from __future__ import annotations import argparse import json import re from collections import Counter, defaultdict from decimal import Decimal, InvalidOperation from pathlib import Path from typing import Any ARROWS = "↑↓▲▼△▽↗↘⇧⇩↟↡" NUM_RE = re.compile(r"^[+-]?\d+(\.\d+)?$") WS_RE = re.compile(r"\s+") SHADE_PREFIXES = {"light", "dark", "medium", "pale", "deep", "bright"} BASE_COLORS = { "blue", "green", "purple", "pink", "red", "orange", "yellow", "gray", "grey", } def strip_arrows(s: str) -> str: return "".join(ch for ch in s if ch not in ARROWS) def normalize_value(x: Any) -> Any: if x is None: return None if isinstance(x, bool): return x if isinstance(x, (int, float)): return normalize_number(str(x)) if isinstance(x, str): s = strip_arrows(x.strip()) while len(s) >= 2 and (s[0], s[-1]) in [("(", ")"), ("[", "]"), ("{", "}"), ("<", ">")]: s = s[1:-1].strip() s = s.strip(" ,;") s = s.replace("−", "-").replace("—", "-").replace("–", "-") s = re.sub(r"\s*±\s*", "±", s) s = re.sub(r"\s*\+\s*/\s*-\s*", "±", s) s = re.sub(r"\s*\+\s*-\s*", "±", s) s = re.sub(r"\s*%\s*", "%", s) s = WS_RE.sub(" ", s).strip() s = normalize_color_name(s) # Optional: uncomment to ignore citation suffixes, e.g. "Foggy Zurich [52]". # s = re.sub(r"\s*\[\d+\]", "", s).strip() if NUM_RE.match(s): return normalize_number(s) return s return x def normalize_color_name(s: str) -> str: color = s.lower().replace("-", "_").replace(" ", "_") parts = [p for p in color.split("_") if p] if len(parts) >= 2 and parts[0] in SHADE_PREFIXES and parts[-1] in BASE_COLORS: return "gray" if parts[-1] == "grey" else parts[-1] if color in BASE_COLORS: return "gray" if color == "grey" else color return s def normalize_number(s: str) -> str: try: d = Decimal(s) except InvalidOperation: return s.strip() if d == 0: return "0" out = format(d.normalize(), "f") return out.rstrip("0").rstrip(".") if "." in out else out def parse_answer(ans: Any) -> Any: if not isinstance(ans, str): return ans text = ans.strip() for _ in range(3): if text.startswith("[") or text.startswith("{") or (text.startswith('"') and text.endswith('"')): try: parsed = json.loads(text) except Exception: break if isinstance(parsed, str): text = parsed.strip() continue return parsed break return ans def normalize_record(x: dict[str, Any]) -> str: return json.dumps({k: normalize_value(v) for k, v in x.items()}, ensure_ascii=False, sort_keys=True) def multiset_score(gt_items: list[Any], pr_items: list[Any]) -> tuple[float, bool, int, int, int]: gt = Counter(gt_items) pr = Counter(pr_items) keys = gt.keys() | pr.keys() tp = sum(min(gt[k], pr[k]) for k in keys) fp = sum(max(pr[k] - gt[k], 0) for k in keys) fn = sum(max(gt[k] - pr[k], 0) for k in keys) exact = fp == 0 and fn == 0 score = 1.0 if exact else (2 * tp / (2 * tp + fp + fn) if (2 * tp + fp + fn) else 1.0) return score, exact, tp, fp, fn def score_answer(gt: Any, pred: Any) -> dict[str, Any]: pred = parse_answer(pred) if isinstance(gt, dict): exact = normalize_record(gt) == normalize_record(pred) if isinstance(pred, dict) else False return {"score": 1.0 if exact else 0.0, "exact": exact} if isinstance(gt, list): if all(isinstance(x, dict) for x in gt): gt_items = [normalize_record(x) for x in gt] pr_list = pred if isinstance(pred, list) else ([pred] if isinstance(pred, dict) else []) pr_items = [normalize_record(x) for x in pr_list if isinstance(x, dict)] else: gt_items = [normalize_value(x) for x in gt] pr_list = pred if isinstance(pred, list) else [pred] pr_items = [normalize_value(x) for x in pr_list] score, exact, tp, fp, fn = multiset_score(gt_items, pr_items) return {"score": score, "exact": exact, "tp": tp, "fp": fp, "fn": fn} exact = normalize_value(gt) == normalize_value(pred) return {"score": 1.0 if exact else 0.0, "exact": exact} def load_jsonl(path: Path) -> list[dict[str, Any]]: return [json.loads(line) for line in path.read_text(encoding="utf-8").splitlines() if line.strip()] def main() -> int: parser = argparse.ArgumentParser(description="Score compact HighlightBench QA predictions.") parser.add_argument("--gt", required=True) parser.add_argument("--pred", required=True) parser.add_argument("--out", default=None) args = parser.parse_args() gt_rows = load_jsonl(Path(args.gt)) pred_rows = load_jsonl(Path(args.pred)) pred_by_qid = {r.get("qid"): r for r in pred_rows} scores = [] exacts = [] per_item = [] by_split = defaultdict(list) for gt in gt_rows: qid = gt.get("qid") pred = pred_by_qid.get(qid, {}) res = score_answer(gt.get("answer"), pred.get("answer")) scores.append(float(res["score"])) exacts.append(bool(res["exact"])) split = gt.get("dataset", "unknown") by_split[split].append(float(res["score"])) per_item.append({"qid": qid, "score": res["score"], "exact": res["exact"]}) summary = { "num_items": len(gt_rows), "overall_mean": sum(scores) / len(scores) if scores else 0.0, "overall_exact_rate": sum(exacts) / len(exacts) if exacts else 0.0, "per_split_mean": {k: sum(v) / len(v) for k, v in by_split.items()}, "missing_predictions": sum(1 for r in gt_rows if r.get("qid") not in pred_by_qid), } out = {"summary": summary, "per_item": per_item} out_path = Path(args.out) if args.out else Path(args.pred).with_suffix(".score.json") out_path.write_text(json.dumps(out, ensure_ascii=False, indent=2) + "\n", encoding="utf-8") print(out_path) return 0 if __name__ == "__main__": raise SystemExit(main())