Question: Describe two situations in which it can make sense to assign lower weight to auxiliary training data, and explain why this can reduce the cost of building the model.

Sample answer: Lower weighting is worth considering when the auxiliary data comes from a distribution that is noticeably different from the development and test data, or when the auxiliary dataset is much larger than the data from the target distribution. In those cases, giving the auxiliary data full influence can force the system to spend a lot of capacity accommodating patterns that are not central to the main task. Reducing its weight helps the model focus more on the target data, which can lower the need for an extremely large network or other costly training setup and therefore save compute.

Key points:

• The auxiliary data distribution is unlike the dev/test distribution.
• The auxiliary data greatly outnumbers the target-distribution data.
• Lower weighting can reduce computational cost.
• It can avoid the need for an unnecessarily large neural network.

Rubric: A strong response identifies both conditions for lowering the weight of auxiliary data and explains that doing so can reduce the model size or training burden, which saves computational resources.