I. Introduction
We define the context as the set of elements of the data which are needed to solve the problem defined by that data. In the case of classification problems represented by real-life benchmark data sets one can say that given data set includes the whole context if he will find an algorithm that solves that problem with 100% accuracy by analyzing only that data. But if we don't know such algorithm, we also do not know what portion of the context exists within the data. This is because we can't say if limited accuracy of ML models, in the case of given data set, is caused by: lack of some part of the context within the data, or by the fact that the data includes all the information needed to solve the problem, but known training algorithms can't find and/or properly use it. In such situation till we don't know the perfect solution, we don't know if it can exist or how close known methods are to the best possible solution. This makes most of real-life benchmark data sets not the best for evaluation of properties of ML algorithms.