Datasets:
|
Download README.md from moganai/TurkishOCR-Bench: direct link, hf CLI and curl.
- Browser
- Download file 8.42 kB
-
https://huggingface.co/datasets/moganai/TurkishOCR-Bench/resolve/main/README.md
- Command line
-
hf download hf://datasets/moganai/TurkishOCR-Bench/README.md
-
curl -L -o README.md https://huggingface.co/datasets/moganai/TurkishOCR-Bench/resolve/main/README.md
8.42 kB
| language: | |
| - tr | |
| license: other | |
| license_name: sayfa-basina-lisans | |
| pretty_name: TurkishOCR-Bench | |
| task_categories: | |
| - image-to-text | |
| tags: | |
| - ocr | |
| - document-understanding | |
| - benchmark | |
| - turkish | |
| size_categories: | |
| - n<1K | |
| configs: | |
| - config_name: default | |
| data_files: | |
| - split: test | |
| path: | |
| - metadata.jsonl | |
| - goruntu/** | |
| # TurkishOCR-Bench | |
| **[English](#english) · [Türkçe](#turkce)** | |
| <a id="english"></a> | |
| ## English | |
| TurkishOCR-Bench is the data of a page-level benchmark for Turkish document OCR: **488 page images and the ground-truth text of each page** (368 documents, 10 categories). | |
| The benchmark measures how accurately a model reads a page and how little manual correction its output needs. Errors specific to Turkish (ı/i, ş/s, İ/I, circumflexed letters, decimal comma, negation suffix) are tested separately. Differences in notation (Markdown/HTML/LaTeX syntax, `$O_2$` / `O₂`, bold, quote style) are not penalised; reading errors are. | |
| The scoring code, unit tests, model outputs and all results are in the GitHub repository: [moganai/TurkishOCR-Bench](https://github.com/moganai/TurkishOCR-Bench). | |
| ### Contents | |
| | File | Content | | |
| |---|---| | |
| | `goruntu/<category>/<id>.png` or `.jpg` | Page image | | |
| | `dogru_metin/<category>/<id>.md` | Ground-truth text of the page | | |
| | `metadata.jsonl` | One row per page: `file_name`, `id`, `kategori` (category), `belge` (document), `baslik` (title), `kaynak` (source), `lisans` (license), `url`, `dogru_metin_dosya` (ground-truth file), `dogru_metin` (ground-truth text) | | |
| ```python | |
| from datasets import load_dataset | |
| ds = load_dataset("moganai/TurkishOCR-Bench", split="test") | |
| example = ds[0] | |
| example["image"] # page image | |
| example["dogru_metin"] # ground-truth text | |
| example["kategori"], example["lisans"] | |
| ``` | |
| This is an **evaluation set** (single split: `test`). It must not be used to fit or fine-tune models. | |
| ### Categories | |
| | Category | Pages | Source | | |
| |---|---|---| | |
| | akademik (academic) | 60 | DergiPark journal articles | | |
| | tablo (table) | 60 | DergiPark journal articles (pages with tables) | | |
| | formul (formula) | 60 | DergiPark science and engineering articles (pages with formulas) | | |
| | taranmis (scanned) | 59 | Scanned DergiPark articles, public-domain old books, old Official Gazette scans | | |
| | kitap (book) | 50 | Open-access books from university presses | | |
| | fis_fatura_form (receipt, invoice, form) | 50 | Receipt, invoice and form images | | |
| | el_yazisi (handwriting) | 50 | Photographs of handwritten lecture notes | | |
| | slayt (slide) | 35 | Presentation slides | | |
| | resmi_yeni (Official Gazette) | 35 | Official Gazette of Türkiye | | |
| | resmi_eski (Official Gazette, old) | 29 | Official Gazette of Türkiye, old issues | | |
| Pages were selected, with per-category quotas and at most 2 pages per document, from pages the evaluated models found difficult. Scores represent hard pages, not a general page distribution. | |
| ### Ground-truth format | |
| - Headings use `#`, tables are HTML `<table>`, formulas are LaTeX (`$…$`, `$$…$$`, equation number as `\tag{}`). | |
| - Page headers are inside `<header>…</header>`; footers and page numbers are inside `<footer>…</footer>`. | |
| - Photographs, charts and diagrams are replaced by `<figure/>`; text inside figures is not transcribed. | |
| - Spelling and printing errors on the page are preserved. Words hyphenated at line ends are joined. | |
| ### License | |
| This dataset has no single license. **Each page is subject to the license of its own source**; the license is in the `lisans` field of `metadata.jsonl` and the source address is in the `url` field. | |
| | Source | Pages | License | | |
| |---|---|---| | |
| | DergiPark journal articles | 144 | CC BY-NC | | |
| | DergiPark journal articles | 33 | CC BY | | |
| | DergiPark journal articles | 17 | CC BY-SA | | |
| | DergiPark journal articles | 4 | CC BY-NC-SA | | |
| | Official Gazette of Türkiye | 90 | Law No. 5846 (FSEK) Art. 31, official legislation | | |
| | Receipts, invoices, forms | 50 | CC BY 4.0 | | |
| | Slides | 33 | CC BY 4.0 | | |
| | Slides | 2 | CC BY-SA 4.0 | | |
| | Istanbul University Press books | 25 | CC BY-NC 4.0 | | |
| | Cappadocia University Press books | 17 | CC BY-NC-SA 4.0 | | |
| | Özgür Publications books | 8 | CC BY 4.0 | | |
| | Public-domain old books (Wikimedia Commons) | 15 | Public domain | | |
| | Handwritten lecture notes (openly available sources on the internet) | 50 | No license stated | | |
| - 190 pages do not allow commercial use (NC); 40 pages require derivatives to be shared under the same license (SA). | |
| - If you are a rights holder and want a page removed, please report it through the repository. | |
| --- | |
| <a id="turkce"></a> | |
| ## Türkçe | |
| Türkçe belge OCR'ı için sayfa düzeyinde bir benchmark'ın verisi: **488 sayfa görüntüsü ve her sayfanın doğru metni** (368 belge, 10 kategori). | |
| Benchmark, bir modelin sayfayı ne kadar doğru okuduğunu ve çıktının elle düzeltmeye ne kadar az ihtiyaç duyduğunu ölçer. Türkçe'ye özgü hatalar (ı/i, ş/s, İ/I, şapkalı harf, ondalık virgül, olumsuzluk eki) ayrıca sınanır. Gösterim farkı (Markdown/HTML/LaTeX yazımı, `$O_2$` / `O₂`, kalın, tırnak türü) ceza almaz; okuma hatası alır. | |
| Puanlama kodu, birim testler, model çıktıları ve bütün sonuçlar GitHub deposundadır: [moganai/TurkishOCR-Bench](https://github.com/moganai/TurkishOCR-Bench). | |
| ### İçerik | |
| | Dosya | İçerik | | |
| |---|---| | |
| | `goruntu/<kategori>/<id>.png` ya da `.jpg` | Sayfa görüntüsü | | |
| | `dogru_metin/<kategori>/<id>.md` | Sayfanın doğru metni | | |
| | `metadata.jsonl` | Sayfa başına bir satır: `file_name`, `id`, `kategori`, `belge`, `baslik`, `kaynak`, `lisans`, `url`, `dogru_metin_dosya`, `dogru_metin` | | |
| ```python | |
| from datasets import load_dataset | |
| ds = load_dataset("moganai/TurkishOCR-Bench", split="test") | |
| ornek = ds[0] | |
| ornek["image"] # sayfa görüntüsü | |
| ornek["dogru_metin"] # doğru metin | |
| ornek["kategori"], ornek["lisans"] | |
| ``` | |
| Bu veri bir **değerlendirme kümesidir** (tek bölüm: `test`). Model eğitiminde kullanılmamalıdır. | |
| ### Kategoriler | |
| | Kategori | Sayfa | Kaynak | | |
| |---|---|---| | |
| | akademik | 60 | DergiPark makaleleri | | |
| | tablo | 60 | DergiPark makaleleri (tablolu sayfalar) | | |
| | formul | 60 | DergiPark fen ve mühendislik makaleleri (formüllü sayfalar) | | |
| | taranmis | 59 | Taranmış DergiPark makaleleri, telifi düşmüş eski kitaplar, eski Resmî Gazete taramaları | | |
| | kitap | 50 | Üniversite yayınevlerinin açık erişim kitapları | | |
| | fis_fatura_form | 50 | Fiş, fatura ve form görüntüleri | | |
| | el_yazisi | 50 | El yazısı ders notu fotoğrafları | | |
| | slayt | 35 | Sunum slaytları | | |
| | resmi_yeni | 35 | Resmî Gazete | | |
| | resmi_eski | 29 | Resmî Gazete, eski sayılar | | |
| Sayfalar, değerlendirilen modellerin zorlandığı sayfalardan kategori kotalı seçildi (belge başına en fazla 2 sayfa). Skorlar genel bir sayfa dağılımını değil, zor sayfaları temsil eder. | |
| ### Doğru metin biçimi | |
| - Başlıklar `#`, tablolar HTML `<table>`, formüller LaTeX (`$…$`, `$$…$$`, denklem numarası `\tag{}`). | |
| - Üst bilgi `<header>…</header>`, alt bilgi ve sayfa numarası `<footer>…</footer>` içindedir. | |
| - Fotoğraf, grafik ve diyagram yerine `<figure/>` vardır; görsel içi yazı aktarılmaz. | |
| - Sayfadaki yazım ve baskı hataları korunur. Satır sonunda tireyle bölünmüş kelimeler birleştirilmiştir. | |
| ### Lisans | |
| Bu veri kümesinin tek bir lisansı yoktur. **Her sayfa kendi kaynağının lisansına tabidir**; lisans `metadata.jsonl`'deki `lisans` alanında, kaynak adresi `url` alanındadır. | |
| | Kaynak | Sayfa | Lisans | | |
| |---|---|---| | |
| | DergiPark makaleleri | 144 | CC BY-NC | | |
| | DergiPark makaleleri | 33 | CC BY | | |
| | DergiPark makaleleri | 17 | CC BY-SA | | |
| | DergiPark makaleleri | 4 | CC BY-NC-SA | | |
| | Resmî Gazete | 90 | FSEK m. 31 (resmî mevzuat) | | |
| | Fiş, fatura, form | 50 | CC BY 4.0 | | |
| | Slaytlar | 33 | CC BY 4.0 | | |
| | Slaytlar | 2 | CC BY-SA 4.0 | | |
| | İstanbul Üniversitesi Yayınevi kitapları | 25 | CC BY-NC 4.0 | | |
| | Kapadokya Üniversitesi Yayınları kitapları | 17 | CC BY-NC-SA 4.0 | | |
| | Özgür Yayınları kitapları | 8 | CC BY 4.0 | | |
| | Telifi düşmüş eski kitaplar (Wikimedia Commons) | 15 | Kamu malı | | |
| | El yazısı ders notları (internet üzerinden açık kaynaklar) | 50 | Lisans belirtilmemiş | | |
| - 190 sayfa ticari kullanıma izin vermez (NC); 40 sayfa türevlerin aynı lisansla paylaşılmasını ister (SA). | |
| - Hak sahibiyseniz ve bir sayfanın kaldırılmasını istiyorsanız depo üzerinden bildirin. | |