scripts/score.py

#!/usr/bin/env python3
"""Score AIME++ JSONL predictions with deterministic exact matching."""

from __future__ import annotations

import argparse
import json
import re
from collections import defaultdict
from pathlib import Path


ROOT = Path(__file__).resolve().parents[1]
CONFIG_FILES = {
    "all": (
        "aime.jsonl",
        "aime-hard.jsonl",
        "aime-graduate.jsonl",
        "aime-researcher.jsonl",
    ),
    "aime": ("aime.jsonl",),
    "aime-hard": ("aime-hard.jsonl",),
    "aime-graduate": ("aime-graduate.jsonl",),
    "aime-researcher": ("aime-researcher.jsonl",),
}
STRICT_ANSWER = re.compile(r"\s*([0-9]{1,3})\s*")
BOXED_ANSWER = re.compile(r"\\boxed\{\s*([0-9]{1,3})\s*\}")


def parse_prediction(value: object, allow_boxed: bool) -> int | None:
    if isinstance(value, bool):
        return None
    if isinstance(value, int):
        return value if 0 <= value <= 999 else None
    if not isinstance(value, str):
        return None

    strict = STRICT_ANSWER.fullmatch(value)
    if strict:
        return int(strict.group(1))
    if allow_boxed:
        boxed = BOXED_ANSWER.findall(value)
        if boxed:
            return int(boxed[-1])
    return None


def load_gold(data_dir: Path, config: str) -> dict[str, dict[str, object]]:
    gold: dict[str, dict[str, object]] = {}
    for filename in CONFIG_FILES[config]:
        path = data_dir / filename
        with path.open(encoding="utf-8") as handle:
            for line_number, line in enumerate(handle, start=1):
                record = json.loads(line)
                record_id = record["id"]
                if record_id in gold:
                    raise ValueError(f"duplicate gold id {record_id!r} in {path}:{line_number}")
                gold[record_id] = record
    return gold


def load_predictions(path: Path) -> dict[str, object]:
    predictions: dict[str, object] = {}
    with path.open(encoding="utf-8") as handle:
        for line_number, line in enumerate(handle, start=1):
            if not line.strip():
                continue
            record = json.loads(line)
            if not isinstance(record, dict) or "id" not in record or "prediction" not in record:
                raise ValueError(f"{path}:{line_number}: expected fields 'id' and 'prediction'")
            record_id = record["id"]
            if not isinstance(record_id, str):
                raise ValueError(f"{path}:{line_number}: id must be a string")
            if record_id in predictions:
                raise ValueError(f"{path}:{line_number}: duplicate prediction id {record_id!r}")
            predictions[record_id] = record["prediction"]
    return predictions


def main() -> int:
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument("predictions", type=Path, help="JSONL with id and prediction fields")
    parser.add_argument("--data-dir", type=Path, default=ROOT / "data")
    parser.add_argument(
        "--config",
        choices=tuple(CONFIG_FILES),
        default="all",
        help="gold configuration to score (default: all)",
    )
    parser.add_argument(
        "--allow-boxed",
        action="store_true",
        help=r"also accept the last \boxed{N} found in a string; strict whole-string matching is the default",
    )
    parser.add_argument("--json", action="store_true", help="emit machine-readable JSON")
    args = parser.parse_args()

    gold = load_gold(args.data_dir, args.config)
    predictions = load_predictions(args.predictions)
    unknown_ids = sorted(set(predictions) - set(gold))

    correct = 0
    valid = 0
    submitted = 0
    by_tier: dict[str, dict[str, int]] = defaultdict(lambda: {"correct": 0, "total": 0})
    for record_id, record in gold.items():
        tier = str(record["tier"])
        by_tier[tier]["total"] += 1
        if record_id not in predictions:
            continue
        submitted += 1
        parsed = parse_prediction(predictions[record_id], args.allow_boxed)
        if parsed is None:
            continue
        valid += 1
        if parsed == record["answer"]:
            correct += 1
            by_tier[tier]["correct"] += 1

    total = len(gold)
    report = {
        "config": args.config,
        "accuracy": correct / total if total else 0.0,
        "correct": correct,
        "total": total,
        "submitted": submitted,
        "valid": valid,
        "invalid": submitted - valid,
        "missing": total - submitted,
        "unknown_ids": unknown_ids,
        "tiers": {
            tier: {
                **counts,
                "accuracy": counts["correct"] / counts["total"] if counts["total"] else 0.0,
            }
            for tier, counts in by_tier.items()
        },
    }

    if args.json:
        print(json.dumps(report, indent=2, sort_keys=True))
    else:
        print(f"overall: {correct}/{total} ({report['accuracy']:.2%})")
        print(
            f"coverage: submitted={submitted}, valid={valid}, "
            f"invalid={submitted - valid}, missing={total - submitted}"
        )
        for tier, counts in report["tiers"].items():
            print(f"- {tier}: {counts['correct']}/{counts['total']} ({counts['accuracy']:.2%})")
        if unknown_ids:
            print(f"unknown prediction ids: {', '.join(unknown_ids)}")
    return 0


if __name__ == "__main__":
    raise SystemExit(main())