Fix four scoring defects surfaced by the DAT-3459 SIG rescore (spotting, KIE syntax, grounding bbox, boxed JSON)

[FHLiang221]

What

Four scoring defects in datbench/datasets_scoring/ that graded answer format instead of content — found by rescoring the DAT-3459 SIG-scaling arms from saved predictions. This package scores db-document (the DatBench composite), so these defects shaped every db-document number.

1. Text spotting was 'simplified to VQA' — the model's box list was string-matched against the flattened polygon answer, which can essentially never match, so every run scored a hard 0. Replaced with the official RRC-style single-image metric (axis-aligned GT polygons, ### ignore regions, greedy IoU>0.5 match, recognition precision/recall h-mean) plus the official box extractor with \boxed{} unwrapping.
2. KIE F1 zeroed the GT's own answer syntax — _generate_combinations unpacks GT list values to scalars but predictions were stringified as-is, so {'key': ['value']} compared "['value']" vs "value". The ground truth fed back as a prediction scored 0.0. Fix: mirror the unwrap on the prediction side (single-element lists only).
3. Grounding/VQA-with-position rows never earned their IoU half where the exported row's bbox field is null (db-document) — the GT box exists only as four numeric answer strings. Fix: rebuild it from the answers when bbox is absent; an explicit bbox still wins.
4. CC-OCR KIE returned {} for \boxed{"k": "v"} — the brace-matching extractor hands json.loads a brace-less body when the boxed braces swallow the JSON braces. Fix: restore the braces.

Impact (DAT-3459 arms, rescored locally from saved predictions)

• Spotting: 8-way tie at 0 → real scores (0.8–3.5%); the all-synthetic arm stays ~0 legitimately (fabricated coordinates)
• OCRBench Extraction: SIG-trained arms 0.9–54.7 → 61.5–72.0; pure-mined arms unchanged (the repair is a no-op where there was no artifact)
• Referring: all real-data arms roughly double (e.g. 5.8 → 13.7); CC-OCR KIE: the all-SIG arm 41.8 → 49.2, others unchanged

None of the repairs can turn a wrong value into a right one. Canary tests included (tests/test_sig_scoring_fixes.py, 14 tests).

Companion

datnanovlm #489 applies the same fixes to datvlmeval's standalone-eval copies of these scorers; after this merges, datnanovlm needs a submodule pointer bump.

🤖 Generated with Claude Code