Our mission
Make under-resourced languages as powerful as English for LLMs.
What we believe
AI is the most revolutionary technology of this generation.
If it only works well in the languages that dominate the internet, its benefits go to a fraction of the world. Making models perform as well in under-resourced languages as they do in English is key to equitable use, and equitable distribution, of this technology.
The industry is focused on English and the few languages that fill the web. We focus on the ones that don’t.
The problem
Across every training phase, models have over-indexed on English. The frontier labs are based in English-speaking countries, and English dominates both the web and the datasets models are trained on. Other languages perform worse as a result.
For languages with a limited online presence, the gaps are clear: models miss dialects, misread culture and fall short in specialist domains. That keeps the people who speak them from getting the full value of LLMs that English speakers already get. See the gap →
Our approach
Native, not translated
Data written by people who live in the language, rather than English content translated into it.
Dialect-specific
A language is many languages. We capture how it’s really spoken, region by region.
Expert-led
Domain experts who work in the language judge quality, so the data reflects real expertise.
Filling the shelves
- NowالعربيةArabicModern Standard Arabic and the major regional dialects.
- NextMore under-resourced languagesLanguages spoken by millions and rarely written online.
- AlwaysLanguages with little online presenceThe languages with the least data deserve the most attention.
Why “Logha”
Logha is the Arabic word for language. Arabic is the first language on our shelf, and it is the one our name comes from.
Our work is to give under-resourced languages, written by the people who speak them, the same place in AI that English has today.
لغة