Training data used for xlm-roberta-base-squad2

[shalakasatheesh]

Hello,

Could you please tell me what German language datasets were used for training the QA model xlm-roberta-base-squad2?

In addition, I was wondering if the model used for fine-tuning is xlm-roberta-base?

Best wishes,
Shalaka

[vblagoje]

Shalaka, you have some of these details at https://huggingface.co/deepset/xlm-roberta-base-squad2
Regarding the dataset, my best guess would be a translation of Squad v2. If you need to know precisely, tell me why, and I'll ask around 😄

[shalakasatheesh]

thank you for the answer @vblagoje :)

As part of my thesis, I want to carry out robustness tests on a monolingual and a multilingual model for the task of German Question Answering. So yes, it would really helpful if I could get some more precise information :)

The monolingual model I am considering is also from deepset: gelectra-base-germanquad. I was able to find relevant training details about the model from this paper. Similarly, if I was wondering I could get some more information about xlm-roberta-base-squad2 so that I can use it as the multi-lingual model.

[vblagoje]

Very cool, I'll ping my colleagues @tholor @mathislucka who are better suited to answer this question.

[tholor]

@shalakasatheesh It's a long time ago, but pretty sure we just took the "xlm-roberta-base" model for it, fine-tuned it on the English Squad2.0 dataset and evaluated it on German MLQA and German XQuAD (as described in the HF model card). So the model has only seen multilingual data at pre-training time.

[shalakasatheesh]

thank you for the quick responses!