Case context: You are training a speech-command classifier for smart earbuds. Your development and test sets contain 4,000 recordings captured on the earbuds themselves. You also have 120,000 extra clips collected from public audio datasets. Your team can only train a small model because of tight memory and compute limits.

Question: How should you use the public audio clips during training, and why does that choice help keep the model small?

Sample answer: Give the 120,000 public clips a much smaller training weight than the 4,000 earbud recordings. The public clips are far more numerous and come from a different distribution, so treating them the same as the target data would push the model to spend capacity learning both domains. Down-weighting the public clips keeps the optimization centered on the earbud distribution and avoids the need for a much larger network.

Key points:

• Assign a lower weight to the auxiliary public audio clips.
• They are much more numerous and come from a different distribution.
• Equal weighting would force the model to represent both domains well.
• Down-weighting them helps keep the network smaller and focused on the target domain.

Rubric: The answer must recommend down-weighting the public clips and explain that this reduces the capacity needed to model the auxiliary distribution, which helps fit the compute budget.