---
license: other
language:
- az
task_categories:
- image-to-text
tags:
- ocr
- azerbaijani
- cyrillic
- line-recognition
size_categories:
- 100K<n<1M
---

# Azerbaijani OCR Lines

Line-level training data for Azerbaijani text recognition, in Latin and
Cyrillic script, extracted from scanned books.

## Fields

| field | description |
|---|---|
| `image` | cropped text line, grayscale, height 48 px |
| `text` | transcription |
| `script` | `az_latin` or `az_cyrillic` |
| `book` | anonymised source-book id |

## How it was built

Pages come from scanned PDFs that already carried an OCR text layer. Line
boxes were taken from that layer, rendered at 300 dpi and cropped. Books
were kept only when at least 85% of a page's ink fell inside the layer's
line boxes, which rejects documents whose text layer is misaligned with the
image.

## Labels are not human-verified

Transcriptions come from the pre-existing OCR layer, so the systematic
errors of that engine are present in this data. Measured against an
independently verified benchmark, roughly 40% of lines differ from what two
vision-language models read from the same crop, most often where the engine
dropped Azerbaijani-specific letters (`a` for `ə`, `к` for `ҝ`, `ч` for `ҹ`).

Train on this; do not evaluate on it. Use
[LocalDoc/azerbaijani-ocr-benchmark](https://huggingface.co/datasets/LocalDoc/azerbaijani-ocr-benchmark)
for evaluation — its books are disjoint from this set.

## Composition

Cyrillic is a small minority of lines. Sample with reweighting, or measure
per-script performance separately, or the recogniser will underfit it.

## Normalisation

Unicode NFC, typographic quotes and apostrophes folded to ASCII, soft
hyphens removed. Pre-2004 orthography (`ən'ənə`), end-of-line hyphens and
original typos are preserved as printed.