Why it matters

Multilingual scaling shapes global LLMs. Understanding shapes design.

Advertisement

The architecture

Data mix per language.

Curriculum + weighting.

Cross-lingual transfer benefits.

Multilingual scalingLanguage mixtraining dataWeightingcurriculumCross-lingual transfershared repsFrontier models English-heavy but growing multilingual; NLLB-200 pushed 200 langs
Multilingual scaling.
Advertisement

How it works end to end

Positive cross-lingual transfer.

Curse of multilinguality above N languages.

Data mix weights (upsampling low-resource).