These small models, having been fine-tuned for the test, achieve frighteningly high scores, yet perform abysmally in real-world scenarios.