Sentence by Sentence

Everlyn Asiko Chimoto, PhD, the arithmetic of enough, and what a community builds when it builds together

Masakhane means we build together. It is a good name, and it is a promise, and every so often someone comes along who keeps it thoroughly.

This is a feature about one such Masakhane Community Member, Everlyn Asiko Chimoto, and it is a joy to write.

Start where she started: with eight thousand sentences.

In 2022, Everlyn and her supervisor Prof Bruce Bassett set out to teach a machine to mine parallel text for Luhya; that is look at a Luhya sentence and find its English twin. The best multilingual models in the world had never really met the language. LASER aligned Luhya correctly 1.5 percent of the time. LaBSE, the stronger of the two, managed 22.

Through creating, roughly 8,000 Luhya–English sentence pairs, gathered and paired by hand; which is to say, hours and hours of patient, unglamorous, deeply motivated work; she fine-tuned LaBSE on a small parallel set and watched the number climb: 53 percent, then past 85 once filtered by similarity threshold. From a standing start to a working resource, in a single paper at LoResMT 2022.

Everlyn does not wait for the conditions to improve. She makes the thing that improves them, and then she turns around and enables access to the broader community.

THE ARITHMETIC OF ENOUGH

Everlyn describes her research as data-efficient and model-efficient techniques for low-resource languages: getting the most out of limited data and limited budgets. In a field that likes to measure progress in billions of parameters, that is a quietly radical bet, a bet that our languages do not have to wait for abundance to be served competently and beautifully. That what we already have is enough to build with, if we are resourceful, innovative, careful and willing to do the work.

Everlyn has been winning that bet, paper by paper.

Critical Learning Periods: Leveraging Early Training Dynamics for Efficient Data Pruning (Findings of ACL 2024), written with Jay Gala, Orevaoghene Ahia, Julia Kreutzer, Bruce Bassett and Sara Hooker, is the clearest demonstration. It introduces Checkpoints Across Time (CAT) which watches how a model behaves in its first few hours of training to work out which examples genuinely matter. Half the training data can then be set aside, and the model does about as well as if it had kept everything. But the best part is what CAT reaches for when left to decide: longer sentences, and sentences carrying rare and unusual words. The difficult, textured, particular material that a blunter tool would sweep away first. Anyone who has ever tried to explain why a proverb is hard to translate will feel something warm at that finding.

The rare thing is not noise. The rare thing is where the meaning lives.

She has carried that same instinct into the era of large models. GrammaMT (ACL 2025, with Rita Ramos, Maartje ter Hoeve and Natalie Schluter) improves translation by feeding models grammar-informed examples, using linguistic structure where others would simply ask for more data. Calibrating Beyond English: Language Diversity for Better Quantized Multilingual LLMs (EACL 2026, with Mostafa Elhoushi and Bruce Bassett) asks a question that matters to anyone trying to run a model on the infrastructure we actually have: when you compress a model to make it small enough and cheap enough to be useful here, how do you make sure our languages come through the squeeze whole? Her answer: choosing your calibration data with language diversity in mind is the sort of finding that quietly makes better technology possible for millions of people who may not read the paper but experience more inclusive technology on account of it.

This is what efficiency research looks like when someone does it out of care; A door held open. Community at the fore.

WE BUILD TOGETHER

Everlyn’s path runs straight through the institutions our community built for itself, and she has given something back to every single one.

A BSc in Computer Science, First Class Honours, from Masinde Muliro University of Science and Technology in Kakamega in the heart of Luhya-speaking Western Kenya. An MSc in Mathematical Sciences, with Distinction, from AIMS South Africa and the University of Cape Town, on a Paul G. Allen Family Foundation Scholarship. Doctoral work at UCT with Bruce Bassett, through Quantum Leap Africa. A research internship at Apple’s machine learning lab in 2024. The Heidelberg Laureate Forum and Imperial’s Global Fellows Programme in 2023. In January 2025, Lelapa AI, as a fundamental research scientist. And in March 2026 her PhD.

Read as a list, it is a fine career. Read carefully, it is a person taking every opportunity she earns and gently converting it into a room that other people can walk into, building for the community.

Inside our Masakhane community, she has co-organised the AfricaNLP workshop in 2023, 2025 and 2026, and co-edited the AfricaNLP 2026 proceedings with Constantine Lignos, Shamsuddeen Muhammad, Idris Abdulmumin, Clemencia Siro and David Ifeoluwa Adelani. She ran the NLP@Indaba workshop and taught an “LLMs for everyone” practical at the Deep Learning Indaba in 2023 putting the newest tools in the world directly into the hands of people who had never touched them, which is a very Everlyn thing to do. In September 2025 she came together with a team and convened the first-ever KenyaNLP meeting: she noticed a room that did not exist, and she built it.

She was also part of the Esethu Framework (ACL 2025), a big, generous collaboration rethinking how datasets in low-resource languages are governed and curated, so that the communities who give their language keep a real stake in whatever is made from it. Data as a relationship rather than raw material. That is participatory research in the truest Masakhane tradition.

That is how a field is actually built. Not by any single result. By people who keep making space and keep saying: come in, there is room, you belong here.

A VERY GOOD YEAR

In July 2026, the Deep Learning Indaba awarded Everlyn the Thamsanqa Kambule Doctoral Award; its recognition of excellence in doctoral research at African universities in the computational and statistical sciences, named for the South African mathematician and headmaster who spent his life defending other people’s right to learn. It is hard to imagine a better fit.

Her thesis was celebrated for precisely the work this community needs: sentence alignment for the Marama dialect of Luhya, data pruning reaching 92 percent of full performance on half the data, and grammar-informed translation for endangered languages.

And characteristically, when she spoke about the award, she spoke about everyone else:

“Receiving this award is deeply meaningful because it recognises not only the years of work invested in this thesis, but also the support of my supervisor, collaborators, family, and the African NLP community that has continually inspired and challenged me.”

— Everlyn Asiko Chimoto, on receiving the 2026 Kambule Doctoral Award

Inspired and challenged. She named both, and both are true of us at our best. We push each other because we believe in what the other person can do. There is no higher form of care in research than community and collaboration.

WHAT WE ARE BUILDING

Not so long ago, the idea that Kiswahili, Yoruba, isiZulu, Bambara and hundreds of African languages would have serious research programmes, real datasets, published methods, their own workshops and award-winning doctoral theses behind them was a hope held by a small number of motivated, hardworking people. Now; There is Masakhane, an AfricaNLP workshop with proceedings. There is a KenyaNLP, Ethio NLP, and many other great collectives mobilising across Africa. There is a framework for governing our own data on our own terms. There is a community. There are researchers all across the continent building models sized for the infrastructure we actually have, for the languages we actually speak.

That is what we have been doing all along. Not making the machines bigger.

Teaching them to hear us.

Everlyn Asiko Chimoto is a research scientist at Lelapa AI, a member of the Masakhane community, and the 2026 recipient of the Deep Learning Indaba’s Thamsanqa Kambule Doctoral Award. Her work is at everlynasiko.github.io.

S O U R C E S

Every citation in this feature is drawn from the public record, Everlyn’s Masakhane Contributions, ArchivX, ACLAnthology. The words attributed to Dr Chimoto are reproduced from the Deep Learning Indaba’s 2026 awards announcement.

• Everlyn Asiko Chimoto — personal site. https://everlynasiko.github.io/

• Deep Learning Indaba, 2026 Awards Announcement. https://deeplearningindaba.com/blog/2026/07/2026-awards-announcement/

• ACL Anthology — publications by Everlyn Asiko Chimoto. https://aclanthology.org/people/everlyn-asiko-chimoto/

• Very Low Resource Sentence Alignment: Luhya and Swahili (LoResMT 2022). https://arxiv.org/abs/2211.00046

• Critical Learning Periods: Leveraging Early Training Dynamics for Efficient Data Pruning (Findings of ACL 2024). https://arxiv.org/abs/2405.19462

• The Esethu Framework: Reimagining Sustainable Dataset Governance and Curation for Low-Resource Languages (ACL 2025). https://aclanthology.org/2025.acl-long.1487/

• Quantum Leap Africa — Everlyn Asiko Chimoto. https://qla.aimsric.org/?peoples=everlyn-asiko-chimoto-kenya

• Stimson Center — Everlyn Asiko Chimoto. https://www.stimson.org/ppl/everlyn-asiko-chimoto/

• Masakhane — a grassroots NLP community for Africa, by Africans. https://www.masakhane.io/home

• State of NLP in Kenya: A Survey. https://arxiv.org/abs/2410.09948