The SVC in the sklearn.svm sub-module claims to use “a squared l2 penalty”. However, the mathematical formulation in the user guide describes an L1 penalty term. Unless I have misunderstood the writing, there must be an error among the two.
Since the implementation claims to use the libsvm, which has similar documentation with linear slack terms, I am assuming the SVC is actually L1 instead of L2 as suggested by the API documentation.