grader.py

"""Deterministic, judge-free grader for BrowseComp-Plus answers.

grade(pred, gold) -> (correct: bool, category: str)
category is one of: exact, alias, numeric, date, gold_in_pred, all_parts, fuzzy, name_subset, no_answer, mismatch
"""
from __future__ import annotations

import re
import unicodedata

ARTICLES = {"a", "an", "the"}
TITLES = {"dr", "mr", "mrs", "ms", "prof", "professor", "sir", "dame", "lord", "lady", "st", "saint", "jr", "sr"}


def extract_answer(text: str | None) -> str | None:
    """Pull the final short answer out of a free-form agent response."""
    if text is None:
        return None
    t = str(text)
    if not t.strip():
        return None
    # "Exact Answer: X" (official BrowseComp response template), tolerate markdown
    m = list(re.finditer(r"exact[ \t]*answer[ \t]*\**[ \t]*[:：][ \t]*\**[ \t]*(.*)", t, re.IGNORECASE))
    if m:
        ans = m[-1].group(1).strip()
        if not ans.strip("* "):
            return None
        ans = re.split(r"\n|\s{2,}confidence\s*:", ans, flags=re.IGNORECASE)[0]
        return ans.strip().strip("*").strip()
    m = list(re.finditer(r"\\boxed\{((?:[^{}]|\{[^{}]*\})*)\}", t))
    if m:
        return m[-1].group(1).strip()
    m = list(re.finditer(r"<answer>(.*?)</answer>", t, re.IGNORECASE | re.DOTALL))
    if m:
        return extract_answer(m[-1].group(1)) or m[-1].group(1).strip()
    m = list(re.finditer(r"final\s*answer\s*\**\s*[:：]\s*\**\s*(.+)", t, re.IGNORECASE))
    if m:
        return m[-1].group(1).strip().strip("*").strip()
    lines = [l.strip() for l in t.strip().splitlines() if l.strip()]
    lines = [l for l in lines if not re.match(r"^\**confidence", l, re.IGNORECASE)]
    return lines[-1] if lines else None


def norm(s: str) -> str:
    s = unicodedata.normalize("NFKD", str(s))
    s = "".join(c for c in s if not unicodedata.combining(c))
    s = s.lower().replace("&", " and ")
    s = re.sub(r"[\u2018\u2019'`]s\b", "s", s)
    s = re.sub(r"[^\w\s.]", " ", s)          # keep dots for decimals for now
    s = re.sub(r"(?<!\d)\.|\.(?!\d)", " ", s)  # drop non-decimal dots
    toks = [w for w in s.split() if w not in ARTICLES and w not in TITLES]
    return " ".join(toks)


def _num(s: str):
    """Return float if s is essentially a single number (with optional unit/commas/%)."""
    x = str(s).lower().replace(",", "").strip()
    x = re.sub(r"^(approximately|approx\.?|about|around|~|\$|usd|eur|£|€)\s*", "", x)
    m = re.fullmatch(r"\$?\s*(-?\d+(?:\.\d+)?)\s*(%|[a-z]{0,12}\.?)?", x)
    if not m:
        return None
    try:
        return float(m.group(1))
    except ValueError:
        return None


def gold_aliases(gold: str) -> list[str]:
    g = str(gold).strip()
    out = {g}
    # parenthetical aliases: "X (Y)" -> X, Y
    m = re.fullmatch(r"(.+?)\s*\((.+)\)\s*", g)
    if m:
        out.add(m.group(1)); out.add(m.group(2))
    for sep in [" / ", "/", " or ", " aka ", " a.k.a. "]:
        if sep in g:
            for part in g.split(sep):
                if part.strip():
                    out.add(part.strip())
    return [a for a in out if norm(a)]


MONTHS = {m: i + 1 for i, m in enumerate(["january", "february", "march", "april", "may", "june", "july", "august",
                                           "september", "october", "november", "december"])}
MONTHS.update({k[:3]: v for k, v in list(MONTHS.items())})
MONTHS["sept"] = 9


def parse_dates(s: str) -> set:
    """All (y, m, d) / (y, m) date mentions in s (month-name forms and ISO / D/M/Y numeric)."""
    t = norm(s)
    out = set()
    mon = r"(" + "|".join(sorted(MONTHS, key=len, reverse=True)) + r")"
    for m in re.finditer(r"(\d{1,2}) (?:of )?" + mon + r" (\d{4})", t):
        out.add((int(m.group(3)), MONTHS[m.group(2)], int(m.group(1))))
    for m in re.finditer(mon + r" (\d{1,2})(?:st|nd|rd|th)? (\d{4})", t):
        out.add((int(m.group(3)), MONTHS[m.group(1)], int(m.group(2))))
    for m in re.finditer(r"(\d{4}) " + mon + r" (\d{1,2})\b", t):
        out.add((int(m.group(1)), MONTHS[m.group(2)], int(m.group(3))))
    for m in re.finditer(r"\b(\d{4})[-/](\d{1,2})[-/](\d{1,2})\b", str(s)):
        out.add((int(m.group(1)), int(m.group(2)), int(m.group(3))))
    for m in re.finditer(r"\b(\d{1,2})[/.](\d{1,2})[/.](\d{4})\b", str(s)):
        out.add((int(m.group(3)), int(m.group(2)), int(m.group(1))))
    return out


def _contains(hay: str, needle: str) -> bool:
    return bool(re.search(r"(?<!\w)" + re.escape(needle) + r"(?!\w)", hay))


def gold_parts(gold: str) -> list[str]:
    """split a multi-part gold ("A, B", "A & B", "A; B", "A and B") into parts."""
    parts = re.split(r"\s*(?:,|;|&|\band\b)\s*", str(gold))
    parts = [norm(p) for p in parts if norm(p)]
    return parts if len(parts) >= 2 else []


def grade(pred: str | None, gold: str) -> tuple[bool, str]:
    if pred is None or not norm(pred):
        return False, "no_answer"
    np_ = norm(pred)
    gn = norm(gold)
    if np_ == gn:
        return True, "exact"
    aliases = gold_aliases(gold)
    # prediction variants: as given, and with parentheticals removed / taken alone
    pvars = {np_}
    pv0 = norm(re.sub(r"\([^)]*\)", " ", str(pred)))
    if pv0:
        pvars.add(pv0)
    for m in re.finditer(r"\(([^)]*)\)", str(pred)):
        if norm(m.group(1)):
            pvars.add(norm(m.group(1)))
    for a in aliases:
        if norm(a) in pvars:
            return True, "alias"
    # numeric tolerance
    gv, pv = _num(gold), _num(pred)
    if gv is not None and pv is not None:
        if gv == pv or (not (gv.is_integer() and pv.is_integer() and abs(gv) < 10000) and gv != 0 and abs(pv - gv) / abs(gv) <= 0.01):
            return True, "numeric"
        return False, "mismatch"
    # dates
    gd = parse_dates(gold)
    if gd:
        pd_ = parse_dates(pred)
        if pd_:
            return (True, "date") if gd <= pd_ else (False, "mismatch")
    ptoks = np_.split()
    # verbose prediction containing the full gold answer (word-bounded); length cap depends on gold specificity
    for a in aliases:
        an = norm(a)
        if not an:
            continue
        specific = len(an) >= 6 and not re.fullmatch(r"[\d. ]+", an)
        cap = 40 if specific else 12
        if (len(an) >= 3 or gv is not None) and _contains(np_, an) and len(ptoks) <= len(an.split()) + cap:
            return True, "gold_in_pred"
    # multi-part gold: every part present in the prediction
    parts = gold_parts(gold)
    if parts and all(_contains(np_, p) for p in parts) and len(ptoks) <= len(gn.split()) + 40:
        return True, "all_parts"
    # near-identical spelling (typos, transliteration)
    import difflib
    for a in aliases:
        an = norm(a)
        for v in pvars:
            if len(an) >= 6 and difflib.SequenceMatcher(None, an, v).ratio() >= 0.92:
                return True, "fuzzy"
    # person-name with a dropped middle name: same first and last token, pred tokens subset of gold
    gt = gn.split()
    for v in pvars:
        vt = v.split()
        if len(vt) >= 2 and len(gt) >= 3 and vt[0] == gt[0] and vt[-1] == gt[-1] and set(vt) <= set(gt):
            return True, "name_subset"
    return False, "mismatch"


def extract_freeform(text: str | None) -> str | None:
    """For long free-form reports with no answer template: bold spans of the first
    substantive paragraph (joined), else that paragraph."""
    if text is None or not str(text).strip():
        return None
    t = str(text)
    if re.search(r"exact[ \t]*answer[ \t]*\**[ \t]*[:：]|\\boxed\{|<answer>", t, re.IGNORECASE):
        return extract_answer(t)
    m = re.search(r"<[｜|]DSML[｜|]answer>", t)
    if m:
        t = t[m.end():]
    paras = [p.strip() for p in re.split(r"\n\s*\n", t) if p.strip()]
    paras = [p for p in paras if not re.fullmatch(r"#+[^\n]*|-{3,}", p)]
    if not paras:
        return None
    p = paras[0]
    bolds = re.findall(r"\*\*(.+?)\*\*", p)
    if bolds:
        return " ; ".join(b.strip() for b in bolds[:4])
    return p


def grade_freeform(report: str | None, pred: str | None, gold: str) -> tuple[bool, str]:
    """Free-form report grading: grade the extracted answer; if that fails, accept when a
    *specific* gold (>=8 chars or >=2 tokens, not purely numeric) appears verbatim (normalized,
    word-bounded) anywhere in the report. Lenient: can over-credit reports that list candidates."""
    c, k = grade(pred, gold)
    if c or report is None:
        return c, k
    nr = norm(report)
    for a in gold_aliases(gold):
        an = norm(a)
        if re.fullmatch(r"[\d. ]+", an):
            continue
        if (len(an) >= 8 or len(an.split()) >= 2) and _contains(nr, an):
            return True, "gold_in_report"
    return c, k