| |
| """Baseline: predict failure_mode from first 3 trajectory iterations.""" |
|
|
| from __future__ import annotations |
|
|
| import argparse |
| import json |
| from pathlib import Path |
|
|
| ROOT = Path(__file__).resolve().parents[1] |
| DEFAULT_JSONL = ROOT / "data" / "seed" / "records.jsonl" |
|
|
|
|
| def load_records(path: Path) -> list[dict]: |
| records = [] |
| with path.open(encoding="utf-8") as handle: |
| for line in handle: |
| line = line.strip() |
| if line: |
| records.append(json.loads(line)) |
| return records |
|
|
|
|
| def extract_features(record: dict, window: int = 3) -> list[float]: |
| trajectory = record["trajectory"][:window] |
| if len(trajectory) < window: |
| trajectory = trajectory + [trajectory[-1]] * (window - len(trajectory)) |
|
|
| features: list[float] = [] |
| for step in trajectory: |
| features.extend( |
| [ |
| step["goal_score"], |
| step.get("primary_quality", step["goal_score"]), |
| step["cost_usd"], |
| step["latency_seconds"], |
| float(step.get("safety_events", 0)), |
| float(step.get("human_intervention", False)), |
| ] |
| ) |
|
|
| features.append(record["metadata"]["goal_target"]) |
| features.append(float(record["metadata"]["worker_count"])) |
| features.append(float(record["metadata"]["evaluator_count"])) |
| return features |
|
|
|
|
| def main(argv: list[str] | None = None) -> int: |
| parser = argparse.ArgumentParser(description=__doc__) |
| parser.add_argument("--jsonl", type=Path, default=DEFAULT_JSONL) |
| parser.add_argument("--window", type=int, default=3) |
| args = parser.parse_args(argv) |
|
|
| try: |
| from sklearn.ensemble import RandomForestClassifier |
| from sklearn.metrics import classification_report |
| from sklearn.pipeline import Pipeline |
| from sklearn.preprocessing import StandardScaler |
| except ImportError as exc: |
| raise SystemExit("scikit-learn required: pip install -e '.[notebook]'") from exc |
|
|
| records = load_records(args.jsonl) |
| labeled = [r for r in records if r.get("failure_mode")] |
|
|
| train = [r for r in labeled if r["split"] == "train"] |
| test = [r for r in labeled if r["split"] == "test"] |
|
|
| if not train or not test: |
| print("Need labeled train and test splits with failure_mode.") |
| return 1 |
|
|
| x_train = [extract_features(r, args.window) for r in train] |
| y_train = [r["failure_mode"] for r in train] |
| x_test = [extract_features(r, args.window) for r in test] |
| y_test = [r["failure_mode"] for r in test] |
|
|
| model = Pipeline( |
| [ |
| ("scaler", StandardScaler()), |
| ("clf", RandomForestClassifier(n_estimators=200, random_state=42)), |
| ] |
| ) |
| model.fit(x_train, y_train) |
| predictions = model.predict(x_test) |
|
|
| print(classification_report(y_test, predictions, zero_division=0)) |
| accuracy = (predictions == y_test).mean() |
| print(f"Accuracy on test split: {accuracy:.1%} ({len(test)} records)") |
| return 0 |
|
|
|
|
| if __name__ == "__main__": |
| raise SystemExit(main()) |
|
|
