evaluate/score.py

#!/usr/bin/env python3
"""Score model outputs on the AudioSpan release.

Two modes (both stdlib-only, no external dependencies):
  accuracy — multiple choice on native audio, per-layer breakdown.
  chain    — first-error truncation over P->U->R chains on anchor audio.

For rubric scoring (which requires a judge model), use score_rubric.py.

The answer file carries only what the model produced, one record per
question: {"qa_id": ..., "answer": "..."}. For multiple choice the scorer
extracts the option letter from the answer text. Correct answers and
question profiles are joined in from metadata/<mode>/{S,M,L}.jsonl;
questions with no answer count as wrong.

Usage:
    python score.py --mode accuracy --input results/<model>/accuracy.jsonl
    python score.py --mode chain    --input results/<model>/chain.jsonl
"""

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

logging.basicConfig(level=logging.INFO, format="%(asctime)s [%(levelname)s] %(message)s")
logger = logging.getLogger(__name__)

RELEASE_ROOT = Path(__file__).resolve().parent.parent
TIERS = ("S", "M", "L")
LAYER_ORDER = ["perception", "understanding", "reasoning"]
LAYER_BY_CODE = {"P": "perception", "U": "understanding", "R": "reasoning"}


def load_questions(mode: str, data_root: Path) -> list[dict]:
    questions = []
    for tier in TIERS:
        path = data_root / "metadata" / mode / f"{tier}.jsonl"
        if not path.is_file():
            sys.exit(f"ERROR: metadata not found: {path}")
        with open(path, encoding="utf-8") as fh:
            questions.extend(json.loads(line) for line in fh if line.strip())
    return questions


def extract_option(answer) -> str | None:
    """Normalize a model answer to an option letter (A-D), or None."""
    if answer is None:
        return None
    if not isinstance(answer, str):
        answer = str(answer)
    text = answer.strip()
    if re.fullmatch(r"[A-Da-d]", text):
        return text.upper()
    m = re.search(r"\b(?:answer|option|choice)\s*(?:is|:)?\s*[\(（]?([A-Da-d])[\)）]?\b",
                  text, re.IGNORECASE)
    if m:
        return m.group(1).upper()
    m = re.search(r"[\(（]([A-Da-d])[\)）]", text)
    if m:
        return m.group(1).upper()
    m = re.fullmatch(r"([A-Da-d])[\.\)、:：].*", text, re.DOTALL)
    if m:
        return m.group(1).upper()
    return None


def load_answers(path: str) -> dict[str, str]:
    """Map qa_id -> raw answer; later duplicates win."""
    preds: dict[str, str] = {}
    with open(path, encoding="utf-8", errors="replace") as fh:
        for lineno, line in enumerate(fh, 1):
            line = line.strip()
            if not line:
                continue
            try:
                rec = json.loads(line)
            except json.JSONDecodeError as e:
                logger.warning("Skipping bad line %d: %s", lineno, e)
                continue
            qa_id = rec.get("qa_id")
            if not qa_id:
                logger.warning("Skipping line %d: missing qa_id", lineno)
                continue
            preds[qa_id] = rec.get("answer")
    return preds


def score_accuracy(questions: list[dict], preds: dict[str, str]) -> dict:
    scored = []
    layer_stats: dict[str, dict[str, int]] = defaultdict(lambda: {"correct": 0, "total": 0})

    for q in questions:
        predicted = extract_option(preds.get(q["qa_id"]))
        is_correct = predicted is not None and predicted == q["correct_option"]
        layer = (q.get("question_profile") or {}).get("level", "unknown")
        layer_stats[layer]["total"] += 1
        if is_correct:
            layer_stats[layer]["correct"] += 1
        scored.append({"qa_id": q["qa_id"], "predicted": predicted, "correct": is_correct})

    total = len(scored)
    correct = sum(1 for s in scored if s["correct"])
    per_layer = {
        layer: {
            "accuracy": round(s["correct"] / s["total"] * 100, 2) if s["total"] else 0,
            "correct": s["correct"], "total": s["total"],
        }
        for layer, s in sorted(layer_stats.items())
    }
    return {
        "mode": "accuracy", "total": total, "correct": correct,
        "accuracy": round(correct / total * 100, 2) if total else 0,
        "per_layer": per_layer, "scored_records": scored,
    }


def score_chain(questions: list[dict], preds: dict[str, str]) -> dict:
    layer_correct: dict[str, int] = defaultdict(int)
    layer_total: dict[str, int] = defaultdict(int)
    chains: dict[str, dict[str, bool | None]] = defaultdict(lambda: {l: None for l in LAYER_ORDER})

    for q in questions:
        chain_key, layer_code = q["qa_id"].rsplit("_", 1)
        layer = LAYER_BY_CODE.get(layer_code)
        if layer not in LAYER_ORDER:
            continue
        predicted = extract_option(preds.get(q["qa_id"]))
        is_correct = predicted is not None and predicted == q["correct_option"]
        chains[chain_key][layer] = is_correct
        layer_total[layer] += 1
        if is_correct:
            layer_correct[layer] += 1

    complete = {cid: layers for cid, layers in chains.items()
                if all(v is not None for v in layers.values())}
    k = len(LAYER_ORDER)

    n_dist = defaultdict(int)
    scored = []
    group_scores = []
    for cid, layers in complete.items():
        n = 0
        for layer in LAYER_ORDER:
            if layers[layer]:
                n += 1
            else:
                break
        n_dist[n] += 1
        group_scores.append(n / k)
        scored.append({
            "chain_id": cid,
            "per_layer": {l: bool(layers[l]) for l in LAYER_ORDER},
            "n_correct": n,
            "correct": n == k,
            "chain_score": round(n / k * 100, 2),
        })

    chain_score = sum(group_scores) / len(group_scores) * 100 if group_scores else 0
    per_layer = {
        layer: {
            "accuracy": round(layer_correct[layer] / layer_total[layer] * 100, 2) if layer_total[layer] else 0,
            "correct": layer_correct[layer], "total": layer_total[layer],
        }
        for layer in LAYER_ORDER
    }
    return {
        "mode": "chain", "total_chains": len(complete),
        "chain_score": round(chain_score, 2),
        "per_layer": per_layer,
        "n_distribution": {str(n): cnt for n, cnt in sorted(n_dist.items())},
        "scored_records": scored,
    }


def main():
    parser = argparse.ArgumentParser(description="Score AudioSpan model outputs (accuracy/chain)")
    parser.add_argument("--input", required=True,
                        help="Prediction JSONL: {\"qa_id\": ..., \"answer\": \"...\"} per line")
    parser.add_argument("--mode", required=True, choices=["accuracy", "chain"])
    parser.add_argument("--data-root", type=Path, default=RELEASE_ROOT,
                        help="release root holding metadata/ (default: parent of evaluate/)")
    parser.add_argument("--output", help="Scored output path (default: <input>_scored.jsonl)")
    args = parser.parse_args()

    preds = load_answers(args.input)
    if not preds:
        print(f"ERROR: no valid answers in {args.input}", file=sys.stderr)
        sys.exit(1)

    questions = load_questions(args.mode, args.data_root.resolve())
    known = {q["qa_id"] for q in questions}
    unknown = sorted(set(preds) - known)
    if unknown:
        logger.warning("Ignoring %d unknown qa_id(s), e.g. %s", len(unknown), unknown[0])
    covered = sum(1 for q in questions if q["qa_id"] in preds)
    logger.info("Answers cover %d/%d questions (missing count as wrong)", covered, len(questions))

    summary = score_accuracy(questions, preds) if args.mode == "accuracy" else score_chain(questions, preds)

    out_path = args.output or args.input.replace(".jsonl", "_scored.jsonl")
    with open(out_path, "w", encoding="utf-8") as fh:
        for s in summary.get("scored_records", []):
            fh.write(json.dumps(s, ensure_ascii=False) + "\n")
    logger.info("Scored records: %s", out_path)

    print(json.dumps({k: v for k, v in summary.items() if k != "scored_records"},
                     indent=2, ensure_ascii=False))


if __name__ == "__main__":
    main()