Ahmad Samuel Gali

BSc Computer Science, University of Lagos

About

I study how to make language models smaller, cheaper to train and more reliable, with a focus on languages that have little data. I trained Yo-ByT5, a byte-level model for Yorùbá diacritic restoration, on a Google Cloud TPU under a TPU Research Cloud grant. It has the lowest character error rate of the seven systems compared, with about half the parameters of mT5-base. I also built AfriNORM, a benchmark that tests whether LLMs reason correctly about cultural norms when the same question is asked in a different language.

My research interests:

Tokenizer-free and byte-level language models
Efficient training and small models for many languages
Evaluation and benchmarks for LLMs

I am applying for PhD programmes that start in 2027.

News

Oct 2026Yo-ByT5 preprint released on arXiv.
Jan 2026Trust but Check accepted at the AfricaNLP Workshop, EACL 2026.
2025Received a Google TPU Research Cloud grant for sequence-to-sequence training in African-language NLP.

Publications

Yo-ByT5: Efficient and High-Fidelity Diacritic Restoration for Yorùbá

Ahmad Samuel Gali, Shamsuddeen Hassan Muhammad

arXiv preprint arXiv:2610.01634, 2026

Lowest character error rate (3.48%) of seven systems on the YAD benchmark. Also reports a Unicode inconsistency in the YAD test set.

arXiv model demo

Trust but Check: LLM-Assisted Review of Human Translations in African Languages

Tadesse Destaw Belay, Henok Biadglign Ademtew, Idris Abdulmumin, Sukairaj Hafiz Imam, Abubakar Juma Chilala, Godfred Agyapong, Chinedu Emmanuel Mbonu, Basil Friday Ovu, Catherine Nana Nyaah Essuman, Alfred Malengo Kondoro, Sonia Adhiambo, Daud Abolade, Ponts'o Mpholle, Nicholaus Dismas Ladislaus, Saminu Mohammad Aliyu, Gali Ahmad Samuel, Fabrice Hakuzimana, Mike Nzirainengwe, Temitayo Olatoye, Sileshi Bogale Haile, Tewodros Achamaleh, Tolulope Olalekan Abiola, Kedir Yassin Hussen, Ibrahim Said Ahmad, Verrah Akinyi Otiende, Seid Muhie Yimam, Shamsuddeen Hassan Muhammad

AfricaNLP Workshop at EACL 2026 (non-archival)

OpenReview

Presentations

Implementing the Agentic Salesman Digital Twin for Real-Time Transactions

Ahmad Samuel Gali, Ebun P. Fasina

Poster, 19th University of Lagos Annual Research Conference and Fair, 2026

Software and models

Yo-ByT5

Byte-level model that restores tone marks and underdots to Yorùbá text.

model demo

Dyon

Python framework (about 13,700 lines) for building agentic digital twins across domains.

PyPI code

Yorùbá AI-generated text detector

Transformer classifier that separates human-written Yorùbá from LLM output.

demo

Education

BSc Computer Science, University of Lagos 2021 – 2026 CGPA 4.49 / 5.00

Experience

Machine Learning Intern, Arewa Data Science Academy, HausaNLP Research Group 2025 Translation-quality review in 11 African languages. Sequence-to-sequence training on TPU v6e-8.

Awards

Google TPU Research Cloud grant 2025