kaa-ml/kaa-cyrl-ocr-synthetic

Karakalpak Cyrillic OCR — synthetic text lines

Line images of Karakalpak (Cyrillic) text for fine-tuning OCR models such as Tesseract LSTM. Each row is one rendered text line (image) and its exact transcription (text).

Config	Lines	How it was made
synthetic_news	30,000	14 system fonts (Arial, Times New Roman, PT Serif/Sans, Courier, Georgia, …), paper tint, noise, blur, ink spread, skew, down/up-scaling. Half of the lines are sampled to contain ў ҳ ә ғ қ ң ө ү.
synthetic_natural	8,000	Same rendering, natural letter distribution.
synthetic_scan	8,000	Serif fonts closer to printed books (Old Standard TT, Literata, Noto Serif, PT Serif, Tinos, Times New Roman, Georgia) with Augraphy scan simulation (ink bleed, low ink, paper texture, bleed-through, bad photocopy, JPEG) followed by blur/tint/noise.

Only fonts that cover the full Karakalpak Cyrillic alphabet were used.

Usage

from datasets import load_dataset

lines = load_dataset("kaa-ml/kaa-cyrl-ocr-synthetic", "synthetic_scan", split="train")
print(lines[0]["text"])
lines[0]["image"]

Export to tesstrain ground truth (*.png + *.gt.txt):

from pathlib import Path

out = Path("gt"); out.mkdir(exist_ok=True)
for i, row in enumerate(lines):
    row["image"].save(out / f"kaa_{i:06d}.png")
    (out / f"kaa_{i:06d}.gt.txt").write_text(row["text"] + "\n")

Then make training MODEL_NAME=kaa_cyrl START_MODEL=kaz TESSDATA=<tessdata_best> GROUND_TRUTH_DIR=gt.

Source text

Sentences come from Karakalpak (Cyrillic) articles of the Karakalpakstan news agency, via kaa-ml/kknews-dataset (kknews.uz, CC-BY-4.0), wrapped into 18–60 character lines.

Results

A Tesseract model fine-tuned from kaz on these lines plus 182 real scanned lines reads Karakalpak-specific letters far better than the Kazakh model (letter recall 88.8% → 92.3% on 10 held-out scanned book pages). On the same pages its overall character error rate (12.87%) does not yet beat the Kazakh model (12.64%). Training on synthetic_scan instead of synthetic_natural gave 13.08% — heavier scan simulation did not help; real scanned lines matter most.

License

CC-BY-4.0, following the source text. Attribute kknews.uz / Qaraqalpaqstan Xabar Agentligi.