KADİM Turkish OCR Dataset (kadim-ocr-data)

1,000 synthetic Turkish document page images with ground-truth text, at four levels of scan degradation. It is the training and evaluation data of furkancmc/kadim-ocr-lora, the reader (OCR) component of KADİM — Kamu Dilekçe ve İnceleme Mekanizması, built for the TEKNOFEST 2026 Language Agents Competition (Scenario 1: Public Documents and Correspondence).

• Project code: github.com/furkancmc/teknofest-kadim-kamu-evrak-ajani (QWEN_VL_ocr/)
• Related: kadim-analiz-data, kadim-yazi-data

What is in it

Each example is a PNG page image plus the exact text that was printed on it.

Split	Images
train	800
validation	100
test	100

The file name prefix tells the degradation group (250 images each across the whole set):

Group	Scan quality
A	Clean print, no degradation
B	Slight blur, small contrast shift
C	Clear blur, reduced contrast, noise
D	Low resolution, heavy noise, skew

Fields in metadata.jsonl (per split): file_name (image), text (ground-truth transcription) and group (A–D).

from datasets import load_dataset
ds = load_dataset("furkancmc/kadim-ocr-data")
ex = ds["train"][0]
print(ex["group"], ex["text"][:200])
ex["image"].show()

How it was made

1. Plain texts were taken from the open erdem-erdem/Turkish-Law-Documents-700k-clustered corpus, which contains decisions published by the Turkish Court of Cassation (Yargıtay) and the Council of State (Danıştay). The courts publish these decisions with party names masked; a random sample of the source texts showed no ID numbers, phone numbers, e-mails or IBANs.
2. generate_dataset_1000.py (GitHub repo) typesets each text as a letterhead page and renders it to PNG.
3. The pages were then degraded to group A–D quality levels to imitate real scans.

The images are synthetic training pages, not scans of real official documents. The text is real, public, court-published language; the page layout, letterhead and degradations are generated.

Intended use

Fine-tuning and evaluating OCR / vision-language models on Turkish text, especially Turkish-specific characters (ı, ş, ğ, ö, ü, ç) under scan-like degradation. Reference results on the test split: kadim-ocr-lora lowers the character error rate from 0.4283 to 0.1076 and raises Turkish-character accuracy from 78.55% to 85.90% (base Qwen2.5-VL-7B-Instruct).

Limitations and ethics

• Synthetic degradations of typeset pages; real scans, handwriting, stamps and unusual layouts are not represented.
• The content is legal-decision language, which is narrower than everyday administrative correspondence.
• Not meant for creating documents that pass as genuine official papers.
• The source corpus has no stated license; the original texts are public court decisions published by Yargıtay and Danıştay (see their Yargıtay and Danıştay decision search systems). Check the source terms for your use case.

License

The images, generation code and annotations produced for this project are MIT-licensed. The underlying decision texts are public court publications whose original terms are not stated in the source corpus.

Citation

@misc{kadim2026,
  title  = {KADİM: Kamu Dilekçe ve İnceleme Mekanizması},
  author = {Camcıoğlu, Furkan},
  year   = {2026},
  url    = {https://github.com/furkancmc/teknofest-kadim-kamu-evrak-ajani}
}