READER can answer "there is no answer" but not all models do

[guillim]

No description provided.

[guillim]

This PR simply suggest a new option for allowing SQuAD 1.0 like - model to be better configured into Haystack.

I have to say I struggled to understand why some of my returned answer array where partially empty. After some digging, I understood that default behavior was "READER SQuAD2.0" format. But that's not always the case, which explains this PR.

[tholor]

Good point. Thanks for raising the PR!
Did you also experience that with the FARMReader when loading a SQuAD 1.0 model? If yes, we should probably add a similar flag there, too. From our experience, the model in these cases simply never predicts a no_answer (start=0, end=0), but we also mostly use SQuAD 2 models so our experience here is limited.

[guillim]

I have to say, I never used FARMReader (at least yet) 🥴 so I cannot tell you more on this side.
I understand that you probably missed it since SQuAD2.0 are the most common, but on our side (https://huggingface.co/etalab-ia/camembert-base-squadFR-fquad-piaf) we are using a SQuAD1.0 style model.

[tholor]

Ok got it. Then let's merge this one for now and investigate it separately for FARMReader.

I would still recommend giving FARMReader a try at some point - it should be a bit faster, often gives slightly better results and you can load the same models from model hub (no conversion needed) :)