Sample questions
Ml Model Training EvaluationDifficulty 1
In machine learning, what does it mean for a model to 'overfit' the training data?
- aThe model fails to capture the underlying pattern even in the training data, producing high error there too.
- bThe model's predictions on unseen data become consistently better than on the data it was trained on.
- cThe model requires more training data because it has too few parameters to represent the pattern.
- dThe model fits noise specific to the training set, so it performs poorly on new data.✓
Explanation:Overfitting means the model captured patterns specific to the training sample (including noise) rather than the true underlying relationship, so it fails to generalize to new data. Option a describes underfitting, and b/c describe the opposite of what actually happens when a model overfits.
Ml Model Training EvaluationDifficulty 1
What is 'underfitting' in the context of a trained model?
- aThe model is too simple to capture the pattern, performing poorly on both training and unseen data.✓
- bThe model performs extremely well on training data but poorly on unseen data.
- cThe model was trained on too much data, which slows down convergence during training.
- dThe model achieves near-perfect accuracy on the test set but noticeably lower accuracy on the training set.
Explanation:Underfitting happens when a model is not expressive enough (or has not learned enough) to capture the real signal, so error stays high everywhere, including on the training data it was fit on. Option b describes overfitting instead.
Ml Model Training EvaluationDifficulty 1
Why do we split a dataset into a training set and a separate test set before evaluating a model?
- aTo reduce the total amount of data the model has to process, which saves compute time during training.
- bBecause most training algorithms technically cannot process more than a fixed number of rows at once.
- cTo estimate performance on unseen data, since evaluating on training data alone can hide overfitting.✓
- dTo guarantee that every class label appears an equal number of times in the dataset.
Explanation:A held-out test set that the model never trains on is the standard way to estimate real-world generalization; evaluating only on training data would hide overfitting and give an overly optimistic score. Options a, b, and d are not the actual reasons for the split.
Ml Model Training EvaluationDifficulty 1
In a binary classification confusion matrix, what does a 'false positive' represent?
- aA case where the model correctly predicted the negative class.
- bA case where the model predicted the positive class, but the true label was actually negative.✓
- cA case where the model predicted the negative class, but the true label was actually positive.
- dA case where the model correctly predicted the positive class.
Explanation:A false positive is a false alarm: the model says 'positive' when the actual label is 'negative'. Option c describes a false negative (a missed positive case), while a and d describe correct predictions.
Ml Model Training EvaluationDifficulty 1
Why can accuracy alone be a misleading metric when a dataset is highly imbalanced (e.g., 99% negative, 1% positive)?
- aAlways predicting the majority class can yield very high accuracy while missing the minority class entirely.✓
- bAccuracy cannot be computed at all when classes are imbalanced, since the formula requires an equal count of each class.
- cAccuracy only applies to regression problems, not to classification problems.
- dAccuracy penalizes the model more heavily for errors on the majority class than for errors on the minority class.
Explanation:With a 99/1 split, a trivial 'always predict negative' model scores 99% accuracy while being useless for the minority class, which is why metrics like precision, recall, or PR-AUC are usually reported alongside accuracy on imbalanced data. Options b, c, and d misdescribe how accuracy is defined.
Ml Model Training EvaluationDifficulty 1
What does 'early stopping' mean during model training?
- aReducing the size of the training dataset before training begins so the process finishes faster.
- bSkipping the validation step entirely to speed up each training iteration.
- cHalting training at a fixed number of iterations decided before the dataset was even inspected.
- dStopping training once the validation metric stops improving, to avoid overfitting.✓
Explanation:Early stopping watches a validation metric during training and halts once it stops improving (or starts degrading), which prevents the model from continuing to fit noise in the training set. Options a, b, and c describe unrelated or fixed procedures that don't react to validation performance.