Elamur Research⌕
☰
home / papers / 2026 / Data-Centric Approach at the LoResMT 2026 Turkic Translation Challenge: Russian-Kyrgyz

Data-Centric Approach at the LoResMT 2026 Turkic Translation Challenge: Russian-Kyrgyz

Summary

This system description targets Russian→Kyrgyz machine translation, a low-resource pair with little parallel data, submitted to the LoResMT 2026 Turkic Languages Translation Challenge. The author takes a data-centric approach. Parallel text from OPUS for Russian, Kyrgyz, Uzbek and Tajik was cleaned with rule-based filters and fastText language ID, then scored with SONAR (blaser_2.0) and split into standard- and high-quality tiers. This data was supplemented with two synthetic sources: Kyrgyz FineWeb-2 text back-translated into Russian by Gemma-3-27B and Qwen3-235B, and SiberianPersonaChat dialogues translated into Kyrgyz by GPT-4o. A vocabulary-pruned mt0-large model (about 800M parameters) was trained through a four-stage curriculum, moving from noisier multilingual data to cleaner, more specialized synthetic sets. The final model scored 44.9 chrF++ on FLORES-200 devtest, beating much larger systems such as NLLB-200 54B. It ranked first in the shared task with 49.1 chrF++. The work shows that careful data curation can let a compact model outperform far bigger ones. The model, synthetic data and a demo are released openly.

Key points

  • Data: About 8.2M standard-quality and 7.1M high-quality OPUS sentence pairs were used, with SONAR thresholds of [3.0, 3.7) and ≥3.7. These were supplemented by about 3.1M synthetic FineWeb-2 pairs and about 113K synthetic dialogue pairs. The back-translated data was further filtered by MADLAD-400-7B-MT loss, keeping pairs in the range (0.1, 4.5).
  • Auxiliary languages: An ablation with mT0-small (300M) showed that adding both ru-uz and ru-tg data to ru-ky gave the best results on FLORES-200 devtest.
  • Model: Vocabulary pruning cut mt0-large from 1.23B to about 800M parameters while leaving the non-embedding weights unchanged.
  • Results on FLORES-200 devtest (ru→ky):
    • Scores rose across the curriculum stages from 32.2 to 44.9 chrF++, reaching 80.5 xCOMET-XXL.
    • The model beat NLLB-200 54B by 3.2 chrF++ and 7.1 xCOMET-XXL.
    • In the reverse direction (ky→ru), it scored 42.4 chrF++ and 82.8 xCOMET-XXL.
  • Shared task: The system ranked first on the official test with 49.1 chrF++ and 69.7 xCOMET-XXL. The released resources on Hugging Face are Novokshanov/ru-ky-mt0-loresmt2026, the ru-ky-synthetic-loresmt2026 dataset, and a Spaces demo.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields: Linguistics & languages · Machine translation

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.loresmt-1.22

@inproceedings{novokshanov2026data,
  title = {{Data-Centric Approach at the LoResMT 2026 Turkic Translation Challenge: Russian-Kyrgyz}},
  author = {Novokshanov, Dmitry},
  booktitle = {Proceedings for the Ninth Workshop on Technologies for Machine Translation of Low Resource Languages (LoResMT 2026)},
  year = {2026},
  doi = {10.18653/v1/2026.loresmt-1.22},
  url = {https://doi.org/10.18653/v1/2026.loresmt-1.22}
}

References

28 references, 11 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Firat, and Sriram Chellappan. 2021. A large-scale study of machine translation in Turkic languages. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 5876–5890, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics… find
  2. Anton Alekseev and Timur Turatali. 2024. Kyr-gyzNLP: Challenges, progress, and future. Preprint, arXiv:2411.05503. — [arXiv:2411.05503](https://arxiv.org/abs/2411.05503) find
  3. Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2014. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473. find our card
  4. Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. 2009. Curriculum learning. In Proceedings of the 26th Annual International Conference on Machine Learning (ICML), pages 41–48. ACM. find our card
  5. Ivan Ramovich Denis Petrov. 2023. Russian dataset for chat models. find
  6. Paul-Ambroise Duquenne, Holger Schwenk, and Benoît Sagot. 2023. SONAR: Sentence-level multimodal and language-agnostic representations. arXiv preprint arXiv:2308.11466. find
  7. Gemma Team. 2025. Gemma 3 technical report. arXiv preprint arXiv:2503.19786. find
  8. Edouard Grave, Piotr Bojanowski, Prakhar Gupta, Armand Joulin, and Tomas Mikolov. 2018. Learning word vectors for 157 languages. In Proceedings of the International Conference on Language Resources and Evaluation (LREC 2018). find our card
  9. Nuno M. Guerreiro, Ricardo Rei, Sara Stymne, Alon Lavie, and André F. T. Martins. 2024. xCOMET: Transparent machine translation evaluation through fine-grained error detection. Transactions of the Association for Computational Linguistics, 12:979– 995. find our card
  10. Barry Haddow, Rachel Bawden, Antonio Valerio Miceli Barone, Jindřich Helcl, and Alexandra Birch. 2022. Survey of low-resource machine translation. Computational Linguistics, 48(3):673–732. find our card
  11. Lars Johanson and Éva Á Csató. 2015. The Turkic Languages. Routledge. find
  12. Sneha Kudugunta, Isaac Caswell, Biao Zhang, Xavier Garcia, Derrick Xin, Aditya Kusber, Romi Stella, Ankur Bapna, and Orhan Firat. 2024. MADLAD-400: A multilingual and document-level large audited dataset. Advances in Neural Information Processing Systems, 36. find our card
  13. Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In Proceedings of the 7th International Conference on Learning Representations (ICLR). find
  14. Jamshidbek Mirzakhalov, Anoop Babu, Duygu Ataman, Sherzod Kariev, Francis Tyers, Otabek Abduraufov, Mammad Hajili, Sardana Ivanova, Abror Khaytbaev, Antonio Laverghetta Jr., Bekhzodbek Moydinboyev, Esra Onal, Shaxnoza Pulatova, Ahsan Wahab, Orhan Niklas Muennighoff, Thomas Wang, Lintang Sutawika, Ad… find our card
  15. Toan Q. Nguyen and David Chiang. 2017. Transfer learning across low-resource, related languages for neural machine translation. In Proceedings of the Eighth International Joint Conference on Natural Language Processing (IJCNLP), pages 296–301, Taipei, Taiwan. Asian Federation of Natural Language Pro… find our card
  16. NLLB Team, Marta R. Costa-jussà, James Cross, Onur Çelebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, Anna Sun, Skyler Wang, Guillaume Wenzek, Alison Youngblood, Bapi Akula, Loïc Barrault, Gabriel Mejia Gonzalez, Prangthip Hansanti, and… find
  17. OpenAI. 2024. GPT-4o system card. find our card
  18. Guilherme Penedo, Hynek Kydlíček, Anton Lozhkov, Margaret Mitchell, Thomas Wolf, and Leandro von Werra. 2024. FineWeb-2: A 14 trillion token multilingual web dataset. Hugging Face Technical Report. find
  19. Maja Popović. 2017. chrF++: Words helping character n-grams. In Proceedings of the Second Conference on Machine Translation (WMT), pages 612–618, Copenhagen, Denmark. Association for Computational Linguistics. find our card
  20. Volker Rybatzki. 2020. The altaic languages: Tungusic, mongolic, turkic. In The Oxford Guide to the Transeurasian Languages, pages 22–28. Oxford University Press. find
  21. Rico Sennrich, Barry Haddow, and Alexandra Birch. 2015. Improving neural machine translation models with monolingual data. arXiv preprint arXiv:1511.06709. find
  22. Jörg Tiedemann. 2012. Parallel data, tools and interfaces in OPUS. In Proceedings of the Eighth International Conference on Language Resources and Evaluation (LREC’12), pages 2214–2218, Istanbul, Turkey. European Language Resources Association (ELRA). find
  23. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems 30 (NeurIPS 2017), pages 5998– 6008. Curran Associates, Inc. find
  24. Leandro von Werra, Lewis Tunstall, Nandan Thakur, Joe Davison, Yacine Jernite, Tristan Moran, and Thomas Wolf. 2022. Evaluate: A library for easily evaluating machine learning models and datasets. find
  25. Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. 2021. mt5: A massively multilingual pre-trained text-to-text transformer. Preprint, arXiv:2010.11934. find
  26. An Yang, Baosong Yang, Beichen Zhang, Binyuan Wang, Bowen Lin, Bwen Hui, and 1 others. 2025. Qwen3 technical report. arXiv preprint arXiv:2505.09388. find our card
  27. Michael Zhu and Suyog Gupta. 2017. To prune, or not to prune: exploring the efficacy of pruning for model compression. Preprint, arXiv:1710.01878. — [arXiv:1710.01878](https://arxiv.org/abs/1710.01878) find
  28. Barret Zoph, Deniz Yuret, Jonathan May, and Kevin Knight. 2016. Transfer learning for low-resource neural machine translation. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1568–1575, Austin, Texas. Association for Computational Linguistics. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/novokshanov-2026-data-centric-approach-loresmt-2026/