sts22-language-pairs

IN premise — summaries/2026/08/24/muennighoff-2022-mteb-s4-additional-modalities-text-embeddings-are.md

Created 2026-08-25T02:58:22+00:00

STS22 in MTEB covers 18 crosslingual language pairs

Summary

The STS22 benchmark in the MTEB suite tests whether a model can recognize that two sentences mean the same thing even when they are in different languages, and it does so across exactly 18 specific language pairings. This sets a hard boundary on which crosslingual combinations are actually covered by any evaluation that relies on this dataset.