Omarrran/Koshur_Pixel

Koshur Pixel

Koshur Pixel is a synthetic optical character recognition (OCR) dataset for Kashmiri (ks), in which Unicode text is rendered to images using Nastaliq / Naksh-style fonts and paired with its exact transcription. It is built for OCR recognition, image-to-text modeling, fine-tuning, and evaluation workflows that need clean, controllable image/label pairs at scale.

Because the text is rendered programmatically, every image ships with a perfectly aligned ground-truth label — making the dataset well suited for pretraining, curriculum learning, and isolating model behavior before introducing noisy real-world scans.

Sentences

Paragraphs

Pages

Dataset summary

Property	Value
Language	Kashmiri (ks)
Script	Perso-Arabic (Nastaliq / Naksh)
Modality	Image → Text (OCR)
Source	Synthetic (font-rendered)
Fonts	Gulmarg Nastaleeq, Afan Koshur Naksh
Total rows (full)	613,078.
License	CC BY-ND 4.0

Configurations and sizes

The dataset is organized by render granularity. Each granularity is exposed as its own config, plus a full config that virtually concatenates all four.

Config	Description	Rows
mixed	Mixed-granularity segments (word, sentence/line, paragraph, and page level)	119,114
sentence_line	Sentence / line-level renders	378,534
paragraph	Longer paragraph-level renders	72,851
page	Page-level renders	42,679
full	Concatenation of all four configs above	613,078.

Data fields

Every row shares the same schema across all configs:

Field	Type	Description
image	Image	Rendered OCR image
text	string	Unicode text label (exact transcription)
sample_type	string	One of mixed, sentence_line, paragraph, page
font_family	string	Renderer font (Gulmarg Nastaleeq or Afan Koshur Naksh)
source_run_id	string	Internal generation run identifier
image_path	string	Source image path in the original generator volume
width	int	Image width in pixels
height	int	Image height in pixels
text_chars	int	Label length in Unicode codepoints
line_count	int	Number of lines in the label

A per-row image path is also retained inside the image struct metadata for traceability. All configs use parquet row-group sizing with small groups and page-index metadata enabled for Dataset Viewer compatibility.

Usage

Install the dependency:

pip install datasets

Load a single config:

from datasets import load_dataset

train = load_dataset("Omarrran/Koshur_Pixel", "sentence_line", split="train")       # you can also use "paragraph" or "page" instead of "sentences"

print(train.column_names)
print(train[0]["text"])
print(train[0]["image"].size)

Load the combined config:

from datasets import load_dataset

full = load_dataset("Omarrran/Koshur_Pixel", "full", split="train")
print(len(full))  # 280657

Stream a config (recommended for the larger paragraph / page renders):

from datasets import load_dataset

ds = load_dataset(
    "Omarrran/Koshur_Pixel",
    name="paragraph",
    split="train",
    streaming=True,
)

for sample in ds:
    img = sample["image"]
    txt = sample["text"]
    # feed into OCR preprocessing / tokenizer pipeline

Inspect a row:

print(mixed[0]["text"], mixed[0]["font_family"], mixed[0]["sample_type"])

Loading each config

from datasets import load_dataset

mixed         = load_dataset("Omarrran/Koshur_Pixel", "mixed",         split="train")
sentence_line = load_dataset("Omarrran/Koshur_Pixel", "sentence_line", split="train")
paragraph     = load_dataset("Omarrran/Koshur_Pixel", "paragraph",     split="train")
page          = load_dataset("Omarrran/Koshur_Pixel", "page",          split="train")
full          = load_dataset("Omarrran/Koshur_Pixel", "full",          split="train")

Recommended use cases

• Fine-tuning image-to-text and OCR models such as TrOCR, Donut, BLIP-2, and PaliGemma-style decoders.
• Pretraining or curriculum stages before introducing noisy scanned or photographed OCR corpora.
• Evaluating normalization, decoding, and post-correction pipelines on Nastaliq scripts.
• Prompting and benchmarking Kashmiri text understanding under controlled synthetic conditions.

Limitations and considerations

• This is synthetic OCR data. Performance on font-rendered text does not guarantee performance on real scans, photographs, or handwriting — validate on in-domain data before any production deployment.
• Only two fonts are represented in this release (Gulmarg Nastaleeq, Afan Koshur Naksh), so visual diversity is narrower than real-world print.
• The paragraph and page configs contain long, multi-line labels and large images; always consume them through the datasets library rather than raw parsing.

License

Released under Creative Commons Attribution-NoDerivatives 4.0 International (CC BY-ND 4.0). You may copy and redistribute the dataset for any purpose with attribution, but you may not distribute modified versions.

Citation

If you use this dataset, please cite it as a synthetic Kashmiri OCR resource and include the exact Hugging Face revision/SHA used:

@misc{koshur_pixel,
  title        = {Koshur Pixel: A Synthetic OCR Dataset for Kashmiri},
  author       = {Malik, Haq Nawaz , Nahfid Nissar, Faizan Iqbal},
  howpublished = {\url{https://arxiv.org/abs/2606.23144}},
  note         = {Synthetic Nastaliq/Naksh-rendered image–text pairs. Cite the exact revision/SHA used.},
  year         = {2026}
}
``