Question: A team trains a vehicle-noise detector using generated audio. In one setup, they duplicate the same 2-minute recording until they have 500 hours of training data. In another, they create 500 hours from recordings taken from only 12 different cars. Explain why the large total amount of audio does not by itself eliminate overfitting. What matters most?

Sample answer: A large training set does not automatically prevent overfitting if the underlying sources are too similar. In the first setup, the model keeps hearing the same 2-minute recording, so it can learn details specific to that recording even though the total hours look large. In the second setup, the audio may be unique in time, but it still comes from only 12 cars, so the model can learn the sound patterns of those few vehicles rather than the broader range of car noises it will face later. The key factor is not just how many hours of audio exist, but how many genuinely different source recordings or source vehicles were used to create them.

Key points:

• Reusing the same short recording many times can still let a model memorize its specific patterns
• A human may not notice the repetition, but the learning algorithm can
• Many hours from only a small set of cars can still cause the model to specialize to those cars
• Total duration alone is not enough to judge training diversity
• Diversity of the original sources is more important than the raw number of hours

Rubric: Full credit requires explaining both the repeated-recording case and the limited-cars case, and stating that source diversity matters more than total audio volume. Partial credit is appropriate if the response explains only one case or mentions diversity without clearly connecting it to overfitting.