A very naive bert/roberta baseline for Hotpot QA (ditractor settings). The model follows select-and-anser style. First, we use a Roberta classifier to select 2 gold pragraphs. Next, we use a Roberta QA model to answer the question. It serves as the base Hotpot model for Evaluating Explanations for Reading Comprehension with Realistic Counterfactuals.