Methods and systems are disclosed herein for accurately training a machine learning model with a reduced training data set. A large number of data records may be parsed. Each record may be reduced to a set of symbols representing the composition of each record. A user may assign a classification to each symbol within each record. Records with identical arrangements and classifications of symbols may be grouped together, and a representative sample of data records from each group may be fed into the model as the reduced training data set.