Question: A classifier has a 10% error rate on the training set, an 11% error rate on a separate training-dev set, and a 12% error rate on the development set. What does this pattern suggest is the model’s main problem, and why?

Sample answer: The model is mainly limited by bias. The training error is already fairly large, and the training-dev and development errors are only slightly higher. That tells us the model is not fitting the training data well enough, so the first issue to address is underfitting or avoidable bias rather than variance.

Key points:

• Identifies bias, especially avoidable bias, as the dominant issue.
• Explains that the model is doing poorly even on the training data.
• Uses the small increases from training to training-dev to development error to show that variance is not the main bottleneck.

Rubric: Answers should identify high bias or avoidable bias and explain that the core weakness is poor performance on the training set, noting that the small differences among the three error rates indicate variance is a secondary concern.