Add SubjQA dataset

[lewtun]

Hello datasetters 🙂!

Here's an interesting dataset about extractive question-answering on subjective product / restaurant reviews. It's quite challenging for models fine-tuned on SQuAD and provides a nice example of domain adaptation (i.e. fine-tuning a SQuAD model on this domain gives better performance).

I found a bug in the start/end indices that I've proposed a fix for here: megagonlabs/SubjQA#2

Unfortunately, the dataset creators are unresponsive, so for now I am using my fork as the source. Will update the URL if/when the creators respond.

[lewtun]

I'm not sure why the windows test fails, but looking at the logs it looks like some caching issue on one of the metrics ... maybe re-run and 🤞 ?

[lhoestq]

Awesome thank you !

Could you also try to reduce the size of the dummy data zip files ? Currently they're quite big (50KB+ each). To do so feel free to take a look inside the books dummy data zip file and remove all the csv files from the other domains, and only keep the files from the books domain. Thanks :)

[yjernite]

Hi @lewtun, thanks for adding this dataset!

If the dataset is going to be referenced heavily, I think it's worth spending some time to make the dataset card really great :) To start, the information that is currently in the Data collection paragraph should probably be organized in the Dataset Creation section.

Here's a link to the relevant section of the guide, let me know if you have any questions!

[lewtun]

If the dataset is going to be referenced heavily, I think it's worth spending some time to make the dataset card really great :) To start, the information that is currently in the Data collection paragraph should probably be organized in the Dataset Creation section.

great idea @yjernite! i've added some extra information / moved things as you suggest and will wrap up the rest tomorrow :)

[lewtun]

hi @yjernite and @lhoestq, i've fleshed out the dataset card and think this is now ready for another round of review!

[lhoestq]

Thanks !
I added my final comments:

[lhoestq]

LGTM thank you !

Merging since the CI error is fixed on master