pndaza/burmese-ocr-1m

Burmese OCR 1M

A dataset of 897,883 rendered line images for Burmese/Pali text recognition (OCR). Images are grayscale/bi-level line crops paired with their transcription. General-purpose: usable with any OCR engine (Tesseract, Kraken, Calamari, or custom HTR models).

Splits

split	rows	files
train	718,515	5 parquet shards
validation	89,625	1 parquet shard
test	89,743	1 parquet shard
total	897,883	7

Columns

column	type	description
image	image	PNG line image (struct: bytes + path)
text	string	Transcription (Burmese/Pali)
image_id	string	Unique zero-padded id, e.g. 000000001

Usage

With datasets

from datasets import load_dataset

ds = load_dataset("pndaza/burmese-ocr-1m")
# ds["train"][0]
# → {"image": <PIL.Image>, "text": "ဧကကနိပါတအဋ္ဌကထာနိဿယ", "image_id": "000000001"}

# The image column returns a PIL.Image directly:
img = ds["train"][0]["image"]
img.show()

Extract to files (for Tesseract / Kraken training)

from datasets import load_dataset
from pathlib import Path

ds = load_dataset("pndaza/burmese-ocr-1m", split="train")
out = Path("train"); out.mkdir(exist_ok=True)
for row in ds:
    (out / f"{row['image_id']}.png").write_bytes(row["image"]["bytes"])
    (out / f"{row['image_id']}.gt.txt").write_text(row["text"], encoding="utf-8")

Notes

• The companion dataset pndaza/burmese-kraken-1m ships the same 897K rows in Kraken's binary PyArrow IPC format for direct use with ketos train -f binary.
• Line images are rendered, not scanned. The degraded subset is intended to bridge the domain gap to real scans; it does not replace genuine scan data.