M ANDOLINE: Model Evaluation under Distribution Shift

     Mayee Chen * 1 Karan Goel * 1 Nimit Sohoni * 2 Fait Poms 1 Kayvon Fatahalian 1 Christopher Re 1

              Abstract                tioners to determine if their models will perform well when
                                  deployed. Unfortunately, standard evaluation falls short
    Machine learning models are often deployed in
                                  of this goal on two counts. First, evaluation data is fre-
    differ ...