crepes 0.9.1

v0.9.1 (12/6/2026)

Features

• The method predict_set of WrapClassifier now includes an argument labels, which if set to True (the default), outputs prediction sets in the form of a list of lists of labels, relying on the labels being provided in the attribute classes_ of the wrapped learner. Alternatively, label names can be specified by providing a list or array of labels (overriding any label names provided by classes_). If instead a binary array indicating presence of labels in each prediction set is desired, labels should set to False. Note that the default behavior of predict_set of WrapClassifier is different from in previous versions, where the only option was to get prediction sets in the form of a binary array.

• The method predict_set of ConformalClassifier now includes an argument labels, which if set to None (the default), generates prediction sets in the form of a binary array. If instead labels is set to a list or array of labels, the output prediction sets will be represented by a list of lists of labels.

• The method semi_online_p_values in the module crepes.martingales for generating (smoothed) p-values from non-conformity scores in the semi-online mode has been optimized for non-unique non-conformity scores, resulting in O(n * k) where k is number of unique values among the n non-conformity scores. It is enabled through the option optimized=True. Thanks to @claude for contributing with the optimization.

• The option return_index is now available for the apply methods of SimpleJumper, SleeperStayer, SleeperDrifter and CompositeMartingale in the module crepes.martingales, which if set to True, results in that the index of the first martingale value that reaches a specified threshold (c) is returned. By setting return_index to False, martingale values up to and including the first that reaches a specified threshold will instead be returned. In the previous version, an index was always returned when a threshold was specified.

• The apply method of the CompositeMartingale has been optimized by terminating the computation of each of the k martingales when reaching k * c, for a specified threshold c.

• The binning method in crepes.extras has been modified to allow for setting a factor (epsilon) to use when multiplying with a random number to handle ties. By setting epsilon to 0, random tiebreaking is not employed. The binning method now also includes an additional argument min_size (default=None), which results in equal-sized binning with the largest number of bins such that the minimum number of values in each bin is at least min_size. A warning will be issued when forming bins if the provided number of bins is larger than the number of values (or the number of unique values if epsilon=0), as well as if min_size allows for forming one bin only. Thanks to @rudymatela for highlighting issues with the binning method.

Fixes

• A minor bug in the documentation has been fixed. Thanks to @KonstantinosTsoumas for pointing it out.