Cheetah is a multilingual natural language generation (NLG) model designed for 517 African languages, addressing the scarcity of resources in these languages. Through extensive evaluations across multiple tasks, Cheetah outperforms existing models in generating coherent and contextually relevant text.

This research highlights the importance of adapting pretrained models to low-resource languages, fostering linguistic diversity and inclusion in NLP. Cheetah's contributions pave the way for improved accessibility and real-world applications for African communities.

AfroLingu-MT is a benchmark for machine translation in low-resource African languages. It includes 43 datasets spanning 84 language pairs from 46 languages, covering diverse linguistic families across 29 African countries.

The dataset features a newly translated government-domain corpus and incorporates widely spoken African languages alongside English, French, and Arabic to support multilingual NLP advancements.

spBLEU^1K is a translation evaluation metric designed to improve on spBLEU by covering 1,003 languages, including 614 African languages. This expansion ensures better representation of low-resource languages in machine translation benchmarks.

Built on SentencePiece tokenization, spBLEU^1K mitigates issues in traditional BLEU scoring, such as tokenization bias and artificially inflated scores. Our approach includes temperature upsampling to enhance the representation of low-resource languages and integrates the model directly into SacreBLEU for seamless benchmarking.

SERENGETI is a suite of pretrained language models designed for African NLP, covering 517 African languages. It achieves state-of-the-art performance across 8 NLU tasks and 20 datasets, and outperforms existing multilingual models on low-resource settings.

Based on Electra and XLM-R architectures, SERENGETI is trained on 42GB of diverse, multi-script, multi-domain text. It promotes inclusivity, supports language preservation, and provides a new foundation for Afrocentric NLP research.

AfroLID is a Neural Language Identification (LID) tool covering 517 African languages and varieties. It is trained on multi-domain web data spanning 14 language families and five different orthographic systems.

AfroLID achieves an F₁-score of 95.89% on a blind test set and outperforms five existing LID tools in recognizing African languages. Additionally, it has been successfully tested on real-world social media data (e.g., Twitter), demonstrating its effectiveness in low-resource settings.