Update tatoeba to v2021-07-22

[KoichiYasuoka]

Tatoeba's latest version is v2021-07-22

[KoichiYasuoka]

How about this? @lhoestq @abhishekkrthakur

[lhoestq]

Hi ! I think it would be nice if people could still be able to load the old version.
Maybe this can be a parameter ? For example to load the old version they could do

load_dataset("tatoeba", lang1="en", lang2="mr", date="v2020-11-09")

If it sounds good to you, we can add this parameter to the TatoebaConfig:

class TatoebaConfig(datasets.BuilderConfig):
    def __init__(self, *args, lang1=None, lang2=None, date="v2021-07-22", **kwargs):
    self.date = date

and then pass the date to the URL

_BASE_URL = "https://object.pouta.csc.fi/OPUS-Tatoeba/{}/moses/{}-{}.txt.zip"

        def _base_url(lang1, lang2, date):
            return _BASE_URL.format(date, lang1, lang2)

What do you think ?

[KoichiYasuoka]

_DATE = "v" + "-".join(s.zfill(2) for s in _VERSION.split(".")) seems rather tricky but works well. How about this? @lhoestq

[lhoestq]

Looks all good thanks !

I just renamed the parent directory of the dummy data to match the new version name, and mentioned how to change the date in the dataset card.

[lhoestq]

The CI is only failing because of the missing sections in the dataset card, and because of an issue with the CER metric that is unrelated to this PR