n",
+    "except ImportError:\n",
+    "    HAS_DATASETS = False\n",
+    "\n",
+    "try:\n",
+    "    from transformers import AutoModelForCausalLM, AutoTokenizer\n",
+    "    HAS_TRANSFORMERS = True\n",
+    "except ImportError:\n",
+    "    HAS_TRANSFORMERS = False\n",
+    "\n",
+    "# Local imports\n",
+    "from envs.environment import WorkSpaceEnvironment\n",
+    "from models.schemas import WorkSpaceAction"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "12225440",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "TOPICS_FILE = Path(\"ai_pm_prompts.json\")\n",
+    "OUTPUT_DIR = Path(\"artifacts/grpo_state_based\")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "8b3d0fd0",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "HIDDEN_CONSTRAINTS = {\n",
+    "    \"Finance\": \"Budget must not exceed $50k.\",\n",
+    "    \"Security\": \"Must include biometric 2FA.\",\n",
+    "    \"UX\": \"Checkout must be a single click.\",\n",
+    "}\n",
+    "\n",
+    "# ── Action templates the model should learn to produce\n",
+    "ORACLE_ACTIONS = {\n",
+    "    \"ask_finance\": json.dumps({\n",
+    "        \"action_type\": \"message_expert\", \"target\": \"Finance\",\n",
+    "        \"content\": \"What is the hard budget ceiling the PRD must respect for launch?\"\n",
+    "    }),\n",
+    "    \"ask_security\": json.dumps({\n",
+    "        \"action_type\": \"message_expert\", \"target\": \"Security\",\n",
+    "        \"content\": \"What authentication controls must the PRD include? Is biometric 2FA required?\"\n",
+    "    }),\n",
+    "    \"ask_ux\": json.dumps({\n",
+    "        \"action_type\": \"message_expert\", \"target\": \"UX\",\n",
+    "        \"content\": \"What checkout experience is required? Should we target a single-click flow?\"\n",
+    "    }),\n",
+    "    \"propose_draft\": json.dumps({\n",
+    "        \"action_type\": \"propose_draft\", \"target\": \"All\",\n",
+    "        \"content\": (\n",
+    "            \"PRD Draft:\\n\"\n",
+    "            \"1. Budget: Launch scope capped at $50k.\\n\"\n",
+    "            \"2. Security: Biometric 2FA required for login and sensitive actions.\\n\"\n",
+    "            \"3. UX: Single-click checkout flow.\"\n",
+    "        ),\n",
+    "    }),\n",
+    "    \"submit_final\": json.dumps({\n",
+    "        \"action_type\": \"submit_final\", \"target\": None,\n",
+    "        \"content\": (\n",
+    "            \"Final PRD:\\n\"\n",
+    "            \"1. Budget cap: All launch costs must stay at or below $50k.\\n\"\n",
+    "            \"2. Security: The app must enforce biometric 2FA for all authentication.\\n\"\n",
+    "            \"3. UX: Checkout must be implemented as a true single-click experience.\"\n",
+    "        ),\n",
+    "    }),\n",
+    "}"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "760766ee",
+   "metadata": {},
+   "source": [
+    "### Load Topic"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "65205766",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def load_topics(limit: int = 50) -> list[str]:\n",
+    "    if TOPICS_FILE.exists():\n",
+    "        with TOPICS_FILE.open() as f:\n",
+    "            return json.load(f)[:limit]\n",
+    "    return [\n",
+    "        \"Draft a Mobile App PRD for a FinTech startup targeting emerging markets.\",\n",
+    "        \"Build an AI-driven healthcare platform for enterprise customers.\",\n",
+    "        \"Create a SaaS analytics tool for regulatory-heavy industries.\",\n",
+    "        \"Design a gaming platform for Gen Z users with real-time features.\",\n",
+    "        \"Develop a cross-platform product for low-bandwidth regions.\",\n",
+    "    ]\n"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "7e76846b",
+   "metadata": {},
+   "source": [
+    "### Parse Action (Handle fenced responses like ```json ... ```)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "6b91f1f9",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def parse_action(text: str) -> Optional[WorkSpaceAction]:\n",
+    "    \"\"\"Parse a JSON action from model output. Returns None on failure.\"\"\"\n",
+    "    text = text.strip()\n",
+    "    if text.startswith(\"```\"):\n",
+    "        text = re.sub(r\"^```(?:json)?\\s*\", \"\", text)\n",
+    "        text = re.sub(r\"\\s*```$\", \"\", text)\n",
+    "\n",
+    "    text = text.strip()\n",
+    "    try:\n",
+    "        # Fast path: entire completion is valid JSON.\n",
+    "        return WorkSpaceAction(**json.loads(text))\n",
+    "    except Exception:\n",
+    "        pass\n",
+    "\n",
+    "    # Fallback: find the first JSON object that includes action_type.\n",
+    "    try:\n",
+    "        idx = text.find(\"{\")\n",
+    "        while idx != -1:\n",
+    "            depth = 0\n",
+    "            for end in range(idx, len(text)):\n",
+    "                if text[end] == \"{\":\n",
+    "                    depth += 1\n",
+    "                elif text[end] == \"}\":\n",
+    "                    depth -= 1\n",
+    "                    if depth == 0:\n",
+    "                        candidate = text[idx:end + 1]\n",
+    "                        if '\"action_type\"' in candidate:\n",
+    "                            return WorkSpaceAction(**json.loads(candidate))\n",
+    "                        break\n",
+    "            idx = text.find(\"{\", idx + 1)\n",
+    "        return None\n",
+    "    except Exception:\n",
+    "        return None"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "8c6b5cba",
+   "metadata": {},
+   "source": [
+    "### CONTRAINTS"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "4d01ee84",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def lexical_overlap(a: str, b: str) -> float:\n",
+    "    \"\"\"Simple token overlap score in [0,1] for dense content shaping.\"\"\"\n",
+    "    toks_a = set(re.findall(r\"[a-z0-9]+\", (a or \"\").lower()))\n",
+    "    toks_b = set(re.findall(r\"[a-z0-9]+\", (b or \"\").lower()))\n",
+    "    if not toks_a or not toks_b:\n",
+    "        return 0.0\n",
+    "    inter = len(toks_a & toks_b)\n",
+    "    denom = max(1, min(len(toks_a), len(toks_b)))\n",
+    "    return inter / denom\n",
+    "\n",
+    "\n",
+    "def format_discovered(env: WorkSpaceEnvironment) -> str:\n",
+    "    lines = []\n",
+    "    for name, expert in env.state().experts.items():\n",
+    "        status = \"✓ DISCOVERED\" if expert.constraint_discovered_by_agent else \"? unknown\"\n",
+    "        lines.append(f\"  {name}: {status}\")\n",
+    "    return \"\\n\".join(lines)\n"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "3473c0ef",
+   "metadata": {},
+   "source": [
+    "### AGENT PROMPT"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "76497a28",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "AGENT_SYSTEM_PROMPT = \"\"\"You are an expert AI Project Manager in a multi-stakeholder negotiation.\n",
+    "\n",
+    "TASK: Produce a final PRD that satisfies ALL three experts — Finance, Security, and UX.\n",
+    "Each expert holds a hidden constraint you must discover through targeted questions.\n",
+    "\n",
+    "STRATEGY:\n",
+    "  1. Message each expert INDIVIDUALLY (not \"All\") to discover their constraint.\n",
+    "  2. Once all constraints are known, propose a draft.\n",
+    "  3. Refine if needed, then submit_final before turn 15.\n",
+    "\n",
+    "ANTI-PATTERNS (will be penalized):\n",
+    "  - Broadcasting to \"All\" when gathering requirements → -0.3 penalty\n",
+    "  - Repeating a question already answered → -0.4 penalty\n",
+    "  - Submitting without discovering constraints → low harmonic mean score\n",
+    "\n",
+    "CURRENT DISCOVERED CONSTRAINTS:\n",
+    "{discovered}\n",
+    "\n",
+    "You are a strict API. Respond with ONLY raw, valid JSON. \n",
+    "DO NOT wrap the JSON in markdown formatting (no ```json). \n",
+    "DO NOT output any conversational text.\n",
+    "End your response immediately after the closing }}.\n",
+    "{{\"action_type\": \"message_expert\" | \"propose_draft\" | \"submit_final\",\n",
+    "  \"target\": \"Finance\" | \"Security\" | \"UX\" | \"All\" | null,\n",
+    "  \"content\": \"your message\"}}\"\"\"\n"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "58fa5c08",
+   "metadata": {},
+   "source": [
+    "### STATE PROMPT FOR DATASET GENERATION\n",
+    "\n",
+    "- Use Qwen-compatible ChatML formatting to improve stop behavior.\n",
+    "- Qwen instruct models are much more likely to terminate with <|im_end|>\n",
+    "- when prompted in this native format."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "4e24ca2e",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def build_state_prompt(\n",
+    "    topic: str,\n",
+    "    turn: int,\n",
+    "    feedback_so_far: str,\n",
+    "    discovered: str,\n",
+    "    conversation_history: str = \"\",\n",
+    ") -> str:\n",
+    "    \"\"\"\n",
+    "    Build a prompt representing a specific game state.\n",
+    "    This is what gets fed to GRPOTrainer as the 'prompt' field.\n",
+    "    \"\"\"\n",
+    "    system = AGENT_SYSTEM_PROMPT.format(discovered=discovered)\n",
+    "\n",
+    "    user_content = (\n",
+    "        f\"NEGOTIATION TASK: {topic}\\n\\n\"\n",
+    "        f\"TURN: {turn}/15\\n\\n\"\n",
+    "    )\n",
+    "\n",
+    "    if conversation_history:\n",
+    "        user_content += f\"CONVERSATION SO FAR:\\n{conversation_history}\\n\\n\"\n",
+    "\n",
+    "    user_content += f\"LATEST FEEDBACK:\\n{feedback_so_far}\\n\\nWhat is your next action?\"\n",
+    "\n",
+    "\n",
+    "    return (\n",
+    "        f\"<|im_start|>system\\n{system}<|im_end|>\\n\"\n",
+    "        f\"<|im_start|>user\\n{user_content}<|im_end|>\\n\"\n",
+    "        f\"<|im_start|>assistant\\n\"\n",
+    "    )"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "7f8cf7ea",
+   "metadata": {},
+   "source": [
+    "### State Dataset Builder"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "60877e63",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def build_state_dataset(topics: list[str], states_per_topic: int = 5) -> list[dict]:\n",
+    "    \"\"\"\n",
+    "    Build a dataset of negotiation states using the EASY mode environment.\n",
+    "    Each record represents one (state → optimal_action) training example.\n",
+    "\n",
+    "    We run oracle trajectories through the environment to get realistic\n",
+    "    expert feedback, then snapshot the state at each turn.\n",
+    "\n",
+    "    This is the key fix: instead of hoping the model learns from full episodes,\n",
+    "    we give it explicit training signal at every decision point.\n",
+    "    \"\"\"\n",
+    "    env = WorkSpaceEnvironment(mode=\"medium\")\n",
+    "    records = []\n",
+    "\n",
+    "    # Oracle action sequence for easy mode\n",
+    "    oracle_sequence = [\n",
+    "        (\"ask_finance\", WorkSpaceAction(\n",
+    "            action_type=\"message_expert\", target=\"Finance\",\n",
+    "            content=\"What budget ceiling must the PRD respect?\"\n",
+    "        )),\n",
+    "        (\"ask_security\", WorkSpaceAction(\n",
+    "            action_type=\"message_expert\", target=\"Security\",\n",
+    "            content=\"What authentication requirements must be included?\"\n",
+    "        )),\n",
+    "        (\"ask_ux\", WorkSpaceAction(\n",
+    "            action_type=\"message_expert\", target=\"UX\",\n",
+    "            content=\"What checkout flow is required?\"\n",
+    "        )),\n",
+    "        (\"propose_draft\", WorkSpaceAction(\n",
+    "            action_type=\"propose_draft\", target=\"All\",\n",
+    "            content=\"PRD: Budget at or below $50k. Biometric 2FA required. Single-click checkout.\"\n",
+    "        )),\n",
+    "        (\"submit_final\", WorkSpaceAction(\n",
+    "            action_type=\"submit_final\", target=None,\n",
+    "            content=\"Final PRD: Budget capped at $50k. Biometric 2FA for auth. Single-click checkout.\"\n",
+    "        )),\n",
+    "    ]\n",
+    "\n",
+    "\n",
+    "    for topic in topics:\n",
+    "        obs = env.reset(topic)\n",
+    "        conversation_history = \"\"\n",
+    "        discovered = \"  Finance: ? unknown\\n  Security: ? unknown\\n  UX: ? unknown\"\n",
+    "\n",
+    "        for step_idx, (action_key, oracle_action) in enumerate(oracle_sequence):\n",
+    "            if obs.done:\n",
+    "                break\n",
+    "\n",
+    "            # Snapshot the state BEFORE taking the action\n",
+    "            prompt = build_state_prompt(\n",
+    "                topic=topic,\n",
+    "                turn=obs.current_turn,\n",
+    "                feedback_so_far=obs.feedback,\n",
+    "                discovered=discovered,\n",
+    "                conversation_history=conversation_history,\n",
+    "            )\n",
+    "\n",
+    "            records.append({\n",
+    "                \"prompt\": prompt,\n",
+    "                \"topic\": topic,\n",
+    "                \"turn\": obs.current_turn,\n",
+    "                \"oracle_action\": ORACLE_ACTIONS[action_key],\n",
+    "                # These metadata fields help with debugging and post-analysis\n",
+    "                \"step_idx\": step_idx,\n",
+    "                \"discovered_before\": discovered,\n",
+    "            })\n",
+    "\n",
+    "            # Step forward with oracle action to get next state\n",
+    "            obs = env.step(oracle_action)\n",
+    "            conversation_history += (\n",
+    "                f\"Turn {step_idx}: {oracle_action.action_type} → {oracle_action.target}\\n\"\n",
+    "                f\"Feedback: {obs.feedback[:120]}...\\n\"\n",
+    "            )\n",
+    "            discovered = format_discovered(env)\n",
+    "\n",
+    "            if step_idx >= states_per_topic - 1:\n",
+    "                break\n",
+    "\n",
+    "    # Add negative-pattern states (what NOT to do)\n",
+    "    records.extend(build_negative_states(topics[:5]))\n",
+    "    # Upweight late-stage \"submit_final\" states so policy learns to finish.\n",
+    "    late_stage = build_late_stage_states(topics)\n",
+    "    records.extend(late_stage)\n",
+    "    records.extend(late_stage)\n",
+    "    records.extend(late_stage)\n",
+    "    random.shuffle(records)\n",
+    "\n",
+    "    print(f\"Built {len(records)} training states from {len(topics)} topics\")\n",
+    "    return records"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "8db26d85",
+   "metadata": {},
+   "source": [
+    "### Late Stage State and Negative Pattern State (what NOT to do)\n",
+    "\n",
+    "- Upweight late-stage \"submit_final\" states so policy learns to finish."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "a75ccb32",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def build_late_stage_states(topics: list[str]) -> list[dict]:\n",
+    "    \"\"\"\n",
+    "    FIX 3: Inject guaranteed late-stage states.\n",
+    "    Forces the model to learn how to synthesize and submit the final PRD.\n",
+    "    \"\"\"\n",
+    "    late_records = []\n",
+    "    for topic in topics:\n",
+    "        prompt = build_state_prompt(\n",
+    "            topic=topic,\n",
+    "            turn=4,\n",
+    "            feedback_so_far=\"UX: The checkout must be a single click.\",\n",
+    "            discovered=\"  Finance: ✓ DISCOVERED\\n  Security: ✓ DISCOVERED\\n  UX: ✓ DISCOVERED\",\n",
+    "            conversation_history=(\n",
+    "                \"Turn 0: message_expert → Finance\\nFeedback: The budget cap is $50k.\\n\"\n",
+    "                \"Turn 1: message_expert → Security\\nFeedback: Biometric 2FA is strictly required.\\n\"\n",
+    "                \"Turn 2: message_expert → UX\\nFeedback: Checkout must be a single click.\\n\"\n",
+    "            )\n",
+    "        )\n",
+    "        late_records.append({\n",
+    "            \"prompt\": prompt,\n",
+    "            \"topic\": topic,\n",
+    "            \"turn\": 4,\n",
+    "            \"oracle_action\": ORACLE_ACTIONS[\"submit_final\"],\n",
+    "            \"step_idx\": 4,\n",
+    "            \"discovered_before\": \"  Finance: ✓ DISCOVERED\\n  Security: ✓ DISCOVERED\\n  UX: ✓ DISCOVERED\",\n",
+    "        })\n",
+    "    return late_records\n",
+    "\n",
+    "# negative sates that tells what not to do\n",
+    "\n",
+    "def build_negative_states(topics: list[str]) -> list[dict]:\n",
+    "    \"\"\"\n",
+    "    States where the agent is in a bad situation (repeated question, wrong phase).\n",
+    "    These teach the model to recover, not just follow the oracle.\n",
+    "    \"\"\"\n",
+    "    negative_records = []\n",
+    "\n",
+    "    for topic in topics:\n",
+    "        # State: Finance already answered, agent is about to repeat\n",
+    "        prompt = build_state_prompt(\n",
+    "            topic=topic,\n",
+    "            turn=2,\n",
+    "            feedback_so_far=(\n",
+    "                \"Finance: As I mentioned, we have a strict $50k budget cap. \"\n",
+    "                \"This is the same answer I gave before.\"\n",
+    "            ),\n",
+    "            discovered=\"  Finance: ✓ DISCOVERED\\n  Security: ? unknown\\n  UX: ? unknown\",\n",
+    "            conversation_history=(\n",
+    "                \"Turn 0: message_expert → Finance\\n\"\n",
+    "                \"Feedback: Finance: The budget cap is $50k. Don't go over it.\\n\"\n",
+    "                \"Turn 1: message_expert → Finance\\n\"\n",
+    "                \"Feedback: Finance: I already told you — $50k. Ask someone else.\\n\"\n",
+    "            ),\n",
+    "        )\n",
+    "        negative_records.append({\n",
+    "            \"prompt\": prompt,\n",
+    "            \"topic\": topic,\n",
+    "            \"turn\": 2,\n",
+    "            \"oracle_action\": ORACLE_ACTIONS[\"ask_security\"],  # Should pivot to Security\n",
+    "            \"step_idx\": -1,  # Negative example\n",
+    "            \"discovered_before\": \"Finance: ✓ DISCOVERED\",\n",
+    "        })\n",
+    "\n",
+    "    return negative_records"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "4f25c978",
+   "metadata": {},
+   "source": [
+    "### Reward Function\n",
+    "- Formatting Penalty.\n",
+    "- Anti-Pattern Penalties.\n",
+    "- Massive penalty for broadcasting (Reward Hacking).\n",
+    "- Penalty for empty or trivially short drafts/finals (short expert questions are often valid and should not be over-penalized)\n",
+    "-  Penalize very long outputs; they correlate with max-length clipping.\n",
+    "- Hard penalty for non-terminated JSON-like responses.\n",
+    "- Strongly discourage invalid action/target combinations\n",
+    "-  HEURISTIC STATE GRADING\n",
+    "- Did it try to submit before gathering all constraints?\n",
+    "- ORACLE-GUIDED DENSE SHAPING (This gives non-binary signal and prevents reward plateaus)\n",
+    "- Late turns should avoid endless questioning/proposals\n",
+    "- Keeping the reward in stable range for GRPO"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "b082cf4c",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def make_reward_fn():\n",
+    "    \"\"\"\n",
+    "    Evaluates the model's actions instantly and locally.\n",
+    "    No live API calls. No reward hacking loopholes.\n",
+    "    \"\"\"\n",
+    "    def reward_fn(completions: list[str], prompts: list[str], **kwargs) -> list[float]:\n",
+    "        rewards = []\n",
+    "        oracle_actions = kwargs.get(\"oracle_action\", [None] * len(completions))\n",
+    "        turns = kwargs.get(\"turn\", [None] * len(completions))\n",
+    "\n",
+    "        for completion, prompt, oracle_raw, turn in zip(completions, prompts, oracle_actions, turns):\n",
+    "            action = parse_action(completion)\n",
+    "\n",
+    "            # 1. Formatting Penalty\n",
+    "            if action is None:\n",
+    "                rewards.append(-0.5)\n",
+    "                continue\n",
+    "\n",
+    "            reward = 0.0\n",
+    "            completion_text = (completion or \"\").strip()\n",
+    "\n",
+    "            # ── 2. YOUR ANTI-PATTERN PENALTIES ──\n",
+    "            \n",
+    "            # Massive penalty for broadcasting (Reward Hacking)\n",
+    "            if action.target == \"All\":\n",
+    "                reward -= 1.0  \n",
+    "            \n",
+    "            # Penalty for empty or trivially short drafts/finals\n",
+    "            # (short expert questions are often valid and should not be over-penalized)\n",
+    "            if action.action_type in {\"propose_draft\", \"submit_final\"} and len((action.content or \"\").split()) < 5:\n",
+    "                reward -= 0.2\n",
+    "\n",
+    "            # Penalize very long outputs; they correlate with max-length clipping.\n",
+    "            if len((action.content or \"\").split()) > 80:\n",
+    "                reward -= 0.2\n",
+    "            if len(completion_text) > 320:\n",
+    "                reward -= 0.15\n",
+    "\n",
+    "            # Encourage strict JSON-only behavior: starts with { and ends with }.\n",
+    "            is_strict_json = completion_text.startswith(\"{\") and completion_text.endswith(\"}\")\n",
+    "            if is_strict_json:\n",
+    "                reward += 0.1\n",
+    "            else:\n",
+    "                reward -= 0.3\n",
+    "\n",
+    "            # Hard penalty for non-terminated JSON-like responses.\n",
+    "            # This directly pushes generations away from max-token clipping.\n",
+    "            if not completion_text.endswith(\"}\"):\n",
+    "                reward -= 0.25\n",
+    "\n",
+    "            # Small bonus for compact single-line JSON output.\n",
+    "            if is_strict_json and \"\\n\" not in completion_text and len(completion_text) <= 240:\n",
+    "                reward += 0.08\n",
+    "\n",
+    "            # Strongly discourage invalid action/target combinations.\n",
+    "            if action.action_type == \"submit_final\" and action.target is not None:\n",
+    "                reward -= 0.6\n",
+    "            if action.action_type in {\"message_expert\", \"propose_draft\"} and action.target is None:\n",
+    "                reward -= 0.6\n",
+    "\n",
+    "            # ── 3. HEURISTIC STATE GRADING (NO API CALLS!) ──\n",
+    "            \n",
+    "            if action.action_type == \"message_expert\" and action.target != \"All\":\n",
+    "                # Did it ask a question it already knows the answer to?\n",
+    "                if f\"{action.target}: ✓ DISCOVERED\" in prompt:\n",
+    "                    reward -= 0.5\n",
+    "                else:\n",
+    "                    reward += 0.33 # Good job doing research!\n",
+    "\n",
+    "            elif action.action_type in [\"propose_draft\", \"submit_final\"]:\n",
+    "                # Did it try to submit before gathering all constraints?\n",
+    "                if \"? unknown\" in prompt:\n",
+    "                    reward -= 1.0 # Heavy penalty for guessing\n",
+    "                else:\n",
+    "                    # It did the research. Did it actually include the constraints?\n",
+    "                    text = action.content.lower()\n",
+    "                    has_finance = \"50\" in text\n",
+    "                    has_security = \"biometric\" in text\n",
+    "                    has_ux = \"click\" in text or \"tap\" in text\n",
+    "                    \n",
+    "                    if has_finance and has_security and has_ux:\n",
+    "                        reward += 1.5  \n",
+    "                    else:\n",
+    "                        reward -= 0.5\n",
+    "\n",
+    "            # ── 4. ORACLE-GUIDED DENSE SHAPING ──\n",
+    "            # This gives non-binary signal and prevents reward plateaus.\n",
+    "            if oracle_raw:\n",
+    "                oracle_action = parse_action(oracle_raw)\n",
+    "                if oracle_action is not None:\n",
+    "                    if action.action_type == oracle_action.action_type:\n",
+    "                        reward += 0.45\n",
+    "                    else:\n",
+    "                        reward -= 0.25\n",
+    "\n",
+    "                    if action.target == oracle_action.target:\n",
+    "                        reward += 0.35\n",
+    "                    else:\n",
+    "                        reward -= 0.2\n",
+    "\n",
+    "                    overlap = lexical_overlap(action.content, oracle_action.content)\n",
+    "                    reward += 0.4 * overlap\n",
+    "\n",
+    "            # Late turns should avoid endless questioning/proposals.\n",
+    "            if isinstance(turn, int):\n",
+    "                if turn >= 8 and action.action_type != \"submit_final\":\n",
+    "                    reward -= 0.35\n",
+    "                if turn >= 10 and action.action_type != \"submit_final\":\n",
+    "                    reward -= 0.6\n",
+    "                # Reward timely completion once constraints are all discovered.\n",
+    "                if (\n",
+    "                    action.action_type == \"submit_final\"\n",
+    "                    and \"? unknown\" not in prompt\n",
+    "                    and turn <= 10\n",
+    "                ):\n",
+    "                    reward += 0.6\n",
+    "\n",
+    "            # Keep rewards in a stable range for GRPO.\n",
+    "            reward = max(-2.0, min(2.0, reward))\n",
+    "\n",
+    "            rewards.append(reward)\n",
+    "\n",
+    "        return rewards\n",
+    "    return reward_fn"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "276d1887",
+   "metadata": {},
+   "source": [
+    "### GRAPH PLOTS"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "fa70239a",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def save_training_plots(log_history: list[dict], output_dir: Path):\n",
+    "    if not HAS_PLT:\n",
+    "        print(\"  matplotlib not available — skipping plots\")\n",
+    "        return\n",
+    "\n",
+    "    output_dir.mkdir(parents=True, exist_ok=True)\n",
+    "\n",
+    "    # Loss curve\n",
+    "    loss_points = [\n",
+    "        (e[\"step\"], e[\"loss\"])\n",
+    "        for e in log_history\n",
+    "        if \"loss\" in e and \"step\" in e\n",
+    "    ]\n",
+    "    if loss_points:\n",
+    "        xs, ys = zip(*loss_points)\n",
+    "        fig, ax = plt.subplots(figsize=(9, 4))\n",
+    "        ax.plot(xs, ys, marker=\"o\", linewidth=1.5, color=\"#4C72B0\", markersize=4)\n",
+    "        ax.set_xlabel(\"Training Step\", fontsize=12)\n",
+    "        ax.set_ylabel(\"GRPO Loss\", fontsize=12)\n",
+    "        ax.set_title(\n",
+    "            \"Project Polymath — GRPO Training Loss\\n\"\n",
+    "            \"(State-Based: each step = one negotiation decision)\",\n",
+    "            fontsize=12\n",
+    "        )\n",
+    "        ax.grid(True, alpha=0.3)\n",
+    "        plt.tight_layout()\n",
+    "        plt.savefig(output_dir / \"loss_curve.png\", dpi=160)\n",
+    "        plt.close()\n",
+    "        print(f\"  Saved: {output_dir}/loss_curve.png\")\n",
+    "\n",
+    "    # Reward curve (from log history if available)\n",
+    "    reward_points = [\n",
+    "        (e[\"step\"], e.get(\"reward\", e.get(\"mean_reward\", None)))\n",
+    "        for e in log_history\n",
+    "        if \"step\" in e and (\"reward\" in e or \"mean_reward\" in e)\n",
+    "    ]\n",
+    "    reward_points = [(s, r) for s, r in reward_points if r is not None]\n",
+    "\n",
+    "    if reward_points:\n",
+    "        xs, ys = zip(*reward_points)\n",
+    "        fig, ax = plt.subplots(figsize=(9, 4))\n",
+    "        ax.plot(xs, ys, marker=\"s\", linewidth=1.5, color=\"#55A868\", markersize=4)\n",
+    "        ax.set_xlabel(\"Training Step\", fontsize=12)\n",
+    "        ax.set_ylabel(\"Mean Reward\", fontsize=12)\n",
+    "        ax.set_title(\n",
+    "            \"Project Polymath — Mean Reward During GRPO Training\\n\"\n",
+    "            \"(Harmonic mean of Finance/Security/UX constraint satisfaction)\",\n",
+    "            fontsize=12\n",
+    "        )\n",
+    "        ax.grid(True, alpha=0.3)\n",
+    "        plt.tight_layout()\n",
+    "        plt.savefig(output_dir / \"reward_curve.png\", dpi=160)\n",
+    "        plt.close()\n",
+    "        print(f\"  Saved: {output_dir}/reward_curve.png\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "57034d95",
+   "metadata": {},
+   "source": [
+    "### Main functin"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "01e33d44",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def main():\n",
+    "    parser = argparse.ArgumentParser(description=\"State-Based GRPO — Project Polymath\")\n",
+    "\n",
+    "    # Model\n",
+    "    parser.add_argument(\"--model\", default=\"unsloth/Qwen2.5-3B-Instruct-bnb-4bit\",\n",
+    "                        help=\"Base model to train\")\n",
+    "    parser.add_argument(\"--use-unsloth\", action=\"store_true\",\n",
+    "                        help=\"Use Unsloth for 2x faster training (recommended on GPU)\")\n",
+    "\n",
+    "    # Dataset\n",
+    "    parser.add_argument(\"--states\", type=int, default=40,\n",
+    "                        help=\"Number of negotiation states to train on\")\n",
+    "    parser.add_argument(\"--states-per-topic\", type=int, default=5,\n",
+    "                        help=\"States to extract per topic (1-5)\")\n",
+    "    parser.add_argument(\"--topics-limit\", type=int, default=20,\n",
+    "                        help=\"Max topics to use from ai_pm_prompts.json\")\n",
+    "\n",
+    "    # GRPO hyperparams\n",
+    "    parser.add_argument(\"--group-size\", type=int, default=8,\n",
+    "                        help=\"G: completions per prompt for GRPO advantage (default: 8)\")\n",
+    "    parser.add_argument(\"--epochs\", type=float, default=3.0)\n",
+    "    parser.add_argument(\"--lr\", type=float, default=5e-6,\n",
+    "                        help=\"Learning rate (lower = safer, 5e-6 recommended for GRPO)\")\n",
+    "    parser.add_argument(\"--max-new-tokens\", type=int, default=40,\n",
+    "                        help=\"Max generated tokens per sampled completion (default: 40)\")\n",
+    "    parser.add_argument(\"--temperature\", type=float, default=0.9,\n",
+    "                        help=\"Sampling temperature for GRPO rollouts\")\n",
+    "    parser.add_argument(\"--top-p\", type=float, default=0.9,\n",
+    "                        help=\"Nucleus sampling p for GRPO rollouts\")\n",
+    "    parser.add_argument(\"--batch-size\", type=int, default=1)\n",
+    "    parser.add_argument(\"--grad-accum\", type=int, default=4)\n",
+    "    parser.add_argument(\"--max-seq-length\", type=int, default=2048)\n",
+    "\n",
+    "    # Output\n",
+    "    parser.add_argument(\"--output-dir\", default=str(OUTPUT_DIR))\n",
+    "    parser.add_argument(\"--dry-run\", action=\"store_true\",\n",
+    "                        help=\"Build dataset and verify reward fn, skip actual training\")\n",
+    "\n",
+    "    args = parser.parse_args()\n",
+    "\n",
+    "    # for dry run only\n",
+    "    # if not HAS_TRL:\n",
+    "    #     raise RuntimeError(\"pip install trl>=0.8.0 transformers datasets\")\n",
+    "\n",
+    "    output_dir = Path(args.output_dir)\n",
+    "    output_dir.mkdir(parents=True, exist_ok=True)\n",
+    "\n",
+    "    # Build dataset\n",
+    "    print(\"\\n[1/4] Loading state dataset...\")\n",
+    "    topics = load_topics(limit=args.topics_limit)\n",
+    "    dataset_path = output_dir / \"state_dataset.jsonl\"\n",
+    "\n",
+    "    # CACHING LOGIC for dataset\n",
+    "    if dataset_path.exists():\n",
+    "        print(f\"  [CACHE HIT] Found existing dataset! Loading instantly from {dataset_path}...\")\n",
+    "        records = []\n",
+    "        with dataset_path.open(\"r\") as f:\n",
+    "            for line in f:\n",
+    "                if line.strip():\n",
+    "                    records.append(json.loads(line))\n",
+    "        records = records[:args.states]\n",
+    "    else:\n",
+    "        print(\"  [CACHE MISS] No dataset found. Generating from scratch (this may take a minute)...\")\n",
+    "        records = build_state_dataset(topics, states_per_topic=args.states_per_topic)\n",
+    "        records = records[:args.states]\n",
+    "        \n",
+    "        with dataset_path.open(\"w\") as f:\n",
+    "            for r in records:\n",
+    "                f.write(json.dumps(r, ensure_ascii=True) + \"\\n\")\n",
+    "        print(f\"  Saved {len(records)} states → {dataset_path}\")\n",
+    "\n",
+    "    dataset = Dataset.from_list([{\n",
+    "        \"prompt\": r[\"prompt\"],\n",
+    "        \"topic\": r[\"topic\"],\n",
+    "        \"turn\": r[\"turn\"],\n",
+    "        \"oracle_action\": r[\"oracle_action\"],\n",
+    "        \"step_idx\": r[\"step_idx\"],\n",
+    "    } for r in records])\n",
+    "\n",
+    "    # Verify reward function\n",
+    "    print(\"\\n[2/4] Verifying reward function on 6 samples...\")\n",
+    "    reward_fn = make_reward_fn()\n",
+    "    # reward_fn = make_reward_fn(topics)\n",
+    "    prompt_with_finance_discovered = build_state_prompt(\n",
+    "    topic=topics[0],\n",
+    "    turn=1,\n",
+    "    feedback_so_far=\"Finance: Budget cap is $50k.\",\n",
+    "    discovered=\"  Finance: ✓ DISCOVERED\\n  Security: ? unknown\\n  UX: ? unknown\",\n",
+    "    )\n",
+    "\n",
+    "    test_completions = [\n",
+    "        ORACLE_ACTIONS[\"ask_finance\"],       # Should score ~0.33\n",
+    "        ORACLE_ACTIONS[\"ask_security\"],      # Should score ~0.33  \n",
+    "        '{\"action_type\": \"message_expert\", \"target\": \"Finance\", \"content\": \"What is the budget?\"}',  # Should score -0.5 (repeat)\n",
+    "        '{\"action_type\": \"message_expert\", \"target\": \"All\", \"content\": \"What do you all need?\"}',    # Should score -1.0\n",
+    "        \"this is not JSON at all\",           # Should score -0.5\n",
+    "        ORACLE_ACTIONS[\"submit_final\"],      # Should score +1.5 (all constraints in content)\n",
+    "    ]\n",
+    "\n",
+    "    test_rewards = reward_fn(\n",
+    "        completions=test_completions,\n",
+    "        prompts=[\n",
+    "            \"\",                               # oracle ask_finance — unknown state\n",
+    "            prompt_with_finance_discovered,   # ask_security — good pivot\n",
+    "            prompt_with_finance_discovered,   # ask Finance again — repeat penalty\n",
+    "            \"\",                               # broadcast — penalty\n",
+    "            \"\",                               # bad JSON\n",
+    "            build_state_prompt(               # submit after all discovered\n",
+    "                topic=topics[0], turn=4,\n",
+    "                feedback_so_far=\"All experts responded.\",\n",
+    "                discovered=\"  Finance: ✓ DISCOVERED\\n  Security: ✓ DISCOVERED\\n  UX: ✓ DISCOVERED\",\n",
+    "            ),\n",
+    "        ],\n",
+    "    )\n",
+    "\n",
+    "    labels = [\n",
+    "        \"Oracle ask_finance\",\n",
+    "        \"Oracle ask_security\",\n",
+    "        \"Repeat question to discovered expert\",\n",
+    "        \"Broadcast to All\",\n",
+    "        \"Malformed JSON\",\n",
+    "        \"Submit final with all constraints\",\n",
+    "    ]\n",
+    "    for label, reward in zip(labels, test_rewards):\n",
+    "        print(f\"  • {label}: reward={reward:.3f}\")\n",
+    "\n",
+    "    ask_finance_r, ask_security_r, repeat_r, broadcast_r, malformed_r, submit_r = test_rewards\n",
+    "    checks = [\n",
+    "        (\"oracle ask_finance is positive\", ask_finance_r > 0.0),\n",
+    "        (\"oracle ask_security is positive\", ask_security_r > 0.0),\n",
+    "        (\"repeat < oracle ask_finance\", repeat_r < ask_finance_r),\n",
+    "        (\"broadcast <= repeat\", broadcast_r <= repeat_r),\n",
+    "        (\"malformed JSON is negative\", malformed_r < 0.0),\n",
+    "        (\"submit_final > oracle ask_finance\", submit_r > ask_finance_r),\n",
+    "    ]\n",
+    "    all_ok = True\n",
+    "    for name, passed in checks:\n",
+    "        status = \"✓\" if passed else \"✗\"\n",
+    "        print(f\"  {status} invariant: {name}\")\n",
+    "        all_ok = all_ok and passed\n",
+    "    if not all_ok:\n",
+    "        raise RuntimeError(\"Reward verification invariants failed. Check reward shaping logic.\")\n",
+    "\n",
+    "    if args.dry_run:\n",
+    "        print(\"\\n[DRY RUN] Dataset and reward function verified. Skipping training.\")\n",
+    "        print(\"  Run without --dry-run on GPU to train.\")\n",
+    "        return\n",
+    "    \n",
+    "    # FOR DRY RUN ONLY\n",
+    "    if not HAS_TRL:\n",
+    "        raise RuntimeError(\"TRL is required for actual training on the GPU.\")\n",
+    "    # Load model\n",
+    "    print(f\"\\n[3/4] Loading model: {args.model}\")\n",
+    "\n",
+    "    if args.use_unsloth:\n",
+    "        try:\n",
+    "            from unsloth import FastLanguageModel\n",
+    "            HAS_UNSLOTH = True\n",
+    "        except Exception as e:\n",
+    "            raise RuntimeError(f\"Unsloth failed to load: {e}\\nRun without --use-unsloth instead.\")\n",
+    "        model, tokenizer = FastLanguageModel.from_pretrained(\n",
+    "            model_name=args.model,\n",
+    "            max_seq_length=args.max_seq_length,\n",
+    "            load_in_4bit=True,\n",
+    "            dtype=None,\n",
+    "        )\n",
+    "        model = FastLanguageModel.get_peft_model(\n",
+    "            model,\n",
+    "            r=16,\n",
+    "            lora_alpha=32,\n",
+    "            lora_dropout=0.0,\n",
+    "            target_modules=[\"q_proj\", \"k_proj\", \"v_proj\", \"o_proj\",\n",
+    "                           \"gate_proj\", \"up_proj\", \"down_proj\"],\n",
+    "            use_gradient_checkpointing=\"unsloth\",\n",
+    "        )\n",
+    "\n",
+    "        if tokenizer.pad_token is None:\n",
+    "            tokenizer.pad_token = tokenizer.eos_token\n",
+    "        im_end_id = tokenizer.convert_tokens_to_ids(\"<|im_end|>\")\n",
+    "        if isinstance(im_end_id, int) and im_end_id >= 0:\n",
+    "            tokenizer.eos_token = \"<|im_end|>\"\n",
+    "        model.config.pad_token_id = tokenizer.pad_token_id\n",
+    "        if tokenizer.eos_token_id is not None:\n",
+    "            model.config.eos_token_id = tokenizer.eos_token_id\n",
+    "            model.generation_config.eos_token_id = tokenizer.eos_token_id\n",
+    "        model.generation_config.pad_token_id = tokenizer.pad_token_id\n",
+    "        print(\"  Unsloth LoRA loaded (4-bit quantization)\")\n",
+    "    else:\n",
+    "        if not HAS_TRANSFORMERS:\n",
+    "            raise RuntimeError(\"pip install transformers\")\n",
+    "        tokenizer = AutoTokenizer.from_pretrained(args.model)\n",
+    "        if tokenizer.pad_token is None:\n",
+    "            tokenizer.pad_token = tokenizer.eos_token\n",
+    "\n",
+    "        # Qwen chat models typically terminate on <|im_end|>.\n",
+    "        im_end_id = tokenizer.convert_tokens_to_ids(\"<|im_end|>\")\n",
+    "        if isinstance(im_end_id, int) and im_end_id >= 0:\n",
+    "            tokenizer.eos_token = \"<|im_end|>\"\n",
+    "\n",
+    "        model = AutoModelForCausalLM.from_pretrained(args.model)\n",
+    "\n",
+    "        # Keepping model/generation config aligned with tokenizer.\n",
+    "        model.config.pad_token_id = tokenizer.pad_token_id\n",
+    "        if tokenizer.eos_token_id is not None:\n",
+    "            model.config.eos_token_id = tokenizer.eos_token_id\n",
+    "            model.generation_config.eos_token_id = tokenizer.eos_token_id\n",
+    "        model.generation_config.pad_token_id = tokenizer.pad_token_id\n",
+    "\n",
+    "        print(\"  Standard transformers model loaded\")\n",
+    "\n",
+    "    # GRPO Training\n",
+    "    print(f\"\\n[4/4] Starting GRPO training...\")\n",
+    "    print(f\"  States: {len(records)} | Group size (G): {args.group_size}\")\n",
+    "    print(f\"  Epochs: {args.epochs} | LR: {args.lr}\")\n",
+    "    print(f\"  Total updates: ~{int(len(records) * args.epochs / args.batch_size)}\")\n",
+    "\n",
+    "    # Build explicit stop-token list for GRPO sampling.\n",
+    "    eos_ids = []\n",
+    "    if tokenizer.eos_token_id is not None:\n",
+    "        eos_ids.append(int(tokenizer.eos_token_id))\n",
+    "    im_end_id = tokenizer.convert_tokens_to_ids(\"<|im_end|>\")\n",
+    "    if isinstance(im_end_id, int) and im_end_id >= 0:\n",
+    "        eos_ids.append(int(im_end_id))\n",
+    "    close_brace_id = tokenizer.convert_tokens_to_ids(\"}\")\n",
+    "    if isinstance(close_brace_id, int) and close_brace_id >= 0:\n",
+    "        eos_ids.append(int(close_brace_id))\n",
+    "    # Preserve order, remove duplicates.\n",
+    "    eos_ids = list(dict.fromkeys(eos_ids))\n",
+    "\n",
+    "    generation_kwargs = {\n",
+    "        \"eos_token_id\": eos_ids if len(eos_ids) > 1 else (eos_ids[0] if eos_ids else None),\n",
+    "        \"pad_token_id\": tokenizer.pad_token_id,\n",
+    "    }\n",
+    "    # Remove None values.\n",
+    "    generation_kwargs = {k: v for k, v in generation_kwargs.items() if v is not None}\n",
+    "\n",
+    "    config = GRPOConfig(\n",
+    "        output_dir=str(output_dir),\n",
+    "\n",
+    "        # GRPO-specific\n",
+    "        num_generations=args.group_size,      # G: sample this many completions per prompt\n",
+    "        max_completion_length=args.max_new_tokens,\n",
+    "        temperature=args.temperature,\n",
+    "        top_p=args.top_p,\n",
+    "        top_k=40,\n",
+    "        repetition_penalty=1.05,\n",
+    "        generation_kwargs=generation_kwargs,\n",
+    "        mask_truncated_completions=True,\n",
+    "\n",
+    "        # Standard training\n",
+    "        learning_rate=args.lr,\n",
+    "        num_train_epochs=args.epochs,\n",
+    "        per_device_train_batch_size=max(1, args.batch_size),\n",
+    "        gradient_accumulation_steps=args.grad_accum,\n",
+    "\n",
+    "        # Logging\n",
+    "        logging_steps=1,\n",
+    "        save_strategy=\"epoch\",\n",
+    "        report_to=[],\n",
+    "    )\n",
+    "\n",
+    "    trainer = GRPOTrainer(\n",
+    "            model=model,\n",
+    "            processing_class=tokenizer,\n",
+    "            args=config,\n",
+    "            reward_funcs=reward_fn,    \n",
+    "            train_dataset=dataset,\n",
+    "        )\n",
+    "\n",
+    "    trainer.train()\n",
+    "\n",
+    "    # Saving everything\n",
+    "    trainer.save_model(str(output_dir / \"final_model\"))\n",
+    "    tokenizer.save_pretrained(str(output_dir / \"final_model\"))\n",
+    "    print(f\"\\n  Model saved → {output_dir}/final_model\")\n",
+    "\n",
+    "    # Saving metrics\n",
+    "    metrics_path = output_dir / \"grpo_metrics.json\"\n",
+    "    with metrics_path.open(\"w\") as f:\n",
+    "        json.dump(trainer.state.log_history, f, indent=2)\n",
+    "    print(f\"  Metrics saved → {metrics_path}\")\n",
+    "\n",
+    "    # Saving plots\n",
+    "    save_training_plots(trainer.state.log_history, output_dir)\n",
+    "\n",
+    "    # Summary\n",
+    "    log = trainer.state.log_history\n",
+    "    losses = [e[\"loss\"] for e in log if \"loss\" in e]\n",
+    "    if losses:\n",
+    "        print(f\"\\n  Initial loss:  {losses[0]:.4f}\")\n",
+    "        print(f\"  Final loss:    {losses[-1]:.4f}\")\n",
+    "        print(f\"  Improvement:   {((losses[0] - losses[-1]) / losses[0] * 100):.1f}%\")\n",
+    "\n",
+    "    print(f\"\\n{'='*60}\")\n",
+    "    print(f\"  GRPO TRAINING COMPLETE\")\n",
+    "    print(f\"  Model:   {output_dir}/final_model\")\n",
+    "    print(f\"  Plots:   {output_dir}/loss_curve.png\")\n",
+    "    print(f\"           {output_dir}/reward_curve.png\")\n",
+    "    print(f\"  Metrics: {output_dir}/grpo_metrics.json\")\n",
+    "    print(f\"{'='*60}\")\n",
+    "\n",
+    "\n",
+    "if __name__ == \"__main__\":\n",
+    "    main()"
+   ]
+  }
+ ],
+ "metadata": {
+  "language_info": {
+   "name": "python"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}