chagatai-project/chagatai-model

Chagatai (Arabic) → English / Kazakh Translation Model

Model Description

This model is a fine-tuned NLLB-200-distilled-600M model for translation from
Chagatai (Arabic script) to English and Kazakh (Cyrillic).

Chagatai is a historical Turkic literary language used in Central Asia until the early 20th century.
The model focuses exclusively on Chagatai written in Arabic script.

Supported Translation Directions

• Chagatai (Arabic script) → English
• Chagatai (Arabic script) → Kazakh (Cyrillic)

Training Data

The model was trained on two combined datasets containing both word-level and sentence-level aligned translations.

1. Sözdikqor + QazCorpus Transliteration Lexicon

• Source: QazCorpus platform
  https://sozdikqor.kz/sozdik/?id=16#page-332
• Origin: Digitized texts of Abulghazi Bahadur Khan
• Languages Used:
  • Chagatai (Arabic script)
  • English
  • Kazakh
• Size: ~2,600 word-level entries

This dataset provides high-quality historical vocabulary aligned with modern Kazakh and English meanings.

2. Digitized Textbook Corpus

• Source:
  An Introduction to Chaghatay: A Graded Textbook for Reading Central Asian Sources
  by Eric Schluessel
  https://quod.lib.umich.edu/m/maize/images/mpub10110094.pdf
• Structure: Chapter-based educational corpus
• Languages Used:
  • Chagatai (Arabic script) — chg_Arab
  • English — eng_Latn
  • Kazakh — kaz_Cyrl

Each chapter includes:

• complete sentences, and
• individual vocabulary items used in those sentences.

This structure allows the model to learn both lexical translations and contextual sentence usage.

Total Training Data

• ~5,300 aligned examples
  • word-level
  • sentence-level

Language Codes

Since Chagatai is not included in the standard NLLB language set, custom tokens were added.

Language	Script	Token	Notes
Chagatai	Arabic	chg_Arab	Custom token added
English	Latin	eng_Latn	Standard NLLB token
Kazakh	Cyrillic	kaz_Cyrl	Standard NLLB token

Note: chg_Latn token was also added but not used during training.

Training Details

• Base Model: facebook/nllb-200-distilled-600M
• Epochs: 5
• Batch Size: 16
• Learning Rate: 2e-5
• Framework: Hugging Face Transformers
• Hardware: NVIDIA A100 (Google Colab)

Performance

Evaluation was conducted on held-out test data (~30% of the dataset).

Direction	Exact Match	BLEU
Chagatai → English	12.86%	35.10
Chagatai → Kazakh	18.70%	41.19

Note:
Exact Match scores are relatively low due to:

• synonym variation,
• morphological differences,
• article usage in English.

BLEU scores indicate strong performance for a low-resource historical language.

Acknowledgements

This work was conducted within the framework of the project:

BR28712621
“Handwritten Heritage of Kazakhstan: registration, restoration, scientific cataloging, digitization, and comprehensive codicological studies.”