mBERT base uncased model trained on 50% SQUAD data. This model can further be used to fine-tune using dev data for QA system on a specific language. The process is similar to what was followed in MLQA paper[https://aclanthology.org/2020.acl-main.421.pdf].
