Omar-2002/kz_benchmark_data

Kazakh Real-Document OCR Benchmark

A test-only benchmark for full-page OCR on real Kazakh documents. It has two subsets:

Subset	Pages	Size	Text characters	Median chars / page	Max chars / page
handwritten	150	100 MB	89,671	542	1,665
non_handwritten	1,429	321 MB	836,742	428	14,336¹
Total	1,579	~421 MB	926,413

¹ This page's reference is mostly underscores (13,051 of 14,336 characters) copied from blank form lines. See "Known limitations".

Each page has one plain-text transcription of the whole page. There are no bounding boxes, line-level crops or layout annotations.

Contents

handwritten: 150 pages

Photos and scans of handwritten pages, such as school notebooks, homework, essays, poems and lecture notes. All pages are Kazakh: every page contains Kazakh-specific letters (ә, ғ, қ, ң, ө, ұ, ү, һ, і). A few pages also contain digits, arithmetic and simple formulas, and two are lecture notes that quote code (MPI/OpenMP).

non_handwritten: 1,429 pages

Printed and digital Kazakh documents, including official letters and certificates («АНЫҚТАМА»), forms and templates, school worksheets and textbook pages, tables, and screenshots. 1,419 of the 1,429 pages contain Kazakh-specific letters. Some pages also include Russian or English text, such as UI labels.

Images

• All files use the .jpg extension. In the non-handwritten subset, 20 of them are actually encoded in another format: 15 WebP, 3 PNG, 1 BMP and 1 MPO. Open images with a library that detects the format from the file contents, such as Pillow, rather than trusting the extension.
• Resolution varies widely (sizes sorted by pixel count):
  • handwritten: median 2480 × 3507 px, smallest 240 × 427 px, largest 3120 × 4160 px
  • non-handwritten: median 1152 × 648 px, smallest 514 × 122 px, largest 5472 × 3648 px
• Colour modes in the non-handwritten subset: 1,401 RGB, 24 grayscale, 2 CMYK, 2 RGBA. Convert to RGB before inference.

File layout

real_handwritten_data/
  data.csv          # columns: image_path, text
  images/0001.jpg … 0150.jpg
real_non_handwritten_data/
  data.csv          # columns: image_path, text
  images/0001.jpg … 1429.jpg

data.csv is UTF-8 with every field quoted. image_path is relative to the subset folder, for example images/0001.jpg. text is the reference transcription and keeps the page's line breaks. A few references are a single line (handwritten page 0042 and 35 short non-handwritten pages). Every image has exactly one row, and no row is empty.

Usage

The CSVs store image paths rather than embedded images, so the most reliable way to load the benchmark is to download the repository and read the files directly:

import csv, os
from huggingface_hub import snapshot_download
from PIL import Image

root = snapshot_download("Omar-2002/kz_benchmark_data", repo_type="dataset")

def load(subset):  # "real_handwritten_data" or "real_non_handwritten_data"
    base = os.path.join(root, subset)
    with open(os.path.join(base, "data.csv"), encoding="utf-8") as f:
        for row in csv.DictReader(f):
            yield Image.open(os.path.join(base, row["image_path"])).convert("RGB"), row["text"]

for image, reference in load("real_handwritten_data"):
    ...

load_dataset("Omar-2002/kz_benchmark_data", "handwritten") also works, but it returns image_path as a string, not as a decoded image.

Evaluation

We recommend reporting CER and WER (character and word error rate, computed from Levenshtein distance) separately for each subset, at two levels:

• corpus-level: total edits divided by total reference length
• per-page mean

Results depend heavily on scoring choices, so state them alongside your numbers:

• Text normalisation. Say whether whitespace is collapsed, whether punctuation and case are kept, and how Markdown/HTML in model output is stripped. Models that emit Markdown tables or LaTeX are penalised under a plain-text reference unless their output is converted first.
• Runaway generations. VLM-based OCR models sometimes loop until they hit the token limit, which can produce per-page error rates of several thousand percent. A few such pages can dominate an uncapped average. We suggest capping per-page rates at 100% and reporting the number of runaway pages as a separate column.
• Confidence intervals. The handwritten subset is small (150 pages), so differences of a few CER points between models may not be significant. A paired bootstrap over pages is a simple way to check.

Known limitations

• Reference quality is uneven. The references were produced with the help of vision-language models and have not all been verified by a human (see Source data). An automated scan of the labels found:

  • handwritten: no Markdown, HTML, meta-commentary or repetition loops. Two pages contain Greek letters (π, α) that are legitimate symbols in formulas.
  • non_handwritten:
    • 46 pages contain English annotations written by a VLM rather than text printed on the page. Most are seal or stamp markers, such as [Seal text] and [Text on the seal], and others describe the layout, such as **Left Side:**, **Image 1 (Left):** and [Modal window].
    • 30 pages contain Markdown tables, and 43 contain Markdown **bold**.
    • 5 pages contain HTML <br> tags.
    • 7 pages have long runs of filler characters (underscores, dashes, non-breaking spaces or table separators) copied from form lines. These pages are 0031, 0093, 0206, 0917, 1030, 1046 and 1080.
    • 3 pairs of pages have identical references: 0003/0007, 0368/0390 and 0749/0753. The image files differ, so these are probably the same page photographed twice.
    • 5 pages contain non-Cyrillic, non-Latin characters (Greek, Arabic, CJK). Some of these are legitimate symbols.

  Strip Markdown/HTML from both the reference and the prediction before scoring, or score tables separately. Treat scores as approximate, especially small differences between strong models.

• Reading order is ambiguous on multi-column pages, tables, two-page spreads and screenshots. A model can read every word correctly and still get a high CER/WER if it chooses a different order.

• Possible overlap with web data. The pages were collected from the public web, so some of them may also appear in web-scraped training corpora.

Source data

• Handwritten:
  • pages 0001–0150: collected by the author from public web image search.
• Non-handwritten: collected by the author from public web image search using Kazakh queries for document types (certificates, letters, forms, worksheets, screenshots), then filtered for readable pages.
• How the references were produced:
  • handwritten, pages 0001–0030: drafted by a vision-language model, then corrected by hand.
  • handwritten, pages 0031–0150: first-pass transcriptions generated by a vision-language model were re-read against the image by Gemini, which corrected misread words and restored the page's line breaks (gemini-3.5-flash-lite for 95 pages, gemini-3.6-flash for 12, gemini-3.7-flash for 12). The edits are small: the corrected text is 94–106% of the first pass's length. Page 0042 still has its first-pass transcription, which is a single line without line breaks. The corrected transcriptions were then checked by a human. Because of time constraints, some pages may still contain small errors.
  • non-handwritten: generated by a vision-language model and not verified by a human. The Markdown tables and English annotations described under "Known limitations" come from this.

Personal and sensitive information

Some pages are official documents, certificates, forms or student work, and may contain names, signatures, ID numbers, addresses, e-mail addresses or phone numbers. Personal identification numbers (IINs) were removed from the transcriptions and blacked out in the images. The images were not otherwise anonymised. If you find a page that should not be public, please open a discussion on this repository and it will be removed.

Licensing

The annotations (data.csv files), the benchmark selection and this card are released under CC BY 4.0.

The images were collected from the web, and their copyright belongs to the original publishers or authors. CC BY 4.0 does not cover them. They are shared here only for non-commercial research evaluation. Rights holders can ask for removal through the repository's discussion tab.

Citation

If you use this benchmark, please cite it:

@misc{polat2026kazakhocr,
  title  = {Kazakh Real-Document OCR Benchmark},
  author = {Omar Polat},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Omar-2002/kz_benchmark_data}
}