From the docs I feel it is not entirely clear that using label indicators means doing multi-label classification.
I think this should be better documented in the multi-class narratives, the OneVsRestClassifier and the LabelBinarizer.
Applying a label-binarizer to your y will result in quite different results, which is obvious if you know that this will switch the problem to a multi-label problem, but is somewhat unintuitive for the uninitiated.