This is the 102 language dataset they used for evaluation: https://huggingface.co/datasets/google/xtreme_s