Kurdish Text Recognition (ktr)

2.5M word images paired with their Central Kurdish transcription — a text-recognition (OCR) training set for Sorani, with the image and its ground-truth word/line text side by side.

At a glance

Rows	2,499,792 (single train split)
Columns	image (JPEG/PNG bytes + path), text (string)
Parquet on disk	7.54 GB across 12 shards (5.60 GB uncompressed in memory)
Row groups	209 per shard (1,000 rows per group)
Script	Arabic script, 100% of rows
Task	text recognition / OCR fine-tuning

Field profile

Measured over a 1,000-row sample (shard 2, first row group):

Column	Mean	Median	Max
text	6.5 chars	6	17

text holds short units — mostly single words, often with the definite/plural suffixes attached (e.g. ودیارینەکراو, مێژووی). Images are image crops stored as bytes in the parquet file (no external image folder to download).

Usage

from datasets import load_dataset

ds = load_dataset("razhan/ktr", split="train")
ex = ds[0]
ex["image"].show()      # PIL image
print(ex["text"])

For recognition training, resize to a fixed height, keep the aspect ratio, and use the short text as the label; for a line-level OCR you would need to concatenate neighbouring crops yourself.

How it was built

The dataset was generated with kurdish-ocr: Kurdish text is rendered with Kurdish-supporting fonts and then cropped/augmented into the word images stored here. Also see razhan/kteb and razhan/kteb-dataset for the underlying book text used as the source.

Known limitations

• Synthetic rendering: fonts, backgrounds and rendering artefacts are those of the generator, so real scanned-book pages will differ (paper texture, skew, bleed).
• Labels are single short units, not full lines or paragraphs.
• 2.5M rows × ~2.2 KB per image make the parquet files large; use streaming=True or download selected shards only.

Related

• razhan/kteb — 1,033 OCR'd books (source text).
• razhan/kteb-dataset — 10,124 books with page and word counts.