Need to verify the compatibility with scikit-learn cross-validation
It seems (on 2 tests) that the first element of the list given by cross_validation_score when a target is given (i.e. perplexity computation on the first folder) gives a value way smaller than on the other folder of the cross-validation.