Need to figure out how to pick the FILLPRED attribute (using tokens with "modra") is not enough.  
We'll use a decision tree classifier to build a model based on the ssj500k mate data.