Leaderboard

davanstrien/ocr-bench-britannica-results

Rankings are computed using Bradley-Terry MLE from pairwise comparisons judged by a vision-language model. The judge sees the original document image alongside two anonymised OCR outputs and picks the more faithful transcription. Browse the comparisons to see the evidence — and vote yourself to build a Human ELO column. Human votes are stored locally for this session only and will reset when the server restarts.

#	Model	Params	Judge ELO	95% CI	Wins	Losses	Ties	Win%
1	dots.mocr	3B	1745	1714–1782	436	141	4	75%
2	LightOnOCR-2-1B	1B	1741	1709–1779	426	141	4	75%
3	GLM-OCR	0.9B	1738	1707–1773	469	157	2	75%
4	olmOCR-2-7B-1025-FP8	—	1719	1688–1753	454	167	4	73%
5	NuExtract3	4B	1690	1657–1725	435	190	0	70%
6	Qianfan-OCR	4.7B	1568	1539–1600	347	279	2	55%
7	FireRed-OCR	2.1B	1557	1528–1588	338	286	4	54%
8	Unlimited-OCR	—	1545	1518–1575	312	281	0	53%
9	PaddleOCR-VL-1.6	0.9B	1463	1435–1493	267	353	1	43%
10	DeepSeek-OCR	4B	1452	1422–1482	260	363	4	41%
11	PP-OCRv6_medium	—	1397	1367–1428	220	393	0	36%
12	DeepSeek-OCR-2	—	1380	1345–1415	211	409	5	34%
13	tesseract-5	—	1113	1062–1153	79	514	0	13%
14	dots.ocr	1.7B	891	797–963	24	604	0	4%

ELO vs Parameter Count

Smaller models can win on the right documents. Error bars show 95% confidence intervals.