Sometimes you have a predetermined subset of the data you want to use as testing or validation instead of letting a Partition Node randomly divide the data. If that is the case, don't use a Partition Node. Instead create a variable with three values (for training, testing and validation) or two values (for training and testing) and, in the Type Node declare its role to be Partition and its type to be Ordinal.
Modeler will then use the records with the smallest value for training, the records with the next one for testing and records with the highest value (if there are three) for validation.