Cognitive-Lab/NayanaOCRBench_Natural

NayanaOCRBench · Natural

Human-verified OCR benchmark on real documents in 22 languages. 10,533 pages of scanned, photographed and born-digital documents (government gazettes, newspapers, forms, books, reports) whose layout and text were machine-proposed and then corrected by human annotators, in OmniDocBench format.

👁️ Browse every model output on every page: https://huggingface.co/spaces/AdithyaSK/NayanaOCRBench

What is in a row

Column	Type	Meaning
image	Image	the document page
id	string	page id, e.g. kn_pdf_0003_page_00226
omnidocbench / omnidocbench_original / omnidocbench_norm1000	struct	OmniDocBench JSON: layout blocks with category, polygon, reading order and text (original pixels and 1000×1000 normalised)
md_norm1000 / html_norm1000	string	full-page ground truth in reading order, Markdown and HTML
layout_norm1000	list	{type, bbox} per block on a 1000×1000 grid
layout_content_norm1000	list	{type, bbox, content} per block

One config per language: English 502 · Hindi 500 · Bengali 467 · Marathi 515 · Gujarati 445 · Punjabi 463 · Odia 301 · Kannada 500 · Tamil 504 · Telugu 461 · Malayalam 492 · Sanskrit 425 · Chinese 526 · Japanese 502 · Korean 454 · German 496 · French 504 · Spanish 500 · Italian 494 · Russian 523 · Arabic 452 · Thai 507.

How it was built

1. Real pages collected per language and de-duplicated.
2. Layout proposed with MinerU, then corrected by annotators (block classes, boxes, reading order).
3. Text extracted per block by frontier VLMs, then reviewed line by line by human annotators.
4. Exported to OmniDocBench format so existing evaluation tooling works unchanged.

Zero-shot results (character accuracy = 100 − CER)

13 open vision-language and OCR models, one shared prompt, full-page plain-text OCR. Per-page predictions, scores and the evaluation harness are public in https://huggingface.co/buckets/Cognitive-Lab/nayanaocrbench-evals; the explorer Space shows the page, ground truth and each model's output side by side.

Model	Langs	Avg (all, missing = 0)	Avg (evaluated)	Indic	East Asian	European	Other
dots.mocr	22/22	71.9	71.9	63.9	69.7	85.3	79.2
Qwen3.5-9B	22/22	66.9	66.9	49.1	75.2	90.9	80.3
Qwen3-VL-8B	22/22	64.7	64.7	44.0	75.1	90.7	85.1
Qwen3.5-4B	22/22	62.2	62.2	45.2	67.6	88.5	68.4
Gemma 4 31B	21/22	61.1	64.0	52.5	61.4	81.4	73.6
Gemma 4 E4B	22/22	59.5	59.5	42.1	62.2	84.9	75.3
Gemma 3 27B	22/22	58.4	58.4	47.1	57.2	76.0	68.9
Qwen3.5-2B	22/22	57.0	57.0	36.5	70.7	86.4	60.9
DeepSeek-OCR	13/22	43.2	73.1	39.1	71.2	86.6	69.1
Gemma 4 E2B	22/22	41.0	41.0	24.0	44.1	67.6	50.3
GLM-OCR 1.3B	9/22	32.7	79.9	—	66.0	86.8	—
Llama 4 Scout 17B	8/22	30.4	83.5	71.5	—	87.1	80.5
GLM-4.6V-Flash	8/22	28.6	78.8	—	83.2	80.8	62.0

Per-language accuracy

Model	en	hi	bn	mr	gu	pa	or	kn	ta	te	ml	sa	zh	ja	ko	de	fr	es	it	ru	ar	th
dots.mocr	75.8	61.8	64.7	54.3	71.8	61.0	54.0	54.1	77.7	71.9	60.6	70.9	79.7	51.5	77.8	86.3	84.5	88.1	87.9	88.9	83.4	75.1
Qwen3.5-9B	82.3	76.1	68.8	62.1	55.0	72.0	3.0	9.5	71.6	43.8	12.1	66.1	87.3	55.8	82.4	93.0	90.2	91.5	93.8	94.3	78.6	81.9
Qwen3-VL-8B	80.0	73.3	74.3	68.4	28.6	66.5	5.4	5.9	60.7	17.0	13.6	70.0	88.1	54.4	82.9	92.0	91.2	92.8	94.0	94.3	86.8	83.4
Qwen3.5-4B	77.0	66.2	61.5	55.3	45.2	68.0	4.1	11.6	70.6	38.2	18.6	57.5	74.4	49.0	79.4	90.5	88.6	91.0	91.9	92.2	67.2	69.6
Gemma 4 31B	74.5	54.9	45.7	51.2	63.6	—	45.4	40.8	73.7	58.7	37.7	53.4	75.9	40.5	67.8	82.4	82.7	83.0	85.3	80.3	72.4	74.7
Gemma 4 E4B	73.8	57.4	47.2	48.6	51.1	39.4	11.9	25.2	62.6	44.2	23.9	52.0	71.8	45.7	69.1	86.1	84.6	88.4	89.6	86.8	81.4	69.1
Gemma 3 27B	66.4	51.4	40.7	47.7	57.3	46.9	39.7	30.9	67.9	54.9	28.0	53.3	69.0	40.6	62.1	77.4	75.0	80.5	80.2	76.8	77.0	60.8
Qwen3.5-2B	75.9	56.4	55.0	43.5	34.7	58.2	1.0	3.5	63.7	32.3	4.8	48.2	84.4	51.8	75.9	88.5	88.2	86.6	90.3	88.8	56.4	65.4
DeepSeek-OCR	76.4	38.1	40.0	—	—	—	—	—	—	—	—	—	85.3	49.7	78.5	87.6	87.0	90.5	91.1	87.1	79.0	59.3
Gemma 4 E2B	60.0	37.1	20.7	33.0	17.3	13.7	5.1	10.7	46.8	30.5	9.2	40.2	50.7	33.5	48.1	66.7	67.5	73.1	72.6	65.8	60.8	39.7
GLM-OCR 1.3B	76.4	—	—	—	—	—	—	—	—	—	—	—	80.5	51.2	66.3	90.0	89.8	91.7	92.0	80.9	—	—
Llama 4 Scout 17B	75.7	71.5	—	—	—	—	—	—	—	—	—	—	—	—	—	89.9	86.1	92.9	90.8	—	77.4	83.7
GLM-4.6V-Flash	65.8	—	—	—	—	—	—	—	—	—	—	—	83.2	—	—	82.4	80.2	82.5	86.5	87.5	62.0	—

Grey / missing cells: the model card does not claim the language, so it was not run. "Avg (all)" counts those as 0, which is what a user of a general multilingual model experiences; "Avg (evaluated)" averages only the languages that were run.

Finding. Latin, Cyrillic and CJK are largely solved. Kannada, Odia, Malayalam and Telugu are not: the strongest general-purpose family scores above 90% on European scripts and under 20% on those four at every model size. The gap is a data gap, which is what the NayanaOCR corpora exist to close.

Related

• Companion held-out synthetic set: Cognitive-Lab/NayanaOCRBench_Synthetic
• Training corpus (public, 1M pages): Cognitive-Lab/NayanaOCR_Corpus_2025
• Retrieval benchmark: Nayana-IR Bench

License and citation

CC BY-NC 4.0. Free for research and education with attribution; commercial use: contact@cognitivelab.in.

@inproceedings{kolavi2025nayanaocr,
  title     = {Nayana {OCR}: A Scalable Framework for Document {OCR} in Low-Resource Languages},
  author    = {Kolavi, Adithya S. and Samarth, P. and Jain, Vyoman},
  booktitle = {Proceedings of the 1st Workshop on Language Models for Underserved Communities (LM4UC), NAACL},
  year      = {2025}, pages = {86--103}, url = {https://aclanthology.org/2025.lm4uc-1.11/}
}
@inproceedings{kolavi2025nayana,
  title     = {Nayana: A Foundation for Document-Centric Vision-Language Models via Multi-Task, Multimodal, and Multilingual Data Synthesis},
  author    = {Kolavi, Adithya S. and Samarth, P. and Jain, Vyoman},
  booktitle = {ICCV Workshops (CV4DC)}, year = {2025}, pages = {1678--1687}
}

Part of the Nayana initiative by Cognitive Lab, a 2025 Meta Llama Impact Grant recipient.