Add MLQA

[KahnSvaer]

• add mlqa_common_yaml

• add 49 tests of mlqa family

• update tasks/README.md

---

resolved: #192

[KahnSvaer]

I am not sure about the doc_to_text prompt that I wrote. I have checked that the code works with dummy model and that the process_result function is working as intended. I have also checked that it works with pythia model though I was not able to completely evaluate it on the dataset due to gpu constraints. Most of the code in tasks/mlqa/utils.py is directly taken from the main repo: https://github.com/facebookresearch/MLQA

[KahnSvaer]

Found the error sending the updated commit as soon as possible

[baberabb]

Hi @KahnSvaer, this is great! small nit: the test failing as it's trying to parse mlqa_common_yaml.yaml. just need to remove the file extension and should be good to go.

Have you come across any papers which test these tasks on any transformer models? I can try to reproduce. But no worries if not.

[KahnSvaer]

@baberabb The original paper for this benchmark https://arxiv.org/pdf/1910.07475 shows some tests on transformer models like M-Bert and XLM.

[baberabb]

LGTM! thanks for the PR!