Today, Meta is releasing two papers with my contributions: Omnilingual Machine Translation and Omnilingual SONAR. They describe a family of models designed to work with ANY language. We are not opensourcing the models yet, but releasing an update to the BOUQuET dataset, with 275 language varieties.

Omnilingual MT and SONAR papers

Today, Meta is releasing two papers with my contributions: Omnilingual Machine Translation and Omnilingual SONAR. This is the main project that I’ve been busy with throughout the last year.

open.substack.com

We (oldi.org) recently released version 3.0 of the FLORES+ dataset: a benchmark for multilingual machine translation. In this version, we added Ladin language (now there are 222 language varieties in the dataset!), corrected the spelling for Chuvash and Dargwa, and fixed sentence order in Aranese.