davanstrien/ocr-board-historical-v0

OCR leaderboard: 20th-century government documents (v0)

9 OCR models ranked on 338 pages. For each page, a vision-language model compared two transcriptions at a time and picked the more faithful one, or called a tie. ocr-bench ran the comparisons and fitted the ratings. The pages, the transcriptions and the judge's reasons stay private: this repo holds scores only.

How to read it. A higher rating is better; 1500 is the average rating on each board. The 95% interval is where the rating stays when the pages are resampled: when two intervals overlap, this data does not clearly separate those models. Size is the parameter count the Hub reports for the model's weights, where it has one (all experts, for a mixture-of-experts model). Pages is the number of pages a model was compared on. Wins, losses and ties are raw counts that depend on how often a model was compared, so compare ratings, not wins.

All pages (338)

Rank	Model	Size	Rating	95% interval	Pages	Wins	Losses	Ties
1	lightonai/LightOnOCR-3-0.8B	0.85B	1643	1620–1666	338	1501	611	208
2	tiiuae/Falcon-OCR	0.27B	1599	1579–1619	338	1361	746	227
3	ATH-MaaS/OvisOCR2	0.85B	1567	1545–1586	338	1283	872	206
4	rednote-hilab/dots.ocr	3B	1557	1537–1576	338	1237	904	178
5	datalab-to/surya-ocr-2.5	0.69B	1549	1529–1569	338	1203	920	190
6	zai-org/GLM-OCR	1.3B	1518	1499–1538	338	1293	1220	151
7	baidu/Unlimited-OCR-2 (tiled) ⚠	3.3B	1508	1489–1528	331	1042	1037	199
8	PaddlePaddle/PaddleOCR-VL-1.6	0.96B	1399	1374–1427	338	806	1577	145
9	tesseract-5	n/a	1161	1127–1192	338	244	2083	16

⚠ marks a model that returned an error message (such as [OCR ERROR]) instead of text on some pages. Those pages were skipped for it, so its rating rests on the pages in the Pages column. Error outputs in the source dataset: baidu/Unlimited-OCR-2 (tiled): 10.

By production

One board for each value of the production column, fitted only on the pages with that value. A model needs 10 of those pages to appear.

blank_or_near_blank (ranked on 60 pages)

Rank	Model	Size	Rating	95% interval	Pages	Wins	Losses	Ties
1	datalab-to/surya-ocr-2.5	0.69B	1699	1626–1789	60	84	21	1
2	baidu/Unlimited-OCR-2 (tiled)	3.3B	1693	1625–1783	58	81	21	1
3	rednote-hilab/dots.ocr	3B	1660	1591–1740	60	80	24	2
4	lightonai/LightOnOCR-3-0.8B	0.85B	1634	1570–1706	60	78	27	1
5	zai-org/GLM-OCR	1.3B	1581	1529–1636	60	263	183	0
6	tesseract-5	n/a	1472	1378–1571	60	69	57	2
7	tiiuae/Falcon-OCR	0.27B	1361	1298–1413	60	51	62	2
8	ATH-MaaS/OvisOCR2	0.85B	1341	1240–1440	60	54	94	5
9	PaddlePaddle/PaddleOCR-VL-1.6	0.96B	1060	954–1138	60	22	293	0

letterpress_or_offset_print (ranked on 118 pages)

Rank	Model	Size	Rating	95% interval	Pages	Wins	Losses	Ties
1	lightonai/LightOnOCR-3-0.8B	0.85B	1629	1595–1668	118	573	265	103
2	ATH-MaaS/OvisOCR2	0.85B	1608	1575–1642	118	549	294	97
3	tiiuae/Falcon-OCR	0.27B	1604	1572–1634	118	544	300	97
4	datalab-to/surya-ocr-2.5	0.69B	1580	1549–1614	118	515	333	91
5	rednote-hilab/dots.ocr	3B	1576	1546–1607	118	519	350	71
6	baidu/Unlimited-OCR-2 (tiled)	3.3B	1542	1511–1576	115	450	372	98
7	zai-org/GLM-OCR	1.3B	1491	1456–1523	118	412	467	61
8	PaddlePaddle/PaddleOCR-VL-1.6	0.96B	1342	1294–1388	118	236	650	55
9	tesseract-5	n/a	1129	1074–1173	118	84	851	5

photocopy_or_microfilm_reproduction (ranked on 13 pages)

Rank	Model	Size	Rating	95% interval	Pages	Wins	Losses	Ties
1	lightonai/LightOnOCR-3-0.8B	0.85B	1662	1558–1838	13	71	23	10
2	zai-org/GLM-OCR	1.3B	1575	1454–1715	13	61	38	5
3	ATH-MaaS/OvisOCR2	0.85B	1539	1465–1625	13	51	40	10
4	tiiuae/Falcon-OCR	0.27B	1539	1466–1612	13	53	41	10
5	rednote-hilab/dots.ocr	3B	1488	1375–1611	13	45	49	10
6	datalab-to/surya-ocr-2.5	0.69B	1487	1383–1587	13	42	47	12
7	PaddlePaddle/PaddleOCR-VL-1.6	0.96B	1463	1339–1566	13	42	54	7
8	baidu/Unlimited-OCR-2 (tiled)	3.3B	1450	1341–1552	13	39	55	6
9	tesseract-5	n/a	1297	1056–1430	13	20	77	4

typed_or_printed_with_handwriting (ranked on 16 pages)

Rank	Model	Size	Rating	95% interval	Pages	Wins	Losses	Ties
1	lightonai/LightOnOCR-3-0.8B	0.85B	1805	1686–2020	16	108	20	0
2	tiiuae/Falcon-OCR	0.27B	1606	1492–1777	16	79	49	0
3	datalab-to/surya-ocr-2.5	0.69B	1591	1506–1725	16	75	50	3
4	ATH-MaaS/OvisOCR2	0.85B	1583	1502–1700	16	75	53	0
5	zai-org/GLM-OCR	1.3B	1539	1457–1649	16	66	59	3
6	rednote-hilab/dots.ocr	3B	1522	1423–1640	16	63	62	3
7	PaddlePaddle/PaddleOCR-VL-1.6	0.96B	1441	1293–1578	16	51	77	0
8	baidu/Unlimited-OCR-2 (tiled)	3.3B	1432	1311–1545	16	48	77	3
9	tesseract-5	n/a	982	678–1106	16	5	123	0

typewritten (ranked on 126 pages)

Rank	Model	Size	Rating	95% interval	Pages	Wins	Losses	Ties
1	lightonai/LightOnOCR-3-0.8B	0.85B	1651	1616–1687	126	649	261	92
2	tiiuae/Falcon-OCR	0.27B	1635	1602–1670	126	618	271	118
3	ATH-MaaS/OvisOCR2	0.85B	1565	1535–1596	126	528	378	94
4	rednote-hilab/dots.ocr	3B	1550	1515–1582	126	509	404	89
5	datalab-to/surya-ocr-2.5	0.69B	1522	1493–1554	126	473	447	80
6	zai-org/GLM-OCR	1.3B	1515	1485–1547	126	465	460	82
7	PaddlePaddle/PaddleOCR-VL-1.6	0.96B	1494	1458–1533	126	432	487	83
8	baidu/Unlimited-OCR-2 (tiled)	3.3B	1479	1448–1511	125	405	501	89
9	tesseract-5	n/a	1089	1019–1142	126	66	936	5

5 judged pages are in none of these boards because they have no label, their label has fewer than 10 pages, or they were only compared with models left out of their board.

How this board was made

• Judge: novita:Qwen/Qwen3.8-27B, with the body-text criteria (the prompts in ocr-bench 0.0.8)
• Pages: 340 pages sampled (seed 42); 338 had at least one comparison
• Comparisons: 10,730; 49 of them had identical text and were tied without a judge call
• Ratings: a Bradley-Terry fit on an Elo-like scale; 95% intervals from resampling pages
• Judged: 2026-11-09

The full settings (prompt hash, text and image caps, page-set fingerprint) are in board.json.

Use the data

In Python, with from datasets import load_dataset: load_dataset("davanstrien/ocr-board-historical-v0") gives the overall board; load_dataset("davanstrien/ocr-board-historical-v0", "slices") gives the boards by label.

Query with DuckDB

Each row is one model on one board. elo is the rating, and elo_low and elo_high are its 95% interval.

-- a private board needs a Hub token first:
-- CREATE SECRET hf (TYPE huggingface, PROVIDER credential_chain);
SELECT rank, model, elo, elo_low, elo_high, pages FROM 'hf://datasets/davanstrien/ocr-board-historical-v0/overall.jsonl' ORDER BY rank;
-- the top model for each label
SELECT slice, model, elo, elo_low, elo_high FROM 'hf://datasets/davanstrien/ocr-board-historical-v0/slices.jsonl' WHERE rank = 1 ORDER BY slice;

Generated by ocr-bench 0.0.8