wer ?

#3
by flutter-painter - opened

Thanks for providing this.
How good is it ?
What is the difference with the rel model ?

What dataset did you use ?
https://huggingface.co/datasets/masakhane/mafand

Masakhane NLP org

Heyy @flutter-painter ,

Check out the paper here:: https://arxiv.org/abs/2205.02022

Thanks @ToluClassics ,
Fascinating !

I understand that REL stands for religious dataset,
I gather that the best outputs for Wolof are :

  • fr -> wolof : BLEU 12.7 / 20.2 CHRF
  • wolof -> fr : BLEU 11.5 / 35.3 CHRF

Similar to https://huggingface.co/facebook/nllb-200-distilled-600M, which claims

eng - wol : 23.5 CHRF
wol -eng : 36.7 CHRF

flutter-painter changed discussion status to closed

Sign up or log in to comment