OK, it seems that in Segment trainer and Nested Trainer, in order to get rid of overflows all metrics are calculated in a log space. Which means that I can't define features that return negative values. I can shift the features but then they wont represent the underlying characteristics correctly. Is there any workaround ?