CI workflows, branch policy alignment, and factor lab bugfixes

.github/workflows/develop.yml

@@ -0,0 +1,32 @@
+name: Develop CI
+
+on:
+  push:
+    branches: [develop]
+  pull_request:
+    branches: [develop]
+
+jobs:
+  fast-tests:
+    name: Lint & Fast Tests
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+          cache: "pip"
+
+      - name: Install dependencies
+        run: |
+          python -m pip install --upgrade pip
+          pip install pandas numpy scipy pyarrow PyYAML omegaconf antlr4-python3-runtime datasets pytest ruff
+
+      - name: Run linter
+        run: ruff check .
+
+      - name: Run fast-tier tests
+        run: PYTHONPATH=. pytest tests/

.github/workflows/main.yml

@@ -0,0 +1,32 @@
+name: Main CI
+
+on:
+  push:
+    branches: [main]
+  pull_request:
+    branches: [main]
+
+jobs:
+  full-tests:
+    name: Lint & Full Test Suite
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+          cache: "pip"
+
+      - name: Install dependencies
+        run: |
+          python -m pip install --upgrade pip
+          pip install pandas numpy scipy pyarrow PyYAML omegaconf antlr4-python3-runtime datasets pytest ruff
+
+      - name: Run linter
+        run: ruff check .
+
+      - name: Run full test suite
+        run: PYTHONPATH=. pytest tests/

config/training_vast_qlora_grpo.yaml

@@ -4,9 +4,9 @@ model:
   adapter: qlora
 
 grpo:
-  script: factor_lab.training.train_grpo_qlora
+  script: src.training.train_grpo_qlora
   objective: executable_factor_reward
-  reward_source: factor_lab.training.grpo_reward.make_grpo_reward_func
+  reward_source: src.training.grpo_reward.make_grpo_reward_func
   reward_components:
     - DSL validity
     - daily cross-sectional RankIC
@@ -23,6 +23,6 @@ grpo:
   loss_type: dapo
 
 data:
-  local_crypto_panel: ClassProject/data/crypto_panel_clean.pkl
+  local_crypto_panel: data/crypto/crypto_panel_clean.pkl
   wrds_crsp_target: daily_cross_sectional_rankic
   taq_target: intraday_microstructure_direction

config/verl_qwen25_7b_grpo.yaml

@@ -26,7 +26,7 @@ model:
 data:
   train_parquet: factor_lab/outputs/verl/crypto_grpo_tasks.parquet
   crypto_panel: data/crypto_panel_clean.pkl
-  tickers: BTC-USD,ETH-USD,XRP-USD
+  tickers: ADA-USD,BNB-USD,BTC-USD,DOGE-USD,ETH-USD,LINK-USD,XLM-USD,XRP-USD
 
 reward:
   function: factor_lab.verl.reward_function.reward_fn

config/verl_qwen3_14b_fullft_a100.yaml

@@ -17,7 +17,7 @@ data:
   train_parquet: outputs/verl/crypto_grpo_tasks.parquet
   val_parquet: outputs/verl/crypto_grpo_val_tasks.parquet
   crypto_panel: data/crypto/crypto_panel_clean.pkl
-  tickers: BTC-USD,ETH-USD,XRP-USD
+  tickers: ADA-USD,BNB-USD,BTC-USD,DOGE-USD,ETH-USD,LINK-USD,XLM-USD,XRP-USD
   max_prompt_length: 1536
   max_response_length: 256
 
@@ -58,4 +58,6 @@ verl:
     nnodes: 1
     n_gpus_per_node: 8
     save_freq: 20
+    test_freq: 20
+    val_before_train: true
     total_epochs: 1

examples/baseline_rollout.py

@@ -12,6 +12,7 @@
 
 from examples.build_seed_bank import crypto_frames_from_panel
 from examples.dsl_smoke_test import DEFAULT_CRYPTO_PANEL, load_dotenv
+from src.data import verify_crypto_panel
 from src.rft import (
     DatabaseSelectionConfig,
     MinerConfig,
@@ -33,19 +34,19 @@
 
 def run_crypto_baseline_rollout(
     *,
-    oracle_model: str,
-    miner_model: str,
+    oracle_model: str = "deepseek/deepseek-chat-v3.1",
+    miner_model: str = "qwen/qwen3-235b-a22b-2507",
     panel_path: Path = DEFAULT_CRYPTO_PANEL,
     tickers: tuple[str, ...] | None = None,
-    oracle_count: int = 12,
+    oracle_count: int = 24,
     miner_count: int = 4,
     top_k_seeds: int = 3,
     output_dir: Path = DEFAULT_OUTPUT_DIR,
 ) -> dict:
     """Run oracle seed generation plus miner baseline rollouts on crypto data."""
 
     load_dotenv()
-    panel = pd.read_pickle(panel_path)
+    panel = verify_crypto_panel(panel_path, tickers=tickers)
     frames = crypto_frames_from_panel(panel, tickers=tickers)
     close = panel["close"]
     scenario = FactorScenario.from_benchmark(
@@ -70,7 +71,7 @@ def run_crypto_baseline_rollout(
         windows=windows,
         raw_candidates=raw_candidates,
         min_history=30,
-        min_assets=max(3, min(5, len(frames))),
+        min_assets=8,
         pool_config=SeedPoolConfig(top_k=top_k_seeds, quality_threshold=-1.0),
     )
 
@@ -84,7 +85,7 @@ def run_crypto_baseline_rollout(
             price_col="close",
             count=miner_count,
             min_history=30,
-            min_assets=max(3, min(5, len(frames))),
+            min_assets=8,
             miner_config=MinerConfig(model=miner_model),
             archive=archive,
             selection_config=DatabaseSelectionConfig(min_score=-1.0, min_reward=-1.0),
@@ -143,8 +144,8 @@ def main() -> int:
     parser.add_argument("--oracle-model", default="deepseek/deepseek-chat-v3.1")
     parser.add_argument("--miner-model", default="qwen/qwen3-235b-a22b-2507")
     parser.add_argument("--crypto-panel", type=Path, default=DEFAULT_CRYPTO_PANEL)
-    parser.add_argument("--tickers", default="BTC-USD,ETH-USD,XRP-USD")
-    parser.add_argument("--oracle-count", type=int, default=12)
+    parser.add_argument("--tickers", default="ADA-USD,BNB-USD,BTC-USD,DOGE-USD,ETH-USD,LINK-USD,XLM-USD,XRP-USD")
+    parser.add_argument("--oracle-count", type=int, default=24)
     parser.add_argument("--miner-count", type=int, default=4)
     parser.add_argument("--top-k-seeds", type=int, default=3)
     parser.add_argument("--output-dir", type=Path, default=DEFAULT_OUTPUT_DIR)

examples/build_seed_bank.py

@@ -7,7 +7,7 @@
 
 import pandas as pd
 
-from src.data import adapt_crypto_ohlcv, adapt_crsp_dsf_v2
+from src.data import adapt_crsp_dsf_v2, adapt_crypto_ohlcv, verify_crypto_panel
 from examples.dsl_smoke_test import (
     DEFAULT_CRYPTO_PANEL,
     fetch_wrds_crsp_dsf_v2_sample,
@@ -25,11 +25,8 @@
 def crypto_frames_from_panel(panel: dict, tickers: tuple[str, ...] | None = None) -> dict[str, pd.DataFrame]:
     """Convert saved crypto panel dict into per-asset DSL-ready frames."""
 
+    verify_crypto_panel(panel, tickers=tickers)
     required = ["open", "high", "low", "close", "volume"]
-    missing = [key for key in required if key not in panel]
-    if missing:
-        raise ValueError(f"crypto panel missing keys: {missing}")
-
     available = tuple(str(col) for col in panel["close"].columns)
     selected = tickers or available
     frames = {}
@@ -51,7 +48,7 @@ def build_crypto_cross_sectional_seed_bank(
 ) -> dict:
     """Build a cross-sectional crypto seed bank from the saved project panel."""
 
-    panel = pd.read_pickle(panel_path)
+    panel = verify_crypto_panel(panel_path, tickers=tickers)
     frames = crypto_frames_from_panel(panel, tickers=tickers)
     close = panel["close"]
     windows = [
@@ -71,7 +68,7 @@ def build_crypto_cross_sectional_seed_bank(
         windows=windows,
         raw_candidates=raw_candidates,
         min_history=30,
-        min_assets=max(3, min(5, len(frames))),
+        min_assets=8,
         pool_config=SeedPoolConfig(top_k=top_k, quality_threshold=quality_threshold),
     )
     return {"result": result, "asset_count": len(frames)}
@@ -128,7 +125,7 @@ def main() -> int:
     parser.add_argument("--candidate-source", choices=("template", "openrouter", "file"), default="template")
     parser.add_argument("--oracle-output-file", type=Path)
     parser.add_argument("--oracle-count", type=int, default=24)
-    parser.add_argument("--openrouter-model", default="deepseek/deepseek-chat")
+    parser.add_argument("--openrouter-model", default="deepseek/deepseek-chat-v3.1")
     parser.add_argument("--wrds-crsp", action="store_true", help="also build a tiny live WRDS CRSP seed bank")
     parser.add_argument("--wrds-permno", type=int, default=14593)
     parser.add_argument("--wrds-start", default="2023-01-01")

examples/check_openrouter_tokens.py

@@ -42,7 +42,7 @@ def main() -> int:
     parser.add_argument("--namespace")
     parser.add_argument("--count", type=int, default=12)
     parser.add_argument("--seed-expr", default="div(ts_mean(crypto.volume(10)), ts_std(crypto.returns(30)))")
-    parser.add_argument("--seed-score", type=float, default=0.655671862964597)
+    parser.add_argument("--seed-score", type=float, default=0.38385972330719526)
     parser.add_argument("--max-tokens", type=int, default=1)
     args = parser.parse_args()
 

examples/dsl_smoke_test.py

@@ -13,7 +13,7 @@
 
 import pandas as pd
 
-from src.data import adapt_crypto_ohlcv, adapt_crsp_dsf_v2
+from src.data import adapt_crsp_dsf_v2, adapt_crypto_ohlcv, verify_crypto_panel
 from src.dsl import PointInTimeContext, evaluate_expr
 
 
@@ -71,11 +71,8 @@ def load_dotenv(path: Path | None = None) -> None:
 def crypto_asset_frame_from_panel(panel: dict, ticker: str) -> pd.DataFrame:
     """Build one asset OHLCV frame from the saved crypto panel dictionary."""
 
+    verify_crypto_panel(panel, tickers=(ticker,))
     required = ["open", "high", "low", "close", "volume"]
-    missing = [key for key in required if key not in panel]
-    if missing:
-        raise ValueError(f"crypto panel missing keys: {missing}")
-
     frame = pd.DataFrame({key: panel[key][ticker] for key in required})
     if "returns" in panel:
         frame["returns"] = panel["returns"][ticker]
@@ -106,7 +103,7 @@ def evaluate_last_available(
 def run_crypto_smoke(panel_path: Path = DEFAULT_CRYPTO_PANEL, ticker: str = "BTC-USD") -> dict[str, float]:
     """Run the crypto namespace smoke test on the saved project panel."""
 
-    panel = pd.read_pickle(panel_path)
+    panel = verify_crypto_panel(panel_path, tickers=(ticker,))
     frame = crypto_asset_frame_from_panel(panel, ticker)
     adapted = adapt_crypto_ohlcv(frame)
     value = evaluate_last_available("crypto", adapted, CRYPTO_EXPR, min_history=25)

examples/evaluate_rollout_library.py

@@ -8,11 +8,12 @@
 
 import pandas as pd
 
-from src.evaluation import PostSelectionConfig, evaluate_factor_library
-from src.evaluation.post_selection import write_evaluation_report
 from examples.baseline_rollout import DEFAULT_OUTPUT_DIR
 from examples.build_seed_bank import crypto_frames_from_panel
 from examples.dsl_smoke_test import DEFAULT_CRYPTO_PANEL
+from src.data import verify_crypto_panel
+from src.evaluation import PostSelectionConfig, evaluate_factor_library
+from src.evaluation.post_selection import write_evaluation_report
 
 
 def load_valid_rollout_exprs(path: Path) -> list[str]:
@@ -46,7 +47,7 @@ def run_crypto_rollout_library_evaluation(
     """Run validation-guided decorrelated selection and equal-weight fusion."""
 
     exprs = load_valid_rollout_exprs(rollout_json)
-    panel = pd.read_pickle(panel_path)
+    panel = verify_crypto_panel(panel_path, tickers=tickers)
     frames = crypto_frames_from_panel(panel, tickers=tickers)
     if validation_start is None or validation_end is None or test_start is None or test_end is None:
         dates = pd.DatetimeIndex(panel["close"].index).sort_values()
@@ -69,7 +70,7 @@ def run_crypto_rollout_library_evaluation(
             correlation_threshold=correlation_threshold,
             top_k=top_k,
             min_history=30,
-            min_assets=max(3, min(5, len(frames))),
+            min_assets=8,
             horizon=1,
         ),
     )
@@ -82,7 +83,7 @@ def main() -> int:
     parser.add_argument("--rollout-json", type=Path, default=DEFAULT_OUTPUT_DIR / "baseline_rollout.json")
     parser.add_argument("--crypto-panel", type=Path, default=DEFAULT_CRYPTO_PANEL)
     parser.add_argument("--output-path", type=Path, default=DEFAULT_OUTPUT_DIR / "paper_style_evaluation.json")
-    parser.add_argument("--tickers", default="BTC-USD,ETH-USD,XRP-USD")
+    parser.add_argument("--tickers", default="ADA-USD,BNB-USD,BTC-USD,DOGE-USD,ETH-USD,LINK-USD,XLM-USD,XRP-USD")
     parser.add_argument("--validation-start")
     parser.add_argument("--validation-end")
     parser.add_argument("--test-start")

examples/launch_verl_qwen3_14b_a100.sh

@@ -10,7 +10,7 @@ set -euo pipefail
 # with a fallback default, so build_dataset below and the eventual Verl
 # reward_fn agree on the same crypto data/tickers/output locations).
 export FACTOR_LAB_CRYPTO_PANEL="${FACTOR_LAB_CRYPTO_PANEL:-data/crypto/crypto_panel_clean.pkl}"
-export FACTOR_LAB_TICKERS="${FACTOR_LAB_TICKERS:-BTC-USD,ETH-USD,XRP-USD}"
+export FACTOR_LAB_TICKERS="${FACTOR_LAB_TICKERS:-ADA-USD,BNB-USD,BTC-USD,DOGE-USD,ETH-USD,LINK-USD,XLM-USD,XRP-USD}"
 export FACTOR_LAB_ARCHIVE_JSONL="${FACTOR_LAB_ARCHIVE_JSONL:-outputs/verl/qwen3_14b_fullft/mined_factors.jsonl}"
 export FACTOR_LAB_REWARD_LOG_JSONL="${FACTOR_LAB_REWARD_LOG_JSONL:-outputs/verl/qwen3_14b_fullft/reward_rollouts.jsonl}"
 
@@ -19,7 +19,7 @@ export FACTOR_LAB_REWARD_LOG_JSONL="${FACTOR_LAB_REWARD_LOG_JSONL:-outputs/verl/
 python -m src.verl_integration.build_dataset \
   --output outputs/verl/crypto_grpo_tasks.parquet \
   --seed-expr "div(ts_mean(crypto.volume(10)), ts_std(crypto.returns(30)))" \
-  --seed-score 0.655671862964597 \
+  --seed-score 0.38385972330719526 \
   --crypto-panel "${FACTOR_LAB_CRYPTO_PANEL}" \
   --tickers "${FACTOR_LAB_TICKERS}" \
   --repeats 400
@@ -29,7 +29,7 @@ python -m src.verl_integration.build_dataset \
 python -m src.verl_integration.build_dataset \
   --output outputs/verl/crypto_grpo_val_tasks.parquet \
   --seed-expr "div(ts_mean(crypto.volume(10)), ts_std(crypto.returns(30)))" \
-  --seed-score 0.655671862964597 \
+  --seed-score 0.38385972330719526 \
   --crypto-panel "${FACTOR_LAB_CRYPTO_PANEL}" \
   --tickers "${FACTOR_LAB_TICKERS}" \
   --repeats 32

examples/run_compared_approaches.py

@@ -14,7 +14,7 @@ def main() -> int:
     parser = argparse.ArgumentParser(description="Run compared alpha-discovery benchmark approaches.")
     parser.add_argument("--model", default="qwen/qwen3-235b-a22b-2507")
     parser.add_argument("--crypto-panel", type=Path, default=DEFAULT_CRYPTO_PANEL)
-    parser.add_argument("--tickers", default="BTC-USD,ETH-USD,XRP-USD")
+    parser.add_argument("--tickers", default="ADA-USD,BNB-USD,BTC-USD,DOGE-USD,ETH-USD,LINK-USD,XLM-USD,XRP-USD")
     parser.add_argument(
         "--approaches",
         default="alphabench,quantaalpha,rd_agent,alpha_jungle,factor_lab",