Model Card for nllb-luo-swa-mt-v1

Model Overview

Model Name: nllb-luo-swa-mt-v1
Model Type: Machine Translation (Luo (Dholuo) to Swahili)
Base Model: NLLB-200-distilled-600M
Languages: Luo (Dholuo), Swahili
Version: 1.0
License: CC0 (Public Domain)
Dataset: SalomonMetre13/luo_swa_arXiv_2501.11003

This model is a fine-tuned version of the NLLB-200-distilled-600M model for translation between Luo (Dholuo) and Swahili. It was trained on a parallel corpus derived from the Dholuo–Swahili corpus created by Mbogho et al. (2025), based on community-driven data collection efforts.

Model Description

The nllb-luo-swa-mt-v1 model performs machine translation from Luo (Dholuo) to Swahili, designed to improve translation capabilities for these low-resource languages. It was fine-tuned using the parallel corpus from the paper "Building low-resource African language corpora: A case study of Kidawida, Kalenjin and Dholuo" by Mbogho et al. (2025). This model is particularly valuable for promoting linguistic diversity and facilitating the development of Natural Language Processing (NLP) tools in African languages.

Key Features:

• Training Data: Fine-tuned on the Dholuo–Swahili parallel text corpus from the dataset SalomonMetre13/luo_swa_arXiv_2501.11003, derived from the grassroots data collection effort by Mbogho et al. (2025).
• Performance: Achieved a BLEU score of 21.56 on the evaluation set, showing strong performance in a low-resource setting.
• Qualitative Analysis: Translations generated by this model are sometimes more fluent and accurate than the provided reference translations.

Intended Use

This model can be used for machine translation applications between Luo (Dholuo) and Swahili. Potential use cases include:

• Educational tools: Enabling educational content in both languages, aiding language learners and teachers.
• Public health and community development: Translating health information, community messages, and official communications.
• Cultural preservation: Supporting the preservation and growth of the Luo language in the digital age.

Model Evaluation

The model was evaluated using the BLEU score, which is commonly used to assess machine translation performance. A BLEU score of 21.56 was achieved, which is a strong result for a low-resource language pair. Qualitative analysis of the translations suggests that, in some cases, the model's outputs outperform the reference translations in terms of fluency and accuracy.

Training Details

• Training Data: The model was trained on the Dholuo–Swahili parallel corpus, based on the dataset SalomonMetre13/luo_swa_arXiv_2501.11003 derived from Mbogho et al.'s (2025) work. The corpus includes text translations and is publicly available for further use and improvement.
• Model Architecture: The model is fine-tuned from the NLLB-200-distilled-600M version of the NLLB model family, which is designed for multilingual translation tasks.

Limitations

• Low-Resource Context: While the model performs well given the limited amount of data, its performance may still lag behind models trained on larger corpora for more widely spoken languages.
• Domain-Specific Use: The model may require additional fine-tuning to perform optimally on domain-specific text such as medical, legal, or technical content.

Future Directions

• Expanding the Dataset: The quality and coverage of the model could be improved by incorporating more diverse and larger datasets.
• Additional Language Pairs: Further fine-tuning of the model to support other language pairs involving Luo and Swahili could make the model even more versatile.
• Real-World Applications: The model could be applied to real-world projects such as translating educational materials, public health information, or community communication platforms.

Acknowledgements

This model was developed based on the Dholuo–Swahili parallel corpus created by Mbogho et al.\ (2025) as part of their work in building low-resource African language corpora. The corpus was made publicly available on platforms like Zenodo and Mozilla Common Voice.

How to Use

You can access the model via the Hugging Face Hub at:
https://huggingface.co/SalomonMetre13/nllb-luo-swa-mt-v1

To load the model using the Hugging Face transformers library, use the following code:

from transformers import pipeline

translator = pipeline("translation", model="SalomonMetre13/nllb-luo-swa-mt-v1")
translation = translator("Jajuok nomaki nyoro gotieno")
print(translation)