Zovda
A model can be excellent in the language of the team that built it and mediocre in the language of half its users. Nobody notices, because the evaluation was written in the first one.
The gap is invisible until a customer complains in the other language, and by then it has been shipping for months.
Score every language you claim
Not a copy of the same fifty questions carried into another language, and not a machine-rendered version of them. Questions written by someone who would actually ask that in that language, about that product, in that market.
A rendered question tests the renderer.
Separate fluency from correctness
A model can produce fluent, grammatical text in a language where its facts are wrong. Score both, because a fluent wrong answer is more damaging than a clumsy right one.
Watch the register
Formality, pronouns, and politeness conventions differ, and a model trained mostly on one language tends to apply its habits to the others. That is a quality issue users hear immediately.
Report per language, every time
A single multilingual score is an average weighted by whichever language has the most rows. Publish the per-language numbers next to their row counts.
Zovda builds these collections through Datuva, where language and native status are separate fields. The audio collection applies the same discipline to speech.
Tell us which languages your users write in. We reply within one business day.