code/koocr_score.py

#!/usr/bin/env python3
"""kobench v1 scorer (copy of kbench score_kbench_mp.py + FOREIGN_W / MD_TABLES / th-td normalisation / bbox).
Score engine outputs on kbench v1. Format-neutral: HTML tags and Markdown syntax are stripped before text metrics
(the 2026-09-29 GT audit showed markup-sensitive CER misranks engines). Per slice:
  tables        TEDS and TEDS-S (apted, odlb's evaluator) of the best-matching predicted table vs the HWPX grid,
                plus grid_exact (row-width sequence equal)
  suneung2026   CER vs PDF text; retention of ㉠~㉤, ⓐ~ⓔ, ①~⑤, <보기> (count in pred / count in truth, capped at 1)
  user_real     CER vs PDF text
  kdi_headings  heading recall (truth bookmark titles found in a heading line: Native title block or Markdown '#'), CER
Usage: score_kbench.py MANIFEST ENGINE=PRED.jsonl ... [--tier core|full]   (PRED rows: {"id": kbench id, "text": output})"""

import json, re, sys, unicodedata, collections, html as H

sys.path.insert(
    0,
    __import__("os").environ.get("ODLB_SRC", "opendataloader-bench/src"),
)
from evaluator_table import TEDSEvaluator
from converter_markdown_table import convert_to_markdown_with_html_tables
from rapidfuzz.distance import (
    Levenshtein,
)  # runs in the odlb uv env (apted + rapidfuzz)

argv = sys.argv[1:]
tier = "full"
if "--tier" in argv:
    i = argv.index("--tier")
    tier = argv[i + 1]
    argv = argv[:i] + argv[i + 2 :]
args = argv
M = [json.loads(l) for l in open(args[0])]
M = [r for r in M if tier in r["tier"]]
import os as _o
FOREIGN_W = float(_o.environ.get("FOREIGN_W", "2"))   # extra errors per foreign char not in truth (1 edit + W)
MD_TABLES = _o.environ.get("MD_TABLES", "zero")  # owner 10/04: our HTML output is the standard    # convert | zero (only real HTML <table> counts)
FOREIGN_RE = re.compile("[㐀-䶿一-鿿豈-﫿぀-ヿㇰ-ㇿЀ-ӿ"
                        "฀-๿؀-ۿ֐-׿，。：；！？]")


NORM = _o.environ.get("NORM", "strong")  # basic = kbench plain() | strong = + NFKC, letters/digits only


def nrm(x):
    # strong: NFKC and keep letters/digits only, so punctuation, bullets, arrows, quotes and dashes never count
    if NORM == "basic":
        return x
    x = unicodedata.normalize("NFKC", x)
    return "".join(ch for ch in x if ch.isalnum())


def foreign_extra(p, t):
    """Foreign-script chars (Han, kana, Cyrillic, Thai, Arabic, Hebrew, Chinese full-width punctuation) the prediction
    has beyond the truth's own count of the same char. Hanja that the truth really contains costs nothing."""
    pc = collections.Counter(FOREIGN_RE.findall(p)); tc = collections.Counter(FOREIGN_RE.findall(t))
    return sum(max(0, n - tc[c]) for c, n in pc.items())


BOX_RE = re.compile(r"<\|det\|>[^<]*?\[\[?\s*\d+\s*,\s*\d+\s*,\s*\d+\s*,\s*\d+\s*\]\]?[^<]*?<\|/det\|>")


def bbox_rate(out):
    """Share of output blocks that carry a 4-number box: boxed det spans / (det spans + unboxed paragraphs)."""
    boxed = len(BOX_RE.findall(out))
    rest = re.sub(r"<\|det\|>.*?<\|/det\|>[^\n]*", "\n", out, flags=re.S)
    rest = re.sub(r"<\|ref\|>.*?<\|/ref\|>", "", rest, flags=re.S)
    loose = [b for b in re.split(r"\n\s*\n|</table>|<table", rest) if plain(b)]
    return boxed / max(1, boxed + len(loose))


def order_score(pp, t):
    """Reading order: truth lines (>=6 chars) located in the prediction by a 10-char key; score = longest increasing
    subsequence of their positions / located lines. None when fewer than 2 lines are located."""
    import bisect
    pos = []
    for ln in t.splitlines():
        k = plain(ln)[:10]
        if len(k) >= 6:
            i = pp.find(k)
            if i >= 0:
                pos.append(i)
    if len(pos) < 2:
        return None
    tails = []
    for x in pos:
        j = bisect.bisect_left(tails, x)
        tails[j:j + 1] = [x]
    return len(tails) / len(pos)


def norm_table(t):
    t = re.sub(r"<(/?)th\b", r"<\1td", t, flags=re.I)
    t = re.sub(r"</?(thead|tbody|tfoot)[^>]*>", "", t, flags=re.I)
    return re.sub(r"<(td|tr|table)\b([^>]*)>", lambda m: "<" + m[1] + "".join(re.findall(r"\s(?:rowspan|colspan)=\"?\d+\"?", m[2])) + ">", t, flags=re.I)


MARKS = list("㉠㉡㉢㉣㉤ⓐⓑⓒⓓⓔ①②③④⑤") + ["<보기>"]


def plain(s):
    s = unicodedata.normalize("NFC", s)
    s = re.sub(r"<\|det\|>.*?<\|/det\|>", "\n", s)
    s = re.sub(
        r"<\|ref\|>.*?<\|/ref\|>", "\n", s
    )  # DeepSeek-OCR grounding labels (table, sub_title, ...)
    s = re.sub(r"<[^>]+>", " ", s)
    s = H.unescape(s)
    s = re.sub(r"(?m)^\s*#{1,6}\s*", "", s)
    s = re.sub(r"[|*_`]|:?-{3,}:?|!\[[^\]]*\]\([^)]*\)|\$\$?|\\\\[()\[\]]", " ", s)
    return re.sub(r"\s+", "", s)


def cer(p, t):
    p, t = plain(p), plain(t)
    return min(1.0, Levenshtein.distance(p, t) / max(1, len(t)))


def grid_html(rows):
    return (
        "<table>"
        + "".join(
            "<tr>"
            + "".join(
                f'<td colspan="{c["cs"]}" rowspan="{c["rs"]}">{H.escape(c["text"])}</td>'
                for c in r
            )
            + "</tr>"
            for r in rows
        )
        + "</table>"
    )


def widths(t):
    return [len(re.findall(r"<td", r)) for r in re.findall(r"<tr>(.*?)</tr>", t, re.S)]


_t, _ts = TEDSEvaluator(structure_only=False), TEDSEvaluator(structure_only=True)
wrap = lambda t: f"<html><body>{t}</body></html>"


class _W:
    def __init__(s, e):
        s.e = e

    def evaluate(s, p, g):
        return s.e.evaluate(wrap(p), wrap(g))


teds, teds_s = _W(_t), _W(_ts)


def tables_of(out):
    out = re.sub(r"<\|det\|>.*?<\|/det\|>", "\n", out)
    if MD_TABLES == "convert":
        out = convert_to_markdown_with_html_tables(out)
    # 2026-10-01: malformed span (rowspan="" / "abc" / "0") crashed odlb int(); treat it as span 1
    out = re.sub(r'''\s(?:rowspan|colspan)\b(?:\s*=\s*(?:"([^"]*)"|'([^']*)'|([^\s>]*)))?''', lambda m: m[0] if re.fullmatch(r'\s*[1-9]\d*\s*', next((g for g in m.groups() if g is not None), '')) else ' ', out, flags=re.I)
    return [norm_table(x) for x in re.findall(r"<table.*?</table>", out, re.S | re.I)]


def headings_of(out):
    hs = [m for m in re.findall(r"<\|det\|>title \[[^\]]*\]<\|/det\|>([^\n]*)", out)]
    hs += re.findall(r"(?m)^\s*#{1,6}\s*(.+)$", out)
    return [plain(h) for h in hs]


def score_one(job):
    eng, r, out = job
    sl = r["slice"]
    S = collections.defaultdict(list)
    if out is None:
        S["missing"].append(1)
        return eng, sl, dict(S), r["id"]
    tr = r["truth"]
    S["bbox"].append(bbox_rate(out))
    if tr["type"] == "table_grid":
        g = grid_html(tr["rows"])
        cands = tables_of(out)
        if not cands:
            S["teds"].append(0); S["teds_s"].append(0); S["grid_exact"].append(0)
            return eng, sl, dict(S), r["id"]
        def _safe(f, a, b):
            try:
                return f.evaluate(a, b)
            except Exception:
                return 0.0  # malformed pred table (e.g. rowspan="") scores 0; the sequential scorer crashes here
        sc = [(_safe(teds, c, g), c) for c in cands]
        best, c = max(sc, key=lambda x: x[0])
        gt_plain = plain(" ".join(x["text"] for row in tr["rows"] for x in row))
        S["foreign"].append(foreign_extra(plain(out), gt_plain))
        best = max(0.0, best - FOREIGN_W * S["foreign"][-1] / max(1, len(gt_plain)))
        S["teds"].append(best); S["teds_s"].append(_safe(teds_s, c, g)); S["grid_exact"].append(float(widths(c) == widths(g)))
    else:
        t = open(tr["text"]).read()
        if r["id"].startswith("suneung2026/math"):
            out = re.sub(r"\\\(.*?\\\)|\\\[.*?\\\]|\$\$.*?\$\$|\$[^$\n]*\$", " ", out, flags=re.S)
        pp, tp = plain(out), plain(t)
        fx = foreign_extra(pp, tp); S["foreign"].append(fx)
        _o_s = order_score(pp, t)
        if _o_s is not None:
            S["order"].append(_o_s)
        np_, nt_ = nrm(pp), nrm(tp)
        S["cer"].append(min(1.0, (Levenshtein.distance(np_, nt_) + FOREIGN_W * fx) / max(1, len(nt_))))
        if sl == "suneung2026":
            for mk in MARKS:
                n = t.count(mk)
                if n:
                    S["mark_keep"].append(min(1.0, out.count(mk) / n))
        if tr["type"] == "headings+page_text":
            hs = headings_of(out)
            for lv, title in tr["headings"]:
                k = plain(title)
                S["heading_recall"].append(float(any(k and (k in h or h in k) and len(h) >= 0.6 * len(k) for h in hs)))
    return eng, sl, dict(S), r["id"]


import multiprocessing as _mp, os as _os

jobs = []
for spec in args[1:]:
    eng, path = spec.split("=", 1)
    P = {}
    for l in open(path):
        d = json.loads(l)
        P[d["id"]] = d.get("text") or ""
    for r in M:
        jobs.append((eng, r, P.get(r["id"])))
res = collections.defaultdict(lambda: collections.defaultdict(list))
with _mp.get_context("fork").Pool(int(_os.environ.get("SCORE_PROCS", "32"))) as pool:
    _dump = open(_os.environ["DUMP"], "w") if _os.environ.get("DUMP") else None  # per-page metrics for re-weighting
    for eng, sl, S, pid in pool.imap(score_one, jobs, chunksize=2):
        if _dump:
            _dump.write(json.dumps({"engine": eng, "slice": sl, "id": pid, **S}, ensure_ascii=False) + "\n")
        for k, v in S.items():
            res[(eng, sl)][k] += v
print(f"tier={tier}")
for (eng, sl), S in sorted(res.items(), key=lambda x: (x[0][1], x[0][0])):
    parts = [f"{k} {sum(v) / len(v):.4f}" for k, v in S.items() if k != "missing" and v]
    print(
        f"{sl:14s} {eng:10s} n={max((len(v) for k, v in S.items() if k != 'missing'), default=0):4d} missing={len(S['missing'])} "
        + " ".join(parts)
    )