Add Pyright configuration, enhance dataset building, and update training notebook

- Created a new Pyright configuration file for type checking.
- Improved error handling in `fetch_ohlcv` function to check for None values.
- Refactored `build_prices` function for better clarity and type safety.
- Updated `build_indicators` to ensure proper type handling for price data.
- Enhanced the training notebook with clearer markdown and structured code cells.
- Added new evaluation and summary files for training runs.
- Included binary files for portfolio and reward curves in evaluation runs.

app/council/llm.py

@@ -180,17 +180,29 @@ class TransformersLLMClient:
         cls,
         model_id: str = "google/gemma-4-E4B-it",
         load_in_4bit: bool = True,
-        device_map: str = "auto",
+        device_map: Any = "auto",
     ) -> "TransformersLLMClient":
+        import torch
         from transformers import AutoModelForCausalLM, AutoProcessor
 
-        kwargs: dict = {"dtype": "auto", "device_map": device_map}
+        # T4 (Turing) doesn't have bf16 hardware — fall back to fp16 there.
+        compute_dtype = (
+            "bfloat16"
+            if (torch.cuda.is_available() and torch.cuda.is_bf16_supported())
+            else "float16"
+        )
+
+        kwargs: dict = {"dtype": compute_dtype, "device_map": device_map}
         if load_in_4bit:
             from transformers import BitsAndBytesConfig
             kwargs["quantization_config"] = BitsAndBytesConfig(
                 load_in_4bit=True,
-                bnb_4bit_compute_dtype="bfloat16",
+                bnb_4bit_compute_dtype=compute_dtype,
                 bnb_4bit_quant_type="nf4",
+                bnb_4bit_use_double_quant=True,
+                # Allow non-quantizable modules (vision tower, embeddings) to
+                # land on CPU when GPU is tight — required on T4 for E4B.
+                llm_int8_enable_fp32_cpu_offload=True,
             )
         processor = AutoProcessor.from_pretrained(model_id)
         model = AutoModelForCausalLM.from_pretrained(model_id, **kwargs)
@@ -244,6 +256,15 @@ class TransformersLLMClient:
     # ----------------------------------------------------------------- helpers
     @staticmethod
     def _prep_messages(messages: list[dict]) -> list[dict]:
+        """Normalize OpenAI-style messages for HF AutoProcessor chat templates.
+
+        Two transforms:
+        1. Plain string `content` is wrapped as `[{"type": "text", "text": ...}]`
+           because multimodal processors iterate `content` and break on strings.
+        2. `{"type": "image_url", "image_url": {"url": "data:image/..."}}` parts
+           are decoded back to PIL.Image objects in `{"type": "image", "image": img}`
+           form expected by Gemma's processor.
+        """
         import base64
         from io import BytesIO
 
@@ -252,6 +273,10 @@ class TransformersLLMClient:
         out = []
         for m in messages:
             content = m.get("content")
+            if isinstance(content, str):
+                # Plain string -> single text part
+                out.append({**m, "content": [{"type": "text", "text": content}]})
+                continue
             if isinstance(content, list):
                 new_parts = []
                 for part in content:

app/data/indicators.py

@@ -17,7 +17,7 @@ def rsi(close: pd.Series, length: int = 14) -> pd.Series:
     avg_gain = gain.ewm(alpha=1 / length, adjust=False, min_periods=length).mean()
     avg_loss = loss.ewm(alpha=1 / length, adjust=False, min_periods=length).mean()
     rs = avg_gain / avg_loss.replace(0, float("nan"))
-    return 100 - (100 / (1 + rs))
+    return pd.Series(100 - (100 / (1 + rs)), index=close.index, name="rsi")
 
 
 def macd(
@@ -33,7 +33,9 @@ def macd(
 
 
 def sma(close: pd.Series, length: int) -> pd.Series:
-    return close.rolling(length, min_periods=length).mean()
+    return pd.Series(
+        close.rolling(length, min_periods=length).mean(), index=close.index
+    )
 
 
 def bbands(close: pd.Series, length: int = 20, std: float = 2.0) -> pd.DataFrame:
@@ -55,4 +57,7 @@ def atr(
         ],
         axis=1,
     ).max(axis=1)
-    return tr.ewm(alpha=1 / length, adjust=False, min_periods=length).mean()
+    return pd.Series(
+        tr.ewm(alpha=1 / length, adjust=False, min_periods=length).mean(),
+        index=close.index,
+    )

app/data/loader.py

@@ -57,7 +57,9 @@ def macro() -> pd.DataFrame:
 def episode_rows(task_id: str) -> pd.DataFrame:
     """Episode prices for a task, sorted by date."""
     df = prices()
-    return df[(df["task_id"] == task_id) & df["in_episode"]].sort_values("date").reset_index(drop=True)
+    mask = (df["task_id"] == task_id) & df["in_episode"]
+    sub = df.loc[mask]
+    return sub.sort_values(by="date").reset_index(drop=True)
 
 
 def lookup_indicators(ticker: str, date: str) -> dict:

pyrightconfig.json

@@ -0,0 +1,9 @@
+{
+  "include": ["app", "scripts", "server", "tests", "training", "inference.py", "client.py"],
+  "exclude": ["**/__pycache__", "**/.venv", "**/.cache", "training/runs"],
+  "extraPaths": ["."],
+  "typeCheckingMode": "basic",
+  "reportMissingImports": "warning",
+  "reportMissingTypeStubs": "none",
+  "pythonVersion": "3.12"
+}

scripts/build_dataset.py

@@ -82,12 +82,14 @@ def fetch_ohlcv(ticker: str, start: str, end: str) -> pd.DataFrame:
     df = yf.download(
         ticker, start=start, end=end, auto_adjust=False, progress=False
     )
-    if df.empty:
+    if df is None or df.empty:
         raise RuntimeError(f"yfinance returned empty for {ticker} {start}..{end}")
     if isinstance(df.columns, pd.MultiIndex):
         df.columns = df.columns.get_level_values(0)
-    df.index = pd.to_datetime(df.index).tz_localize(None).normalize()
-    df = df[["Open", "High", "Low", "Close", "Volume"]]
+    idx = pd.DatetimeIndex(pd.to_datetime(df.index)).tz_localize(None)
+    idx = idx.normalize()  # type: ignore[attr-defined]
+    df.index = idx
+    df = df.loc[:, ["Open", "High", "Low", "Close", "Volume"]]
     df.columns = ["open", "high", "low", "close", "volume"]
     return df
 
@@ -97,38 +99,39 @@ def lookback_start(episode_start: str, days: int) -> str:
 
 
 def build_prices() -> pd.DataFrame:
-    rows = []
+    rows: list[dict] = []
     for task in TASKS:
         start = lookback_start(task["episode_start"], LOOKBACK_DAYS)
         end = (pd.to_datetime(task["episode_end"]) + pd.Timedelta(days=1)).strftime("%Y-%m-%d")
         df = fetch_ohlcv(task["ticker"], start, end)
+        episode_start = pd.to_datetime(task["episode_start"])
+        episode_end = pd.to_datetime(task["episode_end"])
         for date, row in df.iterrows():
+            ts = pd.Timestamp(date)  # type: ignore[arg-type]
             rows.append({
                 "task_id": task["task_id"],
                 "ticker": task["ticker"],
-                "date": date.strftime("%Y-%m-%d"),
-                "open": float(row["open"]),
-                "high": float(row["high"]),
-                "low": float(row["low"]),
-                "close": float(row["close"]),
-                "volume": float(row["volume"]),
-                "in_episode": (
-                    pd.to_datetime(task["episode_start"])
-                    <= date
-                    <= pd.to_datetime(task["episode_end"])
-                ),
+                "date": ts.strftime("%Y-%m-%d"),
+                "open":   float(row["open"]),    # type: ignore[arg-type]
+                "high":   float(row["high"]),    # type: ignore[arg-type]
+                "low":    float(row["low"]),     # type: ignore[arg-type]
+                "close":  float(row["close"]),   # type: ignore[arg-type]
+                "volume": float(row["volume"]),  # type: ignore[arg-type]
+                "in_episode": episode_start <= ts <= episode_end,
             })
     return pd.DataFrame(rows)
 
 
 def build_indicators(prices: pd.DataFrame) -> pd.DataFrame:
     """Compute indicators per ticker on the full lookback+episode window."""
-    out_rows = []
+    out_rows: list[pd.DataFrame] = []
     for ticker, group in prices.groupby("ticker"):
-        df = group.sort_values("date").set_index(pd.to_datetime(group["date"]))
-        close = df["close"]
-        high = df["high"]
-        low = df["low"]
+        df = group.sort_values("date").set_index(
+            pd.to_datetime(group["date"])
+        )
+        close = pd.Series(df["close"].astype(float))
+        high = pd.Series(df["high"].astype(float))
+        low = pd.Series(df["low"].astype(float))
 
         ind = pd.DataFrame(index=df.index)
         ind["rsi14"] = ta.rsi(close, length=14)
@@ -147,10 +150,10 @@ def build_indicators(prices: pd.DataFrame) -> pd.DataFrame:
         ind["ticker"] = ticker
         ind["date"] = ind["date"].dt.strftime("%Y-%m-%d")
         out_rows.append(ind)
-    out = pd.concat(out_rows, ignore_index=True)
+    out: pd.DataFrame = pd.concat(out_rows, ignore_index=True)
     cols = ["ticker", "date", "rsi14", "macd", "macd_signal",
             "sma20", "sma50", "sma200", "bb_lower", "bb_upper", "atr14"]
-    return out[cols]
+    return out.loc[:, cols]
 
 
 def build_peers() -> pd.DataFrame:
@@ -232,8 +235,9 @@ def render_charts(prices: pd.DataFrame) -> int:
     for ticker, group in prices.groupby("ticker"):
         group = group.sort_values("date").reset_index(drop=True)
         group["date_dt"] = pd.to_datetime(group["date"])
-        for i, row in group.iterrows():
-            if not row["in_episode"]:
+        for raw_i, row in group.iterrows():
+            i = int(raw_i)  # type: ignore[arg-type]
+            if not bool(row["in_episode"]):
                 continue
             window_start_idx = max(0, i - 60)
             window = group.iloc[window_start_idx : i + 1].copy()

training/runs/eval_smoke/summary.csv

@@ -0,0 +1,4 @@
+task_id,total_reward,final_portfolio,buy_and_hold,alpha_pct
+task_easy,0.1,10000.0,8755.6,14.21
+task_medium,0.1,10000.0,9007.25,11.02
+task_hard,0.1,10000.0,5668.0,76.43

training/train_grpo.ipynb

@@ -4,23 +4,82 @@
    "cell_type": "markdown",
    "id": "intro",
    "metadata": {},
-   "source": "# Stocker — End-to-end Colab workflow\n\nSelf-contained: clones the repo, installs deps, builds the dataset, loads\n**`google/gemma-4-E4B-it`** in-process via `transformers` (4-bit BnB),\npre-caches all 7 specialist votes, runs **GRPO** on the moderator LoRA via\nTRL, and saves loss / reward plots + the trained adapter.\n\nDesigned for a free **Colab T4** (16 GB VRAM) — no separate vLLM server\nneeded. If you have an L4/A100, drop `load_in_4bit=False` for full bf16.\n\nOutputs live under `training/runs/<timestamp>/`:\n- `moderator-lora/` — trained PEFT adapter\n- `loss.png`, `reward.png` — training curves\n- `eval_pre.json`, `eval_post.json` — pre/post backtest reports\n\n> Tip: `Runtime → Change runtime type → T4 GPU` before running."
+   "source": [
+    "# Stocker — End-to-end Colab workflow\n",
+    "\n",
+    "Self-contained: clones the repo, installs deps, builds the dataset, loads\n",
+    "**`google/gemma-4-E4B-it`** in-process via `transformers` (4-bit BnB),\n",
+    "pre-caches all 7 specialist votes, runs **GRPO** on the moderator LoRA via\n",
+    "TRL, and saves loss / reward plots + the trained adapter.\n",
+    "\n",
+    "Designed for a free **Colab T4** (16 GB VRAM) — no separate vLLM server\n",
+    "needed. If you have an L4/A100, drop `load_in_4bit=False` for full bf16.\n",
+    "\n",
+    "Outputs live under `training/runs/<timestamp>/`:\n",
+    "- `moderator-lora/` — trained PEFT adapter\n",
+    "- `loss.png`, `reward.png` — training curves\n",
+    "- `eval_pre.json`, `eval_post.json` — pre/post backtest reports\n",
+    "\n",
+    "> Tip: `Runtime → Change runtime type → T4 GPU` before running."
+   ]
   },
   {
    "cell_type": "code",
-   "execution_count": null,
+   "execution_count": 7,
    "id": "install",
    "metadata": {},
    "outputs": [],
-   "source": "# Install — keep Colab's pre-baked pandas/numpy (TF + cudf depend on them);\n# only upgrade what we strictly need.\n#\n# Big upgrades (transformers/trl/peft) are intentional — Colab ships old.\n# Everything else uses --upgrade-strategy only-if-needed so we don't\n# break google-colab/tensorflow/gradio/etc.\n!pip install -q -U 'transformers>=4.55' 'trl>=0.11' 'peft>=0.13' 'accelerate>=1.0' 'bitsandbytes>=0.43' 'datasets>=3.0'\n!pip install -q --upgrade-strategy=only-if-needed yfinance mplfinance pyarrow 'pydantic>=2' pydantic-settings 'openai>=1' tensorboard 'huggingface_hub>=1.10'\n# Indicators are computed in-repo (app/data/indicators.py) — no pandas-ta."
+   "source": [
+    "# Install — keep Colab's pre-baked pandas/numpy (TF + cudf depend on them);\n",
+    "# only upgrade what we strictly need.\n",
+    "#\n",
+    "# Big upgrades (transformers/trl/peft) are intentional — Colab ships old.\n",
+    "# Everything else uses --upgrade-strategy only-if-needed so we don't\n",
+    "# break google-colab/tensorflow/gradio/etc.\n",
+    "!pip install -q -U 'transformers>=4.55' 'trl>=0.11' 'peft>=0.13' 'accelerate>=1.0' 'bitsandbytes>=0.43' 'datasets>=3.0'\n",
+    "!pip install -q --upgrade-strategy=only-if-needed yfinance mplfinance pyarrow 'pydantic>=2' pydantic-settings 'openai>=1' tensorboard 'huggingface_hub>=1.10'\n",
+    "# Indicators are computed in-repo (app/data/indicators.py) — no pandas-ta."
+   ]
   },
   {
    "cell_type": "code",
-   "execution_count": null,
+   "execution_count": 8,
    "id": "clone",
    "metadata": {},
-   "outputs": [],
-   "source": "import os, sys, pathlib\n\n# 1) Already in a stocker repo? (running from VSCode on a local clone, or\n#    the repo was manually uploaded to Colab). Detect and skip the clone.\ndef _is_stocker_root(p: str) -> bool:\n    return os.path.isfile(os.path.join(p, \"app\", \"council\", \"specialists.py\"))\n\nCANDIDATES = [os.getcwd(), \"/content/stocker\", \"/workspace/stocker\"]\nWORKDIR = next((c for c in CANDIDATES if _is_stocker_root(c)), None)\n\nif WORKDIR is None:\n    # 2) Not present — clone. EDIT THIS URL before running on a fresh Colab.\n    REPO_URL = \"https://github.com/<your-username>/stocker.git\"\n    WORKDIR  = \"/content/stocker\"\n    assert \"<your-username>\" not in REPO_URL, (\n        \"Edit REPO_URL in this cell to your fork before running on a fresh runtime.\"\n    )\n    !git clone {REPO_URL} {WORKDIR}\n    assert _is_stocker_root(WORKDIR), f\"Clone failed — {WORKDIR}/app/council/ missing.\"\n\nos.chdir(WORKDIR)\nsys.path.insert(0, WORKDIR)\nprint(\"Working dir:\", WORKDIR)"
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Working dir: /content/stocker\n"
+     ]
+    }
+   ],
+   "source": [
+    "import os, sys, pathlib\n",
+    "\n",
+    "# 1) Already in a stocker repo? (running from VSCode on a local clone, or\n",
+    "#    the repo was manually uploaded to Colab). Detect and skip the clone.\n",
+    "def _is_stocker_root(p: str) -> bool:\n",
+    "    return os.path.isfile(os.path.join(p, \"app\", \"council\", \"specialists.py\"))\n",
+    "\n",
+    "CANDIDATES = [os.getcwd(), \"/content/stocker\", \"/workspace/stocker\"]\n",
+    "WORKDIR = next((c for c in CANDIDATES if _is_stocker_root(c)), None)\n",
+    "\n",
+    "if WORKDIR is None:\n",
+    "    # 2) Not present — clone. EDIT THIS URL before running on a fresh Colab.\n",
+    "    REPO_URL = \"https://github.com/<your-username>/stocker.git\"\n",
+    "    WORKDIR  = \"/content/stocker\"\n",
+    "    assert \"<your-username>\" not in REPO_URL, (\n",
+    "        \"Edit REPO_URL in this cell to your fork before running on a fresh runtime.\"\n",
+    "    )\n",
+    "    !git clone {REPO_URL} {WORKDIR}\n",
+    "    assert _is_stocker_root(WORKDIR), f\"Clone failed — {WORKDIR}/app/council/ missing.\"\n",
+    "\n",
+    "os.chdir(WORKDIR)\n",
+    "sys.path.insert(0, WORKDIR)\n",
+    "print(\"Working dir:\", WORKDIR)"
+   ]
   },
   {
    "cell_type": "code",
@@ -28,39 +87,253 @@
    "id": "auth",
    "metadata": {},
    "outputs": [],
-   "source": "# Gemma 4 is Apache-2.0 (not gated) — token just lifts download rate limits.\nfrom huggingface_hub import login\nimport getpass\nlogin(token=getpass.getpass(\"HF token (read scope is enough): \"))"
+   "source": [
+    "# Gemma 4 is Apache-2.0 (not gated) — token just lifts download rate limits.\n",
+    "from huggingface_hub import login\n",
+    "import getpass\n",
+    "login(token=getpass.getpass(\"HF token (read scope is enough): \"))"
+   ]
   },
   {
    "cell_type": "code",
    "execution_count": null,
    "id": "build-data",
    "metadata": {},
-   "outputs": [],
-   "source": "# Build the bundled dataset (yfinance + indicators + chart PNGs).\n# Idempotent — skip if already done.\n!python scripts/build_dataset.py\n!python scripts/validate_tasks.py"
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "[1/6] Fetching OHLCV ...\n",
+      "  prices: 536 rows\n",
+      "[2/6] Computing indicators ...\n",
+      "  indicators: 536 rows\n",
+      "[3/6] Fetching peers + commodity ...\n",
+      "  peers: 374 rows\n",
+      "[4/6] Loading curated news/forums/macro ...\n",
+      "  news: 25 headlines, forums: 17 posts, macro: 12 events\n",
+      "[5/6] Rendering candlestick charts ...\n",
+      "  charts: 126 PNGs in data/charts\n",
+      "[6/6] Done.\n",
+      "task_easy      ticker=AAPL   steps= 43 chart_ok=yes\n",
+      "task_medium    ticker=INTC   steps= 41 chart_ok=yes\n",
+      "task_hard      ticker=META   steps= 42 chart_ok=yes\n",
+      "\n",
+      "All 3 tasks OK.\n"
+     ]
+    }
+   ],
+   "source": [
+    "# Build the bundled dataset (yfinance + indicators + chart PNGs).\n",
+    "# Idempotent — skip if already done.\n",
+    "!python scripts/build_dataset.py\n",
+    "!python scripts/validate_tasks.py"
+   ]
   },
   {
    "cell_type": "code",
+   "execution_count": null,
    "id": "f55a1108",
-   "source": "from app.council.llm import TransformersLLMClient\n\n# 4-bit BnB by default; drop load_in_4bit=False on L4/A100.\nclient = TransformersLLMClient.from_pretrained(\n    \"google/gemma-4-E4B-it\",\n    load_in_4bit=True,\n)\nprint(\"Model loaded on:\", client.model.device)",
    "metadata": {},
-   "execution_count": null,
-   "outputs": []
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "/usr/local/lib/python3.12/dist-packages/huggingface_hub/utils/_auth.py:103: UserWarning: \n",
+      "Error while fetching `HF_TOKEN` secret value from your vault: 'Requesting secret HF_TOKEN timed out. Secrets can only be fetched when running from the Colab UI.'.\n",
+      "You are not authenticated with the Hugging Face Hub in this notebook.\n",
+      "If the error persists, please let us know by opening an issue on GitHub (https://github.com/huggingface/huggingface_hub/issues/new).\n",
+      "  warnings.warn(\n"
+     ]
+    },
+    {
+     "data": {
+      "application/vnd.jupyter.widget-view+json": {
+       "model_id": "09427ffe6e7b4be094cef43d13c33fe6",
+       "version_major": 2,
+       "version_minor": 0
+      },
+      "text/plain": [
+       "processor_config.json: 0.00B [00:00, ?B/s]"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    },
+    {
+     "data": {
+      "application/vnd.jupyter.widget-view+json": {
+       "model_id": "8874cb7e313545a2b79bad9317858169",
+       "version_major": 2,
+       "version_minor": 0
+      },
+      "text/plain": [
+       "chat_template.jinja: 0.00B [00:00, ?B/s]"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    },
+    {
+     "data": {
+      "application/vnd.jupyter.widget-view+json": {
+       "model_id": "37bb28cb29c64f97b2699f780174d9ee",
+       "version_major": 2,
+       "version_minor": 0
+      },
+      "text/plain": [
+       "config.json: 0.00B [00:00, ?B/s]"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    },
+    {
+     "data": {
+      "application/vnd.jupyter.widget-view+json": {
+       "model_id": "459c2365fa5e4cf1ab6897f7a5e9816f",
+       "version_major": 2,
+       "version_minor": 0
+      },
+      "text/plain": [
+       "tokenizer_config.json: 0.00B [00:00, ?B/s]"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    },
+    {
+     "data": {
+      "application/vnd.jupyter.widget-view+json": {
+       "model_id": "0f6eab584b334df89d7aa973094c3e58",
+       "version_major": 2,
+       "version_minor": 0
+      },
+      "text/plain": [
+       "tokenizer.json:   0%|          | 0.00/32.2M [00:00<?, ?B/s]"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    },
+    {
+     "data": {
+      "application/vnd.jupyter.widget-view+json": {
+       "model_id": "7c1bbf3af57b4782b0ea600e652b680a",
+       "version_major": 2,
+       "version_minor": 0
+      },
+      "text/plain": [
+       "model.safetensors:   0%|          | 0.00/16.0G [00:00<?, ?B/s]"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    },
+    {
+     "data": {
+      "application/vnd.jupyter.widget-view+json": {
+       "model_id": "f5cfa2f28b1844cd9808cf55cfdfc9db",
+       "version_major": 2,
+       "version_minor": 0
+      },
+      "text/plain": [
+       "Loading weights:   0%|          | 0/2076 [00:00<?, ?it/s]"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    },
+    {
+     "data": {
+      "application/vnd.jupyter.widget-view+json": {
+       "model_id": "925f4bb541bf49a5b42067ba6e4c10a9",
+       "version_major": 2,
+       "version_minor": 0
+      },
+      "text/plain": [
+       "generation_config.json:   0%|          | 0.00/208 [00:00<?, ?B/s]"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    },
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Model loaded on: cuda:0\n"
+     ]
+    }
+   ],
+   "source": [
+    "from app.council.llm import TransformersLLMClient\n",
+    "\n",
+    "# 4-bit BnB by default; drop load_in_4bit=False on L4/A100.\n",
+    "client = TransformersLLMClient.from_pretrained(\n",
+    "    \"google/gemma-4-E4B-it\",\n",
+    "    load_in_4bit=True,\n",
+    ")\n",
+    "print(\"Model loaded on:\", client.model.device)"
+   ]
   },
   {
    "cell_type": "code",
+   "execution_count": null,
    "id": "a3a9bd4c",
-   "source": "# Pre-cache the 7 specialists' votes for every (task, step). Specialists\n# are FROZEN, so this runs once and is reused across every GRPO step.\nfrom app.council.runner import Council\nfrom app.core.environment import StockerEnv\nfrom app.core.tasks import list_task_ids\n\ncouncil = Council(client=client, use_cache=True)\ntotal = 0\nfor task_id in list_task_ids():\n    env = StockerEnv(task_id=task_id)\n    obs = env.reset().observation\n    while True:\n        for sp in council.specialists:\n            council._cached_vote(sp, obs)   # writes .cache/council/<role>/...\n            total += 1\n        result = env.step({\"side\": \"hold\", \"quantity\": 0})\n        if result.done:\n            break\n        obs = result.observation\nprint(f\"cached {total} specialist votes across {len(list_task_ids())} tasks\")",
    "metadata": {},
-   "execution_count": null,
-   "outputs": []
+   "outputs": [
+    {
+     "ename": "TypeError",
+     "evalue": "string indices must be integers, not 'str'",
+     "output_type": "error",
+     "traceback": [
+      "\u001b[0;31m---------------------------------------------------------------------------\u001b[0m",
+      "\u001b[0;31mTypeError\u001b[0m                                 Traceback (most recent call last)",
+      "\u001b[0;32m/tmp/ipykernel_1481/2753751009.py\u001b[0m in \u001b[0;36m<cell line: 0>\u001b[0;34m()\u001b[0m\n\u001b[1;32m     12\u001b[0m     \u001b[0;32mwhile\u001b[0m \u001b[0;32mTrue\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     13\u001b[0m         \u001b[0;32mfor\u001b[0m \u001b[0msp\u001b[0m \u001b[0;32min\u001b[0m \u001b[0mcouncil\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mspecialists\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m---> 14\u001b[0;31m             \u001b[0mcouncil\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0m_cached_vote\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0msp\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mobs\u001b[0m\u001b[0;34m)\u001b[0m   \u001b[0;31m# writes .cache/council/<role>/...\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m     15\u001b[0m             \u001b[0mtotal\u001b[0m \u001b[0;34m+=\u001b[0m \u001b[0;36m1\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     16\u001b[0m         \u001b[0mresult\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0menv\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mstep\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m{\u001b[0m\u001b[0;34m\"side\"\u001b[0m\u001b[0;34m:\u001b[0m \u001b[0;34m\"hold\"\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m\"quantity\"\u001b[0m\u001b[0;34m:\u001b[0m \u001b[0;36m0\u001b[0m\u001b[0;34m}\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
+      "\u001b[0;32m/content/stocker/app/council/runner.py\u001b[0m in \u001b[0;36m_cached_vote\u001b[0;34m(self, sp, obs)\u001b[0m\n\u001b[1;32m     73\u001b[0m                 \u001b[0;32mpass\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     74\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m---> 75\u001b[0;31m         \u001b[0mvote\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0msp\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mvote\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mobs\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m     76\u001b[0m         \u001b[0mkey\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mparent\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mmkdir\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mparents\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0;32mTrue\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mexist_ok\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0;32mTrue\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     77\u001b[0m         \u001b[0mkey\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mwrite_text\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mvote\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mmodel_dump_json\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
+      "\u001b[0;32m/content/stocker/app/council/specialists.py\u001b[0m in \u001b[0;36mvote\u001b[0;34m(self, obs)\u001b[0m\n\u001b[1;32m     30\u001b[0m     \u001b[0;32mdef\u001b[0m \u001b[0mvote\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mself\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mobs\u001b[0m\u001b[0;34m:\u001b[0m \u001b[0mMarketObservation\u001b[0m\u001b[0;34m)\u001b[0m \u001b[0;34m->\u001b[0m \u001b[0mSpecialistVote\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     31\u001b[0m         \u001b[0mmessages\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mprepare_messages\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mobs\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m---> 32\u001b[0;31m         \u001b[0mtext\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mclient\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mcomplete\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mmessages\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mmax_tokens\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0;36m256\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mtemperature\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0;36m0.2\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m     33\u001b[0m         \u001b[0mdata\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mparse_json_object\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mtext\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     34\u001b[0m         return SpecialistVote(\n",
+      "\u001b[0;32m/content/stocker/app/council/llm.py\u001b[0m in \u001b[0;36mcomplete\u001b[0;34m(self, messages, max_tokens, temperature, extra_body)\u001b[0m\n\u001b[1;32m    209\u001b[0m         \u001b[0;31m# the processor expects.\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m    210\u001b[0m         \u001b[0mprepped\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0m_prep_messages\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mmessages\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m--> 211\u001b[0;31m         inputs = self.processor.apply_chat_template(\n\u001b[0m\u001b[1;32m    212\u001b[0m             \u001b[0mprepped\u001b[0m\u001b[0;34m,\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m    213\u001b[0m             \u001b[0madd_generation_prompt\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0;32mTrue\u001b[0m\u001b[0;34m,\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
+      "\u001b[0;32m/usr/local/lib/python3.12/dist-packages/transformers/processing_utils.py\u001b[0m in \u001b[0;36mapply_chat_template\u001b[0;34m(self, conversation, chat_template, tools, documents, add_generation_prompt, continue_final_message, return_assistant_tokens_mask, tokenize, return_tensors, return_dict, load_audio_from_video, processor_kwargs, **kwargs)\u001b[0m\n\u001b[1;32m   1815\u001b[0m                     \u001b[0mcontent\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mmessage\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mget\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m\"content\"\u001b[0m\u001b[0;34m)\u001b[0m \u001b[0;32mor\u001b[0m \u001b[0;34m[\u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m   1816\u001b[0m                     visuals = [\n\u001b[0;32m-> 1817\u001b[0;31m                         \u001b[0mcontent_block\u001b[0m \u001b[0;32mfor\u001b[0m \u001b[0mcontent_block\u001b[0m \u001b[0;32min\u001b[0m \u001b[0mcontent\u001b[0m \u001b[0;32mif\u001b[0m \u001b[0mcontent_block\u001b[0m\u001b[0;34m[\u001b[0m\u001b[0;34m\"type\"\u001b[0m\u001b[0;34m]\u001b[0m \u001b[0;32min\u001b[0m \u001b[0;34m[\u001b[0m\u001b[0;34m\"image\"\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m\"video\"\u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m   1818\u001b[0m                     ]\n\u001b[1;32m   1819\u001b[0m                     audio_fnames = [\n",
+      "\u001b[0;31mTypeError\u001b[0m: string indices must be integers, not 'str'"
+     ]
+    }
+   ],
+   "source": [
+    "# Pre-cache the 7 specialists' votes for every (task, step). Specialists\n",
+    "# are FROZEN, so this runs once and is reused across every GRPO step.\n",
+    "from app.council.runner import Council\n",
+    "from app.core.environment import StockerEnv\n",
+    "from app.core.tasks import list_task_ids\n",
+    "\n",
+    "council = Council(client=client, use_cache=True)\n",
+    "total = 0\n",
+    "for task_id in list_task_ids():\n",
+    "    env = StockerEnv(task_id=task_id)\n",
+    "    obs = env.reset().observation\n",
+    "    while True:\n",
+    "        for sp in council.specialists:\n",
+    "            council._cached_vote(sp, obs)   # writes .cache/council/<role>/...\n",
+    "            total += 1\n",
+    "        result = env.step({\"side\": \"hold\", \"quantity\": 0})\n",
+    "        if result.done:\n",
+    "            break\n",
+    "        obs = result.observation\n",
+    "print(f\"cached {total} specialist votes across {len(list_task_ids())} tasks\")"
+   ]
   },
   {
    "cell_type": "code",
+   "execution_count": null,
    "id": "6af95f46",
-   "source": "# Pre-training baseline rollout (specialists from cache, moderator = base Gemma)\n!python -m training.eval_rollout --out training/runs/eval_pre\n!cat training/runs/eval_pre/summary.csv",
    "metadata": {},
-   "execution_count": null,
-   "outputs": []
+   "outputs": [],
+   "source": [
+    "# Pre-training baseline rollout (specialists from cache, moderator = base Gemma)\n",
+    "!python -m training.eval_rollout --out training/runs/eval_pre\n",
+    "!cat training/runs/eval_pre/summary.csv"
+   ]
   },
   {
    "cell_type": "code",
@@ -68,7 +341,20 @@
    "id": "train",
    "metadata": {},
    "outputs": [],
-   "source": "# GRPO on the moderator LoRA. Specialist votes are read from cache —\n# only the moderator is re-rolled per training step, so this is fast.\n#\n# Knobs: --num-generations is K candidates per prompt (GRPO group size).\n# On a free T4 keep batch_size=1, grad_accum=8.\n!python -m training.train_grpo \\\n    --epochs 2 \\\n    --num-generations 8 \\\n    --batch-size 1 \\\n    --grad-accum 8 \\\n    --lora-rank 16 \\\n    --lr 5e-6"
+   "source": [
+    "# GRPO on the moderator LoRA. Specialist votes are read from cache —\n",
+    "# only the moderator is re-rolled per training step, so this is fast.\n",
+    "#\n",
+    "# Knobs: --num-generations is K candidates per prompt (GRPO group size).\n",
+    "# On a free T4 keep batch_size=1, grad_accum=8.\n",
+    "!python -m training.train_grpo \\\n",
+    "    --epochs 2 \\\n",
+    "    --num-generations 8 \\\n",
+    "    --batch-size 1 \\\n",
+    "    --grad-accum 8 \\\n",
+    "    --lora-rank 16 \\\n",
+    "    --lr 5e-6"
+   ]
   },
   {
    "cell_type": "code",
@@ -76,27 +362,63 @@
    "id": "eval",
    "metadata": {},
    "outputs": [],
-   "source": "# Post-training eval — load the trained adapter into the same client.\nimport glob, os\nRUN_DIR = sorted(glob.glob(\"training/runs/grpo_*\"))[-1]\nLORA_DIR = os.path.join(RUN_DIR, \"moderator-lora\")\nprint(\"Using LoRA:\", LORA_DIR)\n\n# Attach adapter and re-run eval. The adapter name \"moderator\" matches what\n# Moderator.decide() requests via extra_body={'lora_request': {'name': ...}}\nfrom peft import PeftModel\nclient.model = PeftModel.from_pretrained(client.model, LORA_DIR, adapter_name=\"moderator\")\nclient.moderator_lora = \"moderator\"\n\n!python -m training.eval_rollout --moderator-lora moderator --out training/runs/eval_post\n!cat training/runs/eval_post/summary.csv"
+   "source": [
+    "# Post-training eval — load the trained adapter into the same client.\n",
+    "import glob, os\n",
+    "RUN_DIR = sorted(glob.glob(\"training/runs/grpo_*\"))[-1]\n",
+    "LORA_DIR = os.path.join(RUN_DIR, \"moderator-lora\")\n",
+    "print(\"Using LoRA:\", LORA_DIR)\n",
+    "\n",
+    "# Attach adapter and re-run eval. The adapter name \"moderator\" matches what\n",
+    "# Moderator.decide() requests via extra_body={'lora_request': {'name': ...}}\n",
+    "from peft import PeftModel\n",
+    "client.model = PeftModel.from_pretrained(client.model, LORA_DIR, adapter_name=\"moderator\")\n",
+    "client.moderator_lora = \"moderator\"\n",
+    "\n",
+    "!python -m training.eval_rollout --moderator-lora moderator --out training/runs/eval_post\n",
+    "!cat training/runs/eval_post/summary.csv"
+   ]
   },
   {
    "cell_type": "code",
+   "execution_count": null,
    "id": "b1d888ed",
-   "source": "# Show training + eval curves inline\nfrom IPython.display import Image, display\nimport os\nfor p in [f\"{RUN_DIR}/loss.png\", f\"{RUN_DIR}/reward.png\",\n          \"training/runs/eval_pre/reward_curve.png\",\n          \"training/runs/eval_post/reward_curve.png\",\n          \"training/runs/eval_pre/portfolio_curve.png\",\n          \"training/runs/eval_post/portfolio_curve.png\"]:\n    if os.path.exists(p):\n        print(p)\n        display(Image(p))",
    "metadata": {},
-   "execution_count": null,
-   "outputs": []
+   "outputs": [],
+   "source": [
+    "# Show training + eval curves inline\n",
+    "from IPython.display import Image, display\n",
+    "import os\n",
+    "for p in [f\"{RUN_DIR}/loss.png\", f\"{RUN_DIR}/reward.png\",\n",
+    "          \"training/runs/eval_pre/reward_curve.png\",\n",
+    "          \"training/runs/eval_post/reward_curve.png\",\n",
+    "          \"training/runs/eval_pre/portfolio_curve.png\",\n",
+    "          \"training/runs/eval_post/portfolio_curve.png\"]:\n",
+    "    if os.path.exists(p):\n",
+    "        print(p)\n",
+    "        display(Image(p))"
+   ]
   }
  ],
  "metadata": {
   "kernelspec": {
-   "display_name": "Python 3",
+   "display_name": "Python 3 (ipykernel)",
    "language": "python",
    "name": "python3"
   },
   "language_info": {
-   "name": "python"
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.12.13"
   }
  },
  "nbformat": 4,
  "nbformat_minor": 5
-}
+}