Selecting Fine-Tuning Examples by Quizzing VLMs