| ---
|
| license: cc-by-4.0
|
| language:
|
| - az
|
| task_categories:
|
| - image-to-text
|
| tags:
|
| - ocr
|
| - azerbaijani
|
| - cyrillic
|
| - historical-documents
|
| size_categories:
|
| - 1K<n<10K
|
| ---
|
|
|
| # Azerbaijani OCR Benchmark
|
|
|
| Line-level OCR benchmark for Azerbaijani in both Latin and Cyrillic script,
|
| built from scanned books.
|
|
|
| ## Fields
|
|
|
| | field | description |
|
| |---|---|
|
| | `image` | cropped text line, grayscale, height 48 px |
|
| | `text` | verbatim transcription |
|
| | `script` | `az_latin` or `az_cyrillic` |
|
| | `book` | anonymised source-book id |
|
|
|
| ## How labels were produced
|
|
|
| Every line carries a label agreed on independently by three sources: the
|
| OCR text layer already present in the source PDF, and two separate passes
|
| of a vision-language model that saw only the image. Lines where the three
|
| disagreed were excluded.
|
|
|
| ## Known limitations
|
|
|
| Excluding disagreements removes the hardest lines, so absolute CER measured
|
| here is optimistic relative to unconstrained material. Comparisons between
|
| models on this benchmark remain valid. Disagreements concentrate on
|
| `ҹ/ч`, `ҝ/к`, `ү/у` and on the vertical-stroke family `и/н/п/м`, which are
|
| genuinely ambiguous at the print resolution of the sources.
|
|
|
| Transcriptions are verbatim: pre-2004 orthography (`ən'ənə`), end-of-line
|
| hyphens and original typos are preserved as printed.
|
|
|
| ## Evaluation
|
|
|
| Report CER and WER overall, and separately per `script`. A diacritics-only
|
| CER over `ə ı ğ ş ç ö ü İ` and their Cyrillic counterparts is the most
|
| informative single number for this language.
|
|
|
