We offer one explanation for the historically low performance of French in the SIGMORPHON-UniMorph shared tasks. We conducted experiments replicating the 2023 task on French with the non-neural and neural baselines, first using the original task splits, and then using splits that excluded Old and Middle French lemmas. We applied a taxonomy on our errors using a framework based on Kyle Gorman’s “Weird Inflects but OK” 2019 annotation scheme, finding that a high portion of the French errors produced with the original splits were due to the inclusion of Old French forms, which was resolved with cleaned data.