schift-io/koocr-bench-v2

KoOCR-Bench v2.0c

Korean document OCR benchmark: tables, official gazette, statistics, KDI reports, user documents, 2026 CSAT (수능), handwriting, accounting disclosures, drug labels, patents, medical forms. 12 slices, 877 public pages + 193 held out (1,070 in all). v2.0c (2026-10-29): handwritten_rw11 is reduced to the 111 pages whose truth was reviewed by hand.

Not for training: every page here is eval-only.

Slices

slice	pages (this repo)	held out	license / source
gwanbo_table	81	19	관보 (public official gazette)
handwritten_owner20	18	2	국립한글박물관 소장자료 (rights per source); 교보 손글씨 (rights per source); realworld photo (owner-provided); AIHub 81 (NIA AI Hub; rights per AIHub terms)
handwritten_rw11	90	21	국립한글박물관 소장자료 (rights per source); 교보 손글씨 (rights per source); realworld photo (owner-provided); AIHub 81 (NIA AI Hub; rights per AIHub terms)
kdi_headings	78	22	KDI (owner-cleared for public release 2026-09-30)
kostat_wide	80	20	KOGL (per-article type) — 1유형 (출처표시), logo in attachment
suneung2026	87	17	KICE 2026 수능 문제지 (owner-provided, eval only)
tables	86	14	KOGL 1유형 (공공누리 제1유형: 출처표시)
user_real	58	17	ggugeo ToS (owner-cleared for public release 2026-09-30)
accounting_table	87	13	DART 공시 (public disclosure)
drug_label	80	20	MFDS 의약품안전나라 허가정보 (owner-cleared 2026-10-27)
patent_exp	85	15	KIPO 특허공보 (published gazette; owner-cleared 2026-10-27)
medical_form_synth	47	13	synthetic (generated, no real patient data)
total	877	193

Suneung: owner reports KICE OK for this release (2026-10-27).

handwritten_owner20 is a 20-page subset of the handwriting pages with owner-corrected truth (same images as rw11).

Total (v3 weights, 2026-10-31)

bash score.sh A=a.jsonl | python3 scorer/koocr_total_v3.py prints one weighted total per engine (higher is better; table slices use teds, text slices 1 - cer). v3 raises tables and handwriting; slice ratios inside each group are the v2 ones.

group	slices (weight)	sum
tables	kostat_wide .159 · tables .136 · gwanbo_table .136 (scored by CER) · accounting_table .068	.50
text	suneung2026 .056 · kdi_headings .044 · user_real .044 · patent_exp .022 · drug_label .022 · medical_form_synth .011	.20
handwriting	handwritten_owner20 .10 · handwritten_rw11 .10	.20
reading order	page-weighted mean of order over every slice that reports it	.10

A slice an engine has no rows for is dropped and the remaining weights are renormalised (row marked partial). The v2 weights (tables .44 · text .36 · handwriting .10 · order .10) stay available as scorer/koocr_total_v2.py.

Results (v3 total, 2026-11-01; CLOVA and kordoc added 2026-11-09; LightOnOCR-3 added 2026-11-11; PepperOCR-VL added 2026-11-12)

Scored on all 1,070 pages (public + holdout). Higher is better. tbl_form / tbl_text split the grid-table slices into grid shape and cell text; they are shown, not added to the total.

engine	total	tables	text	hand	order	tbl_form	tbl_text
Claude Opus 5.5 (claude-opus-5-5, API)	0.924	0.920	0.901	0.948	0.941	0.900	0.902
OpenAI Codex (gpt-6-sol, API)	0.918	0.913	0.897	0.945	0.934	0.895	0.889
Schift OCR 2 (API)	0.899	0.896	0.894	0.892	0.935	0.877	0.859
Schift OCR 1 beta	0.862	0.862	0.869	0.823	0.928	0.873	0.842
datalab-to/chandra-ocr-2	0.860	0.886	0.804	0.816	0.933	0.870	0.850
NAVER CLOVA OCR (General + table extraction, API)	0.853	0.874	0.736	0.943	0.800	0.884	0.875
kordoc 4.24.57 (GitHub 1d3dab1, local pipeline)	0.853	0.814	0.856	0.917	0.915	0.774	0.766
Qwen/Qwen3.8-27B-FP8	0.852	0.883	0.862	0.724	0.931	0.868	0.850
PaddlePaddle/PaddleOCR-VL-1.6	0.851	0.879	0.814	0.755	0.976	0.872	0.864
sionic-ai/PepperOCR-VL	0.810	0.751	0.849	0.836	0.976	0.726	0.673
lightonai/LightOnOCR-3-0.8B	0.776	0.848	0.861	0.435	0.928	0.845	0.795
MinerU 2.5 (local pipeline)	0.758	0.804	0.810	0.478	0.984	0.759	0.750
OpenGVLab/InternVL3_5-38B	0.753	0.753	0.688	0.738	0.910	0.797	0.704
mistralai/Mistral-Small-3.2-24B-Instruct-2506	0.696	0.756	0.676	0.435	0.962	0.772	0.678
google/gemma-4-26B-A4B-it	0.689	0.686	0.612	0.656	0.919	0.753	0.612

Schift OCR rows are our models (weights not released).

NAVER CLOVA OCR: General OCR V2 API with enableTableDetection: true, lang: ko, run 2026-11-09, score only (commercial API, predictions not published). Output is the API's words in API order (lineBreak ends a line); each detected table is written once as HTML <table> (colspan/rowspan from the API) where its first word falls, and words inside a table box are not repeated. Three handwriting photos over the API's 8,000 px limit were shrunk to 8,000 px on the long side. Its low text comes mostly from suneung2026: the API returns two-column exam pages as one large table, so the two columns interleave.

sionic-ai/PepperOCR-VL (AGPL-3.0, fine-tune of Qwen/Qwen3.5-4B): run 2026-11-12 with the repo's prompt.txt verbatim, processor default image size, thinking off, temperature 0, top_p 1, repetition_penalty 1.0, max_tokens 8192, one run. The card's defaults (repetition_penalty 1.05, no_repeat_ngram_size 30, retries at a higher temperature) were not used, so it runs under the same plain greedy decoding as every other engine here. 47 printed pages stopped at the 8,192-token cap; no handwriting page did. Its low tables comes mostly from kostat_wide (0.598) and tables (0.685); its order (0.976) is among the highest on the board.

lightonai/LightOnOCR-3-0.8B (Apache-2.0): run 2026-11-11 with the model card's image-only prompt (no text instruction), longest side resized to 2048 px, thinking off, temperature 0 (the card suggests 0.2), repetition_penalty 1.0, max_tokens 8192. 27 pages (7 printed, 20 handwriting) stopped at the 8,192-token cap. Its low hand is handwriting CER (owner20 0.549, rw11 0.581).

kordoc: open-source npm package by chrisryugj (not ours), GitHub HEAD 1d3dab1 (2 commits after npm 4.24.57), built from source and run on page images with default options (--jobs), output markdown with HTML tables, run 2026-11-09. Its CHANGELOG reports tuning against the public 877 pages; the score here is on all 1,070 pages, including the 193-page holdout it never saw.

Holdout

20% of every slice is held out and is not in this repo: a page is held out when int(sha1(id), 16) % 5 == 0 (for owner20 the hash is taken on its rw11 id, so a page is in the same split in both slices). Only counts are published.

Files

• manifest.jsonl — one row per page: id, slice, tier, image, truth, source, license (+ math_text_only, gt, split). truth.type is table_grid (rows of cells with text, cs, rs), page_text (text = path under truth/), or headings+page_text (adds headings = [[level, title], ...]).
• images/, truth/
• scorer/ — KoOCR-Bench v1.3r scorer (koocr_score_v13r.py), GriTS on rendered grids (grits_render.py, unmodified microsoft/table-transformer grits.py, MIT), TEDS from opendataloader-bench (odlb_src/).

Scoring

pip install -r requirements.txt
bash score.sh MYENGINE=preds.jsonl      # preds rows: {"id": "<manifest id>", "text": "<engine output>"}

Output is format-neutral: Markdown syntax and HTML tags are stripped before text metrics; tables must be HTML <table>. Per slice: tables/kostat_wide → GriTS-based teds (0.5 form + 0.5 content) plus TEDS/TEDS-S/grid_exact; text slices → cer (pred blocks re-ordered to truth order), order, foreign; suneung → mark_keep; kdi → heading_recall. Per-page metrics go to scores.pages.jsonl.