Questions to answer: How to gather training data for the IDM? How much data is enough to train something meaningful? How to annotate our data?