Has it been trained on enough reverse text training data, or is this it actually reversing the text and parsing it backwards? If so - why are the answers so odd?