Omarrran/Persian_Pixel

Persian Pixel

Persian Pixel is a synthetic optical character recognition (OCR) dataset for Persian / Farsi (fa), in which Unicode text is rendered to images 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.

Authors: Haq Nawaz Malik · PouriaMahdi84

Sentences

Paragraphs

Pages

Dataset summary

Property	Value
Language	Persian / Farsi (fa)
Script	Perso-Arabic
Modality	Image → Text (OCR)
Source	Synthetic (font-rendered)
Total rows (full)	343,246
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 three.

Config	Description	Rows	Typical width
sentence	Short line / sentence-level crops	251,000	640 px
paragraph	Multi-line paragraph blocks	110,138	512 px
page	Page-level renders	31,108	768 px
full	Concatenation of all three configs above	343,246	—

Verified from Parquet shard metadata on 2026-06-25 05:34:30 UTC (320 Parquet files, ~31.8 GB).

Page-config statistics

Metric	Min	Max	Mean
Height (px)	38	1100	597.92
Text length (chars)	13	1896	824.22
Line count	2	75	7.88

Data fields

Every row shares the same schema across all configs:

Field	Type	Description
bytes	bytes	Encoded image bytes
path	string	Image path/name stored with the image bytes
text	string	Exact Persian Unicode OCR target / transcription
sample_type	string	One of sentence, paragraph, or page
source_run_id	string	Internal generation / publish run identifier
image_path	string	Original relative image path
width	int	Rendered image width in pixels
height	int	Rendered image height in pixels
text_chars	int	Number of Unicode characters in text
line_count	int	Number of text lines in the label

Usage

Install the dependency:

pip install datasets pillow

Load a single config:

from datasets import load_dataset

train = load_dataset("Omarrran/Persian_Pixel", "sentence", split="train")  # or "paragraph" / "page"

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

Load the combined config:

from datasets import load_dataset

full = load_dataset("Omarrran/Persian_Pixel", "full", split="train")
print(len(full))  # 343246

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

from datasets import load_dataset

ds = load_dataset(
    "Omarrran/Persian_Pixel",
    name="page",
    split="train",
    streaming=True,
)

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

If your loader sees raw bytes and path columns rather than a decoded image object, reconstruct a PIL image:

from io import BytesIO
from PIL import Image

row = train[0]
img = Image.open(BytesIO(row["bytes"]))
print(img.size, row["text"])

Loading each config

from datasets import load_dataset

sentence  = load_dataset("Omarrran/Persian_Pixel", "sentence",  split="train")
paragraph = load_dataset("Omarrran/Persian_Pixel", "paragraph", split="train")
page      = load_dataset("Omarrran/Persian_Pixel", "page",      split="train")
full      = load_dataset("Omarrran/Persian_Pixel", "full",      split="train")

Recommended training split strategy

The repo publishes data as train shards. For model development, create your own deterministic split:

ds = load_dataset("Omarrran/Persian_Pixel", "sentence", split="train")
split = ds.train_test_split(test_size=0.02, seed=42)
train_ds, val_ds = split["train"], split["test"]

For page-level OCR, keep validation smaller and representative because page images are much heavier than sentence crops.

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 Perso-Arabic script.
• Benchmarking Persian text recognition under controlled synthetic conditions and variable line counts.

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.
• Synthetic rendering can introduce artifacts, unusual line wrapping, punctuation variation, or source-text noise.
• 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

@dataset{persian_pixel_2026,
  title     = {Persian Pixel: A Synthetic OCR Dataset for Persian/Farsi},
  author    = { Haq Nawaz Malik, Pouria Mahdi},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/Omarrran/Persian_Pixel}
}

Repository

• Dataset repo: https://huggingface.co/datasets/Omarrran/Persian_Pixel
• Format: Parquet shards
• Primary task: Persian OCR / image-to-text