Google

In a development set with 1,000 examples, a classifier with 98.0% accuracy makes about 20 mistakes and gets about 980 examples right. Because it is much faster to inspect the 20 mistakes than to audit all 980 correct predictions, teams often revise only the mislabeled cases. That convenience can introduce label bias into the development set.

Reviewing Only Mistakes Can Skew a Dev Set

What is the main reason error-checking can become biased when revising labels in a dev set?

Reviewing the Mistakes

Counting Misclassifications in a Development Set

Match Each Validation-Set Quantity to Its Meaning

Question:
A dev set contains 800 examples, and the model reaches 96.0% accuracy. Explain why teams often end up correcting labels mainly among the 32 examples the model misses instead of checking all 768 examples it classifies correctly.

Sample answer:
At 96.0% accuracy, there are 32 mistakes and 768 correct predictions. It is much faster to inspect 32 cases than to review 768, so people naturally start with the smaller error set. That behavior is driven by time and effort, not by a deliberate decision that the error set is the only place worth checking. If corrections are made only in the smaller group, the dev set can become biased because many label problems in the larger correct group are never examined.

Key points:
- 800 examples at 96.0% accuracy means 32 misclassified and 768 correctly classified examples
- Reviewing 32 examples is far less work than reviewing 768
- The pattern comes from convenience, not an intentional labeling policy
- Correctly classified examples can still contain label errors that remain undiscovered

Rubric:
Full credit requires stating the 32 versus 768 split, explaining why the smaller group is easier to inspect, and noting that the resulting bias is an unintended consequence of convenience.

Explain why label checks often get focused on the examples a model gets wrong.

Case context:
A product team maintains a validation set of 2,400 labeled examples. Their classifier reaches 97.5% accuracy, so about 60 examples are predicted incorrectly. To speed up review, the team checks and fixes labels only for those 60 error cases and does not inspect the remaining 2,340 predictions.

Question:
What bias does this create in the validation set, and what is the reason?

Sample answer:
This approach creates label-correction bias. The examples that the model gets wrong are the only ones that receive a label audit, so any labeling mistakes among them are more likely to be found and repaired. The much larger group of correctly predicted examples is left alone, even though some of those labels could still be wrong. Because the review effort is driven by convenience and limited time, the corrections are applied unevenly instead of being distributed across the full set.

Key points:
- Only the 60 error cases are checked for label problems
- The 2,340 predicted-correct cases may still contain bad labels
- The review strategy is chosen for efficiency, not to sample the whole set fairly
- The result is uneven correction across the validation data, which introduces bias

Rubric:
Full credit requires stating that only the mispredicted subset is reviewed and corrected, that the unreviewed correctly predicted subset may still have label mistakes, and that the unequal review pattern arises from practical convenience rather than a balanced auditing plan.

Explain the bias created when a team inspects only the validation errors.

Question:
In one to three sentences, explain why it is usually more efficient to examine only the wrong predictions in a development set instead of checking every prediction.

Sample answer:
If a model is correct on most examples, the wrong predictions are a much smaller set to inspect. For example, if a system gets 97% accuracy on 2,500 items, only 75 items are incorrect, so reviewing those 75 takes far less time than reviewing the 2,425 correct ones.

Key points:
- A small number of errors is much easier to review than the full set.
- Less reviewing time means the analysis is more practical.

Rubric:
Full credit requires noting the small error count compared with the much larger number of correct cases and explaining that this reduces review effort.

Learn Before

Related