The model might perform well, but without a robust cross-validation schema, it's hard to be sure. Given the small dataset, careful target encoding could help extract signal. Dropping three columns might be risky, every feature counts when data is limited.