Elamur Research⌕
☰
home / papers / 2026 / AfriNLLB: Efficient Translation Models for African Languages

AfriNLLB: Efficient Translation Models for African Languages

Paper · 2026 · Yasmin Moslem, Aman Kassahun Wassie and Amanuel Gizachew Abebe

Summary

Machine translation for African languages is held back by two problems. Parallel data is scattered across many sources, and existing models are often too heavy for low-resource deployment, for example in government or healthcare services. The authors present AfriNLLB, a family of compact translation models derived from NLLB-200 600M. It covers 15 language pairs (30 directions):

  • ten African languages, including Swahili, Hausa, Amharic, Wolof and Egyptian Arabic, paired with English or French;
  • MSA Arabic, French, Portuguese and Spanish, paired with English.

They gathered parallel corpora mostly from OPUS and Hugging Face and cleaned them with a four-step pipeline: rules, language identification, embedding similarity and quality estimation. Training then proceeds in three stages:

  1. The base model is fine-tuned.
  2. Decoder layers are removed one at a time, each time dropping the layer whose loss costs the least chrF++ on Flores200 dev.
  3. The smaller network is retrained with knowledge distillation from a larger teacher.

The resulting 548M model nearly matches or slightly exceeds the original baseline in quality. With float16 quantization it runs about 57% faster. Models (Transformers and CTranslate2 formats), data and code are all openly released, which lowers the cost of building practical MT for these languages.

Key points

  • The filtering pipeline uses a 0.6 threshold and the following tools:
    • AfroLID and fastText for language detection;
    • LaBSE and DistilUSE for semantic similarity (no Lingala embedding model was available, so Lingala skipped this step);
    • COMET and AfriCOMET-QE-STL for quality estimation.
  • About 7.2M raw pairs shrank to 6.4M after cleaning. High-resource pairs were then downsampled to 200k each, giving roughly 1.6M training pairs (3.2M once reversed). Distillation added 568k filtered segments for African languages.
  • Evaluation used Flores200 dev (997 segments) for validation and layer selection, and devtest (1,012 segments) for testing. Metrics were BLEU, chrF++, AfriCOMET and COMET.
  • On average for en→xx, chrF++ went from 47.89 (baseline) to 50.05 (pruned + FT). For xx→en it went from 56.22 to 56.98. Pruning alone gave about a 23% speedup.
  • In the ablations, greedy iterative pruning beat removing middle layers 4–7. Fine-tuning after pruning proved essential. Dropping up to 6 of the 12 decoder layers still roughly matched or beat the baseline.

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

 Institutions

🎓Trinity College DublinDublin · Ireland24 works · 1,231 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

🧭 Topics 2Research subjects and languages the work is about.Machine translation · 1,087 worksLow-resource languages · 267 works

Cite

Cite as Yasmin Moslem, Aman Kassahun Wassie and Amanuel Gizachew Abebe. 2026. AfriNLLB: Efficient Translation Models for African Languages. Proceedings of the 7th Workshop on African Natural Language Processing (AfricaNLP 2026).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.africanlp-main.30

@inproceedings{moslem2026afrinllb,
  title = {{AfriNLLB: Efficient Translation Models for African Languages}},
  author = {Yasmin Moslem, Aman Kassahun Wassie and Amanuel Gizachew Abebe},
  booktitle = {Proceedings of the 7th Workshop on African Natural Language Processing (AfricaNLP 2026)},
  year = {2026},
  doi = {10.18653/v1/2026.africanlp-main.30},
  url = {https://doi.org/10.18653/v1/2026.africanlp-main.30}
}

References

55 references, 26 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Idris Abdulmumin, Satya Ranjan Dash, Musa Abdullahi Dawud, Shantipriya Parida, Shamsuddeen Muhammad, Ibrahim Sa’id Ahmad, Subhadarshi Panda, Ondřej Bojar, Bashir Shehu Galadanci, and Bello Shehu Bello. 2022. Hausa visual genome: A dataset for multi-modal English to Hausa machine translation. In Pro… find
  2. Ife Adebara, AbdelRahim Elmadany, Muhammad Abdul-Mageed, and Alcides Inciarte. 2022. AfroLID: A neural language identification tool for African languages. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 1958–1981, Abu Dhabi, United Arab Emirates. Asso… find
  3. David Adelani, Jesujoba Alabi, Angela Fan, Julia Kreutzer, Xiaoyu Shen, Machel Reid, Dana Ruiter, Dietrich Klakow, Peter Nabende, and 14 others. 2022. A few thousand translations go a long way! leveraging pre-trained models for African news translation. In Proceedings of the 2022 Conference of the N…
  4. David Adelani, Dana Ruiter, Jesujoba Alabi, Damilola Adebonojo, Adesina Ayeni, Mofetoluwa Adeyemi, Ayodele Awokoya, and Cristina Espa na-Bonet. 2021. MENYO-20k: A multi-domain English-Yor‘ub’a corpus for machine translation and domain adaptation. In Proceedings of the Second Workshop on African Natu… find
  5. David Ifeoluwa Adelani, Alison Chi, Simbiat Aderibigbe, Butoyi Beatrice, Tumaini Balikwisha, Barkwende Hugues Diallo, Tunde Oluwaseyi Ajayi, Joseph K. O. Oaminu, Ruqayya Nasir Iro, and 12 others. 2025a. AFRIDOC-MT: Document-level MT Corpus for African Languages. arXiv preprint arXiv:2501.06374.
  6. David Ifeoluwa Adelani, Jessica Ojo, Israel Abebe Azime, Jian Yun Zhuang, Jesujoba Oluwadara Alabi, Xuanli He, Millicent Ochieng, Sara Hooker, Andiswa Bukula, and 18 others. 2025b. IrokoBench: A new benchmark for African languages in the age of large language models. In Proceedings of the 2025 Confe…
  7. Rania Al-Sabbagh. 2024. Arzen-multigenre: An aligned parallel dataset of egyptian arabic song lyrics, novels, and subtitles, with english translations. Data in Brief, 54:110271.
  8. Duarte Miguel Alves, Jose Pombal, Nuno M. Guerreiro, Pedro Henrique Martins, Joao Alves, Amin Farajian, Ben Peters, Ricardo Rei, Patrick Fernandes, and 4 others. 2025. WMT24++: Expanding the Language Coverage of WMT24 to 55 Languages & Dialects. arXiv preprint arXiv:2502.12404.
  9. Antonios Anastasopoulos, Alessandro Cattelan, Zi-Yi Dou, Marcello Federico, Christian Federmann, Dmitriy Genzel, Franscisco Guzmán, Junjie Hu, Macduff Hughes, and 9 others. 2020. TICO-19: the Translation Initiative for COvid-19. In Proceedings of the 1st Workshop on NLP for COVID-19 (Part 2) at EMNL…
  10. Israel Abebe Azime, Atnafu Lambebo Tonja, Tadesse Destaw Belay, Mitiku Yohannes Fuge, Aman Kassahun Wassie, Eyasu Shiferaw Jada, Yonas Chanie, Walelign Tewabe Sewunetie, and Seid Muhie Yimam. 2024. Walia-LLM: Enhancing Amharic-LLaMA by integrating task-specific and generative datasets. In Findings o…
  11. Isaac Caswell, Elizabeth Nielsen, Jiaming Luo, Colin Cherry, Geza Kovacs, Hadar Shemtov, Partha Talukdar, Dinesh Tewari, Baba Mamadi Diane, and 13 others. 2025. SMOL: Professionally translated parallel data for 115 under-represented languages. arXiv preprint arXiv:2502.12301. find
  12. Mauro Cettolo, Christian Girardi, and Marcello Federico. 2012. WIT3: Web inventory of transcribed and translated talks. In Proceedings of the 16th Annual Conference of the European Association for Machine Translation, pages 261–268. European Association for Machine Translation.
  13. Christos Christodouloupoulos and Mark Steedman. 2015. A massively parallel corpus: the bible in 100 languages. Language Resources and Evaluation, 49(2):375–395.
  14. Marta R Costa-jussà, James Cross, Onur Çelebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, and 29 others. 2022. No Language Left Behind: Scaling human-centered machine translation. arXiv [cs.CL]. find
  15. Josep Crego and Jean Senellart. 2016. Neural Machine Translation from Simplified Translations. arXiv [cs.CL]. find
  16. Andreas Eisele and Yu Chen. 2010. MultiUN: A multilingual corpus from united nation documents. In Proceedings of the Seventh International Conference on Language Resources and Evaluation (LREC’10). European Language Resources Association (ELRA).
  17. Kenneth Enevoldsen, Isaac Chung, Imene Kerboua, Márton Kardos, Ashwin Mathur, David Stap, Jay Gala, Wissam Siblini, Dominik Krzemiński, and 77 others. 2025. Mmteb: Massive multilingual text embedding benchmark. arXiv preprint arXiv:2502.13595.
  18. Angela Fan, Shruti Bhosale, Holger Schwenk, Zhiyi Ma, Ahmed El-Kishky, Siddharth Goyal, Mandeep Baines, Onur Celebi, Guillaume Wenzek, and 7 others. 2021. Beyond English-Centric Multilingual Machine Translation. Journal of Machine Learning Research, 22(107):1–48. find
  19. Muhammad Hazim Al Farouq, Aman Kassahun Wassie, and Yasmin Moslem. 2025. Bemba Speech Translation: Exploring a Low-Resource African Language. In Proceedings of the 22nd International Conference on Spoken Language Translation (IWSLT 2025), pages 354–359, Vienna, Austria (in-person and online). Associ… find
  20. Christian Federmann, Tom Kocmi, and Ying Xin. 2022. NTREX-128 – a benchmark for evaluating machine translation performance. In Proceedings of the First Workshop on Scaling Up Multilingual Evaluation, pages 21–24. Association for Computational Linguistics. find
  21. Fangxiaoyu Feng, Yinfei Yang, Daniel Cer, Naveen Arivazhagan, and Wei Wang. 2022. Language-agnostic BERT sentence embedding. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 878–891, Stroudsburg, PA, USA. Association for Comput…
  22. Sanchit Gandhi, Patrick von Platen, and Alexander M Rush. 2023. Distil-Whisper: Robust knowledge distillation via large-scale pseudo labelling. arXiv [cs.CL]. find
  23. Barbara F. Grimes. 1996. Ethnologue: Languages of the World, 13th edition. SIL International, Dallas, TX. Summer Institute of Linguistics. find
  24. Bernd Heine and Derek Nurse, editors. 2000. African Languages: An Introduction. Cambridge University Press, Cambridge. find
  25. Andreea Iana, Goran Glavoči’c, and Heiko Paulheim. 2023. News without borders: Domain adaptation of multilingual sentence embeddings for cross-lingual news recommendation. In Proceedings of the 17th ACM Conference on Recommender Systems. ACM. find
  26. Armand Joulin, Edouard Grave, Piotr Bojanowski, and Tomas Mikolov. 2017. Bag of Tricks for Efficient Text Classification. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers, pages 427–431, Valencia, Spain. Associatio… find
  27. Yoon Kim and Alexander M Rush. 2016. Sequence-Level Knowledge Distillation. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1317–1327, Austin, Texas. Association for Computational Linguistics.
  28. Guillaume Klein, Dakun Zhang, Clément Chouteau, Josep Crego, and Jean Senellart. 2020. Efficient and high-quality neural machine translation with Open-NMT. In Proceedings of the Fourth Workshop on Neural Generation and Translation, pages 211–217, Stroudsburg, PA, USA. Association for Computational L… find
  29. Laban Kumbuga, Joyce Nakatumba-Nabende, Jonathan Mukiibi, and Andrew Katumba. 2024. SALT: Sunbird African language translation corpus. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 5462– 5472. ELRA… find
  30. Quentin Lhoest, Albert Villanova del Moral, Yacine Jernite, Abhishek Thakur, Patrick von Platen, Suraj Patil, Julien Chaumond, Mariama Drame, Julien Plu, and 23 others. 2021. Datasets: A community library for natural language processing. In Proceedings of the 2021 Conference on Empirical Methods in…
  31. Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, and Luke Zettlemoyer. 2020. Multilingual Denoising Pre-training for Neural Machine Translation. Transactions of the Association for Computational Linguistics, 8:726–742. find
  32. Yasmin Moslem. 2025. Efficient speech translation through model compression and knowledge distillation. In Proceedings of the 22nd International Conference on Spoken Language Translation (IWSLT 2025), pages 379–388, Stroudsburg, PA, USA. Association for Computational Linguistics. find
  33. Yasmin Moslem, Muhammad Hazim Al Farouq, and John Kelleher. 2025. Iterative layer pruning for efficient translation inference. In Proceedings of the Tenth Conference on Machine Translation, pages 1022–1027, Suzhou, China. Association for Computational Linguistics. find
  34. Jessica Ojo, Odunayo Ogundepo, Akintunde Oladipo, Kelechi Ogueji, Jimmy Lin, Pontus Stenetorp, and David Ifeoluwa Adelani. 2025. AfroBench: How good are large language models on African languages? In Findings of the Association for Computational Linguistics: ACL 2025, pages 19048–19095, Vienna, Aust… find
  35. Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a Method for Automatic Evaluation of Machine Translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311–318, Philadelphia, Pennsylvania, USA. Association for Computational…
  36. David Peer, Sebastian Stabinger, Stefan Engl, and Antonio Rodríguez-Sánchez. 2022. Greedy-layer pruning: Speeding up transformer models for natural language processing. Pattern Recognit. Lett., 157:76–82. find
  37. Maja Popović. 2017. chrF++: words helping character n-grams. In Proceedings of the Second Conference on Machine Translation, pages 612–618, Copenhagen, Denmark. Association for Computational Linguistics.
  38. Matt Post. 2018. A Call for Clarity in Reporting BLEU Scores. In Proceedings of the Third Conference on Machine Translation: Research Papers, pages 186– 191, Brussels, Belgium. Association for Computational Linguistics.
  39. Ricardo Rei, Craig Stewart, Ana C Farinha, and Alon Lavie. 2020. COMET: A Neural Framework for MT Evaluation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2685–2702, Online. Association for Computational Linguistics.
  40. Nils Reimers and Iryna Gurevych. 2019. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3982–3992,…
  41. Nils Reimers and Iryna Gurevych. 2020. Making monolingual sentence embeddings multilingual using knowledge distillation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4512–4525, Stroudsburg, PA, USA. Association for Computational Linguistics…
  42. Pedram Rostami and Mohammad Javad Dousti. 2024. CULL-MT: Compression using language and layer pruning for machine translation. arXiv [cs.CL]. find
  43. Ralf Steinberger, Bruno Pouliquen, Anna Widiger, Camelia Ignat, Tomaž Erjavec, Dan Tufiş, and Dániel Varga. 2006. The JRC-Acquis: A multilingual aligned parallel corpus with 20+ languages. In Proceedings of the Fifth International Conference on Language Resources and Evaluation (LREC’06), Genoa, It…
  44. 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
  45. Jörg Tiedemann. 2020. The Tatoeba Translation Challenge – Realistic Data Sets for Low Resource and Multilingual MT. In Proceedings of the Fifth Conference on Machine Translation, pages 1174–1182, Online. Association for Computational Linguistics.
  46. Jiayi Wang, David Adelani, Sweta Agrawal, Marek Masiak, Ricardo Rei, Eleftheria Briakou, Marine Carpuat, Xuanli He, Sofia Bourhim, and 49 others. 2024. AfriMTE and AfriCOMET: Enhancing COMET to embrace under-resourced African languages. In Proceedings of the 2024 Conference of the North American Cha… find
  47. Aman Kassahun Wassie. 2024. Machine translation for ge’ez language. arXiv preprint arXiv:2311.14530, arXiv:2311.14530.
  48. Aman Kassahun Wassie, Mahdi Molaei, and Yasmin Moslem. 2025. Domain-specific translation with open-source large language models: Resource-oriented analysis. arXiv preprint arXiv:2412.05862, arXiv:2412.05862. — [arXiv:2412.05862](https://arxiv.org/abs/2412.05862) find
  49. Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, and 13 others. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing:…
  50. Yinfei Yang, Daniel Cer, Amin Ahmad, Mandy Guo, Jax Law, Noah Constant, Gustavo Hernandez Abrego, Steve Yuan, Chris Tar, and 3 others. 2020. Multilingual Universal Sentence Encoder for Semantic Retrieval. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: Sys…
  51. Biao Zhang, Philip Williams, Ivan Titov, and Rico Sennrich. 2020. Improving massively multilingual neural machine translation and zero-shot translation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1628– 1639. Association for Computational Linguis…
  52. NLLB – FP16 NLLB + FT – FP16 35.15 35.10 57.61 57.61 71.87 71.87 1530.94 2808.90 20.39 11.15 – FP16 34.01 34.05 56.98 56.99 71.20 71.19 1807.61 3513.32 17.38 8.96 – FP16 33.35 33.32 56.48 56.45 70.79 70.79 2028.18 4000.25 15.41 7.82 – FP16 32.03 32.01 55.62 55.60 69.71 69.71 2257.03 4589.42 13.77 6.… find
  53. NLLB – FP16 22.70 22.68 47.89 47.88 69.36 69.38 1530.10 2898.38 28.09 15.33 NLLB + FT – FP16 24.28 24.14 49.97 49.84 70.91 70.90 1610.23 2811.34 26.98 18.82 – FP16 24.17 24.15 50.05 50.06 70.37 70.41 1946.61 3732.72 22.51 11.98 – FP16 23.48 23.49 49.34 49.35 68.98 69.00 2265.87 4428.68 18.50 9.65 –… find
  54. NLLB – FP16 16.41 16.33 38.83 38.83 17.34 17.23 1475.48 2850.66 26.46 13.71 NLLB + FT – FP16 17.91 17.83 40.45 40.42 18.47 18.37 1524.32 2749.45 26.12 14.68 – FP16 17.43 17.38 40.21 40.18 14.52 14.53 1845.09 3569.23 21.61 11.17 – FP16 16.52 16.54 39.44 39.42 11.78 11.68 2044.27 3953.51 19.21 9.92 –… find
  55. NLLB + FT – FP16 10.98 10.48 35.68 35.05 21.33 21.49 1081.84 1700.25 51.56 51.31 – FP16 10.20 10.11 35.21 35.13 20.04 20.03 1261.66 2313.85 49.91 31.15 – FP16 10.07 9.99 35.14 35.08 19.83 19.78 1416.33 2465.60 30.89 18.68 – FP16 7.57 7.57 32.42 32.38 14.16 14.29 1207.06 2069.52 38.75 23.25 – FP16 9.… find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/moslem-2026-afrinllb-efficient-translation-models-african/