Yo-ByT5 achieves high-fidelity Yorùbá diacritic restoration
October 2, 2026
Yo-ByT5, a byte-level model fine-tuned from ByT5-small, matches mT5-base performance on the YAD benchmark with a 10.14% DER and 3.48% CER. The model achieves this while using approximately half the parameter count of its larger counterpart.
HOW THIS AFFECTS YOU
●
builderYou can deploy a much smaller, more efficient model for Yorùbá NLP tasks without losing accuracy.
●
researcherNote the effectiveness of byte-level fine-tuning for low-resource tonal languages.