Add CER/WER ground-truth scoring for Kat57

[tadad]

Results

We scored all 16 OCR models on the same deterministic 500-card Kat57 sample (8,000 model-card evaluations, zero failed outputs). This benchmark uses CER/WER only—no VLM judge or ELO.

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

Lower is better. These are strict sequential comparisons with the PAGE XML references; nonlinear reading order, structured Markdown/LaTeX, imperfect references, and repetition loops all affect the scores. The published results card contains distributional statistics, per-card measurements, examples, and full limitations.

Reference sample · OCR outputs · CER/WER results · Complete Kat57 conversion

Code added

• ocr-bench score: independently computes corpus and per-card CER/WER from a reference column, with no model or judge calls.
• Metric publishing: writes aggregate metrics, per-card metric_details, and run metric_metadata configs to a Hugging Face dataset.
• Kat57 conversion: converts Lund University Library's 10,695 paired scans/PAGE XML files into bounded-memory Parquet shards and supports deterministic samples.

Related to #3.