tadad/kat57-ocr-bench-500-results

Kat57 OCR benchmark — CER/WER

Strict reference-based evaluation of 16 OCR models on a deterministic 500-card sample from Lund University Library's Kat57 catalogue-card collection. This result set contains only Character Error Rate (CER) and Word Error Rate (WER); it does not contain VLM judging or ELO ratings.

The sample was drawn with seed 57 from tadad/kat57-ground-truth and is published as tadad/kat57-ground-truth-500. The OCR outputs are retained in tadad/kat57-ocr-bench-500.

Rank	Model	CER	WER	Cards	Failed outputs
1	rednote-hilab/dots.mocr	0.1890	0.3781	500	0
2	zai-org/GLM-OCR	0.1980	0.4095	500	0
3	allenai/olmOCR-2-7B-1025-FP8	0.2018	0.3787	500	0
4	rednote-hilab/dots.ocr	0.2055	0.4013	500	0
5	PaddlePaddle/PaddleOCR-VL-1.6	0.2481	0.4691	500	0
6	PaddlePaddle/PP-OCRv6_medium	0.3030	0.5468	500	0
7	tiiuae/Falcon-OCR	0.3833	0.6273	500	0
8	numind/NuExtract3	0.4248	0.4830	500	0
9	baidu/Unlimited-OCR	0.4687	0.9121	500	0
10	tesseract-5	0.5279	0.9108	500	0
11	ATH-MaaS/OvisOCR2	0.5831	0.9452	500	0
12	baidu/Qianfan-OCR	0.7060	1.7825	500	0
13	deepseek-ai/DeepSeek-OCR-2	0.7219	1.1728	500	0
14	deepseek-ai/DeepSeek-OCR	0.9048	1.8508	500	0
15	FireRedTeam/FireRed-OCR	1.1686	1.5431	500	0
16	lightonai/LightOnOCR-2-1B	2.6549	2.5598	500	0

Lower is better. Scores use corpus-level edit totals rather than averaging per-card rates. Normalized scoring canonicalizes Unicode, flattens known HTML, and collapses whitespace while preserving case and punctuation.

What this benchmark measures

CER/WER measure sequential string agreement with the human PAGE XML reference transcription. They are useful for evaluating literal plain-text transcription, but they should not be interpreted as a complete ranking of document-understanding or layout-preservation quality.

The corpus scores above can be dominated by a small number of extremely long, repetitive outputs. The per-card distribution provides essential context:

Model	Corpus CER	Median card CER	95th percentile	Maximum	Exact cards	Cards CER > 1
rednote-hilab/dots.mocr	0.1890	0.1749	0.4343	0.600	10	0
zai-org/GLM-OCR	0.1980	0.1843	0.4451	0.756	8	0
allenai/olmOCR-2-7B-1025-FP8	0.2018	0.1951	0.4386	0.679	7	0
rednote-hilab/dots.ocr	0.2055	0.1965	0.4646	0.902	7	0
PaddlePaddle/PaddleOCR-VL-1.6	0.2481	0.2288	0.6154	1.846	1	5
PaddlePaddle/PP-OCRv6_medium	0.3030	0.3032	0.6937	0.864	3	0
tiiuae/Falcon-OCR	0.3833	0.2188	0.5455	17.201	4	7
numind/NuExtract3	0.4248	0.2475	0.5856	108.020	5	1
baidu/Unlimited-OCR	0.4687	0.2903	0.6479	231.635	6	3
tesseract-5	0.5279	0.5667	1.0000	1.321	1	4
ATH-MaaS/OvisOCR2	0.5831	0.2355	0.5506	187.966	6	2
baidu/Qianfan-OCR	0.7060	0.1953	0.4851	100.765	6	4
deepseek-ai/DeepSeek-OCR-2	0.7219	0.3137	0.7121	82.294	4	9
deepseek-ai/DeepSeek-OCR	0.9048	0.2676	0.6000	112.638	5	6
FireRedTeam/FireRed-OCR	1.1686	0.2155	0.5797	171.647	7	5
lightonai/LightOnOCR-2-1B	2.6549	0.5571	3.3672	130.712	4	128

Limitations and observed failure modes

• Nonlinear reading order: catalogue cards contain marginal classifications and spatially separated entries. PAGE XML imposes one sequence, while another output order may be equally usable but receive a much worse CER/WER.
• Structured output is penalized: Markdown and LaTeX that preserve underlining, grouping, or layout add characters absent from the plain-text reference. This can make a useful rendered representation score poorly.
• Editorial output is penalized: model notes about uncertain text, diagrams, or scan quality count as insertions even when they help a human reviewer.
• Reference quality varies: the human PAGE XML includes apparent transcription errors and debatable line ordering, especially on densely handwritten cards.
• Generation failures distort corpus scores: repetition loops produce unbounded insertions, so a few cards can make corpus CER much larger than typical per-card performance.

Representative examples include kat57-00047-00054 (LightOnOCR repetition loop), kat57-00047-00077, kat57-00047-00106, kat57-00047-00163, and kat57-00047-00182 (LaTeX, Markdown, or editorial representation mismatches), kat57-00047-00125 (questionable reference transcription and reading order), and kat57-00047-00156 (poor Tesseract recognition on handwriting).

These results are therefore best read as a strict PAGE-reference baseline and a stress test for messy archival catalogue cards, not as a universal measure of OCR usefulness.

Configs

• metrics: corpus-level CER/WER with edit-count numerators and denominators
• metric_details: per-card CER/WER and edit counts for all 8,000 model-card evaluations
• metric_metadata: source dataset, scoring mode, sample parameters, and timestamp

Generated with ocr-bench score.