Gh3275: sample_missing_splits in SST-2

[plonerma]

See #3275

[alanakbik]

@plonerma this PR should make it so that GLUE datasets have the option of being loaded without sampling the test split?

I am getting an error when running this:

from flair.datasets import GLUE_SST2

# default setting: SST-2 with sampling missing splits
corpus = GLUE_SST2()
print(corpus)

# without splits
corpus = GLUE_SST2(sample_missing_splits=False)
print(corpus)

[plonerma]

My bad. I fixed the error: it should work now.

Some background:

(Most?) GLUE datasets do not have gold labels for the test split. Hence, in flair they are no treated as test-splits but as a separate split (validation) which is used separately to create a tsv file which can be submitted to the GLUE-portal.

Since the test-split is considered to be missing, by the default, flair samples a new test-split from the training data.

The naming clearly is a bit unfortunate then (as e.g. saying "SST2 does not have a test split" is wrong / a bit ambiguous), but functionality-wise imo it makes sense as the test-split is handled very differently (e.g. no final evaluation on that split). With the other GLUE datasets I implemented, I was following the path already taken in existing ones.

[alanakbik]

Thanks for fixing this @plonerma!