| """Score named typed answers for Structured State Reasoning Probe.""" |
| import argparse |
| import json |
| from pathlib import Path |
|
|
|
|
| def load_jsonl(path): |
| return [json.loads(line) for line in Path(path).read_text().splitlines() if line.strip()] |
|
|
|
|
| def normalize(spec, raw): |
| if spec["type"] == "choice": |
| return raw.get("choice") if isinstance(raw, dict) else raw |
| if spec["type"] == "noul": |
| if isinstance(raw, bool): |
| return raw |
| if isinstance(raw, dict) and isinstance(raw.get("noul"), (int, float)): |
| return raw["noul"] >= 0.5 |
| return None |
|
|
|
|
| def main(): |
| parser = argparse.ArgumentParser() |
| parser.add_argument("predictions", help="JSONL rows: task_id and answers map") |
| parser.add_argument("--dataset", default=str(Path(__file__).resolve().parent / "test.jsonl")) |
| args = parser.parse_args() |
| dataset = load_jsonl(args.dataset) |
| predictions = load_jsonl(args.predictions) |
| tasks = {row["task_id"]: row for row in dataset} |
| pred = {row["task_id"]: row for row in predictions} |
| if len(tasks) != len(dataset) or len(pred) != len(predictions): |
| raise SystemExit("Duplicate task_id in data or predictions") |
| if set(tasks) != set(pred): |
| raise SystemExit(f"Prediction coverage mismatch: missing={sorted(set(tasks)-set(pred))}, extra={sorted(set(pred)-set(tasks))}") |
| task_pass = question_correct = question_total = 0 |
| per_task = [] |
| for tid, row in tasks.items(): |
| questions = json.loads(row["questions_json"]) |
| gold = json.loads(row["gold_json"]) |
| answers = pred[tid]["answers"] |
| if set(answers) != set(questions): |
| raise SystemExit(f"Question coverage mismatch for {tid}") |
| checks = {qid: normalize(spec, answers[qid]) == gold[qid] for qid, spec in questions.items()} |
| passed = all(checks.values()) |
| task_pass += passed |
| question_correct += sum(checks.values()) |
| question_total += len(checks) |
| per_task.append({"task_id": tid, "pass": passed, "questions": checks}) |
| print(json.dumps({"tasks_passed": task_pass, "tasks_total": len(tasks), |
| "questions_correct": question_correct, "questions_total": question_total, |
| "per_task": per_task}, indent=2)) |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
