This is probably not entirely invalidating the result, but the language samples in the dataset seem to be extremely badly translated from english, with unnatura

This is probably not entirely invalidating the result, but the language samples in the dataset seem to be extremely badly translated from english, with unnatural, verbose and grammatically wrong sentences. That would not help with good tokenisation. For example the english text > please add milk to the grocery list Is compared to the french text > s'il vous plaît ajouter du lait à la liste d' épicerie But a native would say > veuillez ajouter du lait à la liste de courses