ranjanhr1/nayanabench-rendered

NayanaBench Rendered Dataset

Dataset Description

This dataset contains 2D rendered text images for document analysis and OCR tasks across 22 languages. Each language split contains document images with text rendered in bounding boxes using authentic fonts.

Dataset Statistics

• Total Samples: 4,400
• Languages: 22
• Splits: 22 language-based splits

Language Splits

Language	Code	Samples
EN	en	200
HI	hi	200
KN	kn	200
TA	ta	200
TE	te	200
MR	mr	200
PA	pa	200
BN	bn	200
OR	or	200
ML	ml	200
GU	gu	200
SA	sa	200
JA	ja	200
KO	ko	200
ZH	zh	200
DE	de	200
FR	fr	200
IT	it	200
RU	ru	200
AR	ar	200
ES	es	200
TH	th	200

Dataset Structure

Data Fields

• image: The rendered document image
• image_id: Unique identifier for the sample
• language: Language code (ISO 639-1)
• font_used: Font file used for rendering
• num_regions: Number of text regions in the image
• regions: JSON string containing region metadata (bounding boxes, text content)

Data Splits

This dataset has 22 splits, one per language. Each split can be loaded separately:

from datasets import load_dataset

# Load a specific language
dataset = load_dataset("ranjanhr1/nayana_rendered", split="en")

# Load multiple languages
dataset_dict = load_dataset("ranjanhr1/nayana_rendered")

# Load specific languages
en_data = load_dataset("ranjanhr1/nayana_rendered", split="en")
hi_data = load_dataset("ranjanhr1/nayana_rendered", split="hi")

Dataset Creation

This dataset was created by rendering text onto document images using authentic fonts for each language. The rendering process uses:

• Tight canvas rendering (exact bounding box size)
• Optimal font size calculation
• Multi-line text wrapping
• Centered text alignment

Source Dataset

Based on Nayana-cognitivelab/NayanaBench

Citation

If you use this dataset, please cite:

@dataset{nayana_rendered,
  title={NayanaBench Rendered Dataset},
  author={Your Name},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/ranjanhr1/nayana_rendered}
}

License

Please refer to the original NayanaBench dataset for licensing information.