Cognitive-Lab/NayanaOCRBench_Synthetic

NayanaOCRBench · Synthetic

Held-out evaluation split of the NayanaOCR 2026 pipeline, 22 languages, ~770 pages per language. Every page is the same source document re-typeset in each language with the layout preserved, so the set is parallel across languages and carries the full annotation stack of the training corpus: layout, reading-order OCR ground truth, translated tables, VQA pairs.

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

What is in a row

Same 14 columns as the page-level NayanaOCR corpus: image, font_used, page_size, original_id (shared across all 22 language configs), rendered_layout (per region: type, coordinates, original and translated text, fit ratio), translated_tables (HTML), vqa_translated, omnidocbench / omnidocbench_original / omnidocbench_norm1000, html_norm1000, md_norm1000, layout_content_norm1000, layout_norm1000.

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

Same protocol and models as the Natural benchmark. English is excluded from the tables below while its ground truth is being revised (text leaked into the latex field on some text blocks); all other languages are unaffected. Predictions and scores: https://huggingface.co/buckets/Cognitive-Lab/nayanaocrbench-evals.

Model	Langs	Avg (all, missing = 0)	Avg (evaluated)	Indic	East Asian	European	Other
dots.mocr	21/21	61.8	61.8	54.6	67.3	83.9	38.2
Qwen3.5-4B	21/21	59.5	59.5	43.1	76.1	89.6	49.3
Qwen3.5-9B	21/21	55.0	55.0	40.8	66.3	78.4	57.3
Gemma 4 E4B	21/21	52.7	52.7	34.9	66.3	83.9	52.7
Gemma 3 27B	21/21	50.3	50.3	39.4	56.4	71.4	48.1
Qwen3-VL-8B	21/21	50.2	50.2	33.6	66.1	78.3	47.1
Gemma 4 31B	21/21	49.7	49.7	37.8	62.5	69.3	47.3
Qwen3.5-2B	21/21	46.6	46.6	28.6	63.5	81.1	34.4
Gemma 4 E2B	21/21	36.4	36.4	19.0	48.3	68.7	33.9
DeepSeek-OCR	12/21	31.8	55.7	27.7	60.1	74.1	31.4
GLM-OCR 1.3B	8/21	30.5	80.0	—	68.2	87.1	—
GLM-4.6V-Flash	9/21	25.3	58.9	—	49.4	73.4	14.9
Llama 4 Scout 17B	7/21	22.1	66.4	62.4	—	75.0	51.3

Per-language accuracy

Model	hi	bn	mr	gu	pa	or	kn	ta	te	ml	sa	zh	ja	ko	de	fr	es	it	ru	ar	th
dots.mocr	63.8	60.1	63.0	65.5	55.6	33.4	47.8	50.6	63.4	47.4	50.3	69.5	54.7	77.7	87.6	86.7	86.0	86.3	72.9	38.9	37.5
Qwen3.5-4B	72.9	65.7	66.1	48.5	63.2	2.8	4.7	61.7	18.6	10.5	58.9	73.4	67.2	87.6	89.5	91.0	89.6	89.3	88.6	39.3	59.3
Qwen3.5-9B	66.9	61.9	62.2	42.9	56.2	4.0	3.0	49.6	16.2	7.0	78.8	63.3	58.9	76.6	78.3	79.2	78.2	76.7	79.5	38.8	75.7
Gemma 4 E4B	62.6	40.8	53.7	44.6	29.1	8.8	18.0	38.9	24.2	19.1	44.0	61.7	59.8	77.3	84.2	85.4	84.3	83.8	81.6	55.4	50.1
Gemma 3 27B	56.3	43.1	51.8	46.6	34.8	26.2	22.9	46.4	33.1	23.6	49.1	54.2	51.3	63.7	71.3	72.4	72.6	71.0	69.7	53.4	42.8
Qwen3-VL-8B	67.8	62.2	63.8	13.9	37.1	6.8	2.6	40.2	8.9	7.7	59.0	63.0	58.6	76.8	78.7	79.3	78.0	76.5	79.0	49.6	44.6
Gemma 4 31B	52.2	35.2	51.4	41.2	29.3	25.2	25.7	50.7	35.3	27.0	42.8	57.5	58.1	71.9	67.2	69.2	68.2	66.7	74.9	47.9	46.7
Qwen3.5-2B	53.3	43.1	53.3	30.4	37.2	3.5	1.6	36.7	10.2	3.3	41.4	61.0	56.8	72.7	76.4	90.5	76.5	75.2	87.1	28.8	39.9
Gemma 4 E2B	39.6	22.6	35.3	15.1	9.6	6.3	5.9	20.1	14.4	6.5	33.1	44.6	46.1	54.1	67.0	71.9	71.4	69.7	63.3	36.9	30.9
DeepSeek-OCR	34.3	21.0	—	—	—	—	—	—	—	—	—	58.4	53.9	67.9	74.6	76.5	74.5	73.4	71.4	32.3	30.5
GLM-OCR 1.3B	—	—	—	—	—	—	—	—	—	—	—	73.4	64.3	67.0	88.2	90.6	88.7	88.5	79.4	—	—
GLM-4.6V-Flash	—	—	—	—	—	—	—	—	—	—	—	58.2	49.8	40.3	74.4	75.7	74.6	73.7	68.8	14.9	—
Llama 4 Scout 17B	62.4	—	—	—	—	—	—	—	—	—	—	—	—	—	75.2	75.7	75.0	74.1	—	51.7	50.9

Synthetic pages are typeset with a wide rotation of fonts per script and are, on average, harder for current models than the natural pages; use the two sets together.

Related

• Human-verified benchmark on real documents: Cognitive-Lab/NayanaOCRBench_Natural
• Training corpus (public, 1M pages): Cognitive-Lab/NayanaOCR_Corpus_2025

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.