esengul3/turkish-word-ocr

Turkish Word OCR (Synthetic)

Synthetic text-image dataset for Turkish OCR / text recognition: rendered word images paired with their text labels.

• Total images: 250,000 (train 225,000 · val 12,500 · test 12,500)
• Vocabulary: ~89,000 unique Turkish words
• Character set: 58 chars — 29 uppercase + 29 lowercase Turkish letters, including ı i İ ç ğ ö ş ü. No digits or punctuation.
• Images: grayscale, 48 px height, variable width.

Versions

version	splits	notes
v2.0 (latest)	word-disjoint	every word appears in only ONE by split — for example test/val words are unseen during training.
v1.0	random	a word may appear across train/val/test (only the image differs).

Usage

from datasets import load_dataset

#latest (word-disjoint splits)
ds = load_dataset("esengul3/turkish-word-ocr")

# pin a specific version
ds = load_dataset("esengul3/turkish-word-ocr", revision="v1.0") # random splits

ds["train"][0] 

Sources and credits

Vocabulary compiled from two public word lists:

• hermitdave — FrequencyWords (OpenSubtitles Turkish frequency list; content licensed CC BY-SA 4.0)
• mertemin — turkish-word-list (TDK dictionary words)

Only rendered images are distributed. Released under CC BY-SA 4.0, matching the share-alike terms of the upstream content.

Citation

If you use this dataset, please cite it:

@misc{turkish_word_ocr_synthetic,
  title  = {Turkish Word OCR (Synthetic)},
  author = {esengul3},
  year   = {2026},
  url    = {https://huggingface.co/datasets/esengul3/turkish-word-ocr}
}