Avar Tesseract model (ava_v9)
A Tesseract 5 LSTM model for Avar (авар мацӀ) in Cyrillic script, fine-tuned from
tessdata_best/Cyrillic. It targets scanned
Avar–Russian linguistic material: bilingual dictionaries, grammars and textbooks.
It exists because the ava.traineddata in circulation was never trained on Avar. That file
is stock Cyrillic.traineddata with two unicharset entries relabelled (U+0406 І → U+04C0
Ӏ, U+0456 і → U+04CF ӏ); its network weights are byte-identical to upstream. It
drops the palochka, emits ], or produces U+04CF where Avar requires U+04C0.
Usage
cp ava_v9.traineddata "$TESSDATA_PREFIX/avar.traineddata"
tesseract page.png out -l avar --psm 3
Use --psm 3 (Tesseract's default). Forcing --psm 6 costs more than a point of CER on
books with mixed type sizes or columns.
Float model, 11 MB, Tesseract ≥ 5.0.
Results
Character error rate at psm 3, palochka lookalikes folded before scoring. Pages were rendered at 300 DPI and scored against each PDF's own text layer.
| model | dictionary (1993) | textbook (2018) | monograph | Alekseev & Ataev |
|---|---|---|---|---|
stock rus |
4.93 % | 4.40 % | 4.91 % | 4.01 % |
PaddleOCR cyrillic_PP-OCRv5 |
5.62 % | 1.74 % | 3.47 % | 2.44 % |
relabelled ava (circulating) |
3.41 % | 2.31 % | 3.60 % | 3.17 % |
ava_v9 (this model) |
1.84 % | 1.83 % | 3.43 % | 2.70 % |
The first three books contributed training lines, with the benchmark pages excluded. Alekseev & Ataev was never used for training in any form, so it is the only test of generalisation.
Codepoint correctness
CER hides the defect that matters most for search: emitting the wrong character for the
palochka. In an Avar Quran, which typesets the palochka as the digit 1, characters in
palochka position over eight pages (pp. 900–907):
| model | Ӏ U+04C0 |
digit 1 |
other | correct |
|---|---|---|---|---|
relabelled ava (circulating) |
121 | 396 | 30 | 22 % |
ava_v9 |
523 | 4 | 1 | 99 % |
The model normalises the palochka to U+04C0 whether the page prints Ӏ, Latin I or the
digit 1.
Training
- Base:
tessdata_best/Cyrillic(Apache-2.0). - Data: synthetic line renders of Avar text in several serif faces, plus text lines harvested from scanned Avar books and corrected where two OCR engines disagreed. The scanned books are under copyright and are not distributed; neither are the line images.
- Augmentation: synthetic stress accents (to stop accented Russian glosses being read
as
ё), and palochka substitution: drawing1,I,lor|in palochka position while transcribingӀ.
Limitations
- On modern offset print a generic model does better. PaddleOCR PP-OCRv5 wins on the 2018 textbook and on Alekseev. This model's advantage is Soviet-era letterpress, and emitting Avar codepoints rather than lookalikes.
- Generalisation rests on one held-out book. On Alekseev the model is within noise of its predecessors, not clearly better.
- The palochka supervision is synthetic. The Quran check is the only evidence it transfers to real print, and that is one book.
- Residual errors: stray
е→ёon bold all-caps headwords, comma/period confusion at small point sizes, and occasional doubled vowels in all-caps headwords (БАЧӀИИНforБАЧӀИН).
Licence
Apache-2.0, the same as the base model.