jhu-clsp / mmBERT-base

TL;DR: A state-of-the-art multilingual encoder trained on 3T+ tokens across 1800+ languages, introducing novel techniques for learning low-resource languages during the decay phase.