- hellaswag for te/th
- xquad tasks for th/ar/ru/arcd
- fixed arabic_mmlu
- xnli2.0
- fixed ocnli
- fix arabic hellaswag (empty choices)
- add token_norm metric to xwinograde + correct prompt
- for hellaswag correctly format the the activity label
- separate space into word_space and setence_space (which fully affects zh/th, other languages should be uefected)
- add space for all generative tasks at the end of query (affects all generative tasks)
- small nits
