maroxtn/mt5-M2M-comparison

Comparing M2M and mT5 on a rare language pairs, blog post: medium.com # Comparing mT5 and M2M models In this project, I fine tuned both of mT5 and M2M on a small dataset (10k) of Yoruba - English sentence pairs, and compared the results. Both models were trained on multi-lingual data, and both are being used for translation. They were released during relatively the same, mT5 is a little bit bigger than M2M. I use both `fairseq` and `simpletransformers` for training and it took around two hours on a P100 GPU. TLDR; M2M is much better More details, blog post: medium.com If you find this useful please star the repo.