AWS Machine Learning Engineer – Associate (new) (MLA-C02): 5 Practice Questions on ML Model and Foundation Model (FM) Development
5 questions, one domain: ML Model and Foundation Model (FM) Development from the AWS Machine Learning Engineer – Associate (new) (MLA-C02) exam. It makes up about 24% of the exam. Pick your answer, then check it against the explanation.
AWS Machine Learning Engineer – Associate (new) (MLA-C02) · ML Model and Foundation Model (FM) Development · 5 questions with answers and explanations
📊 Set profile
Exam: MLA-C02 · 65 questions · 130 minutes · passing score 720
This set: 5 advanced · 2 scenario, 2 trade-off, 1 concept check
Topics: Amazon SageMaker, Amazon SageMaker Clarify, SageMaker Image Classification
🧭 How ML Model and Foundation Model (FM) Development is tested
Classic models and foundation models share this domain. A scenario sets a goal (adapt a large model on limited hardware, satisfy a regulator, avoid one costly kind of error, label images in a certain way) and asks which technique, metric or approach fits. Good answers respect the stated constraint; weak ones chase overall accuracy or set compliance aside.
Tips for this set
• Map the costly mistake to a metric. If false alarms hurt most, one metric matters; if missed cases hurt most, its counterpart does. Plain accuracy rarely settles the choice.
• When a large model cannot be fully fine-tuned on the hardware available, look at parameter-efficient methods that train small adapters on a frozen, possibly quantized, base.
• Strong overall results can hide a weak customer segment, and regulators may require explanations of single decisions. Check segment metrics and interpretability before picking a model on accuracy alone.
Question 1 of 5
A team must adapt a 13-billion-parameter open language model to its support-ticket style on SageMaker. Its budget allows one ml.g5.2xlarge instance with a single 24 GB GPU. Full fine-tuning runs out of GPU memory immediately. Which approach fits BEST?
Show answer
✅ Answer: B
Difficulty: advanced · Topic: Amazon SageMaker
Where fine-tuning memory goes
Full fine-tuning keeps, per parameter: the weight, its gradient and the optimizer state (Adam keeps two values). In mixed precision this is roughly 16 bytes per parameter, so a 13B model needs around 200 GB before counting activations. One 24 GB GPU cannot hold it.
Parameter-efficient fine-tuning (PEFT)
• LoRA freezes the base weights and trains small low-rank matrices added to selected layers, often well under 1% of the parameters. Only those need gradients and optimizer state.
• QLoRA also loads the frozen base model in 4-bit precision: 13B x 0.5 bytes is about 6.5 GB, which leaves room for adapters and activations on a 24 GB GPU.
The result is a small adapter file that is loaded on top of the base model for inference, or merged into it.
Why each option works or fails
• ✗ A. Batch size 1 with gradient checkpointing: these reduce activation memory only; the weights, gradients and Adam states of 13B parameters still need far more than 24 GB.
• ✓ B. QLoRA (correct): quantizing the frozen base model and training only LoRA adapters cuts memory by an order of magnitude, so the 13B model fits on one 24 GB GPU while still learning the ticket style.
• ✗ C. Higher dropout: dropout zeroes activations randomly for regularization. All weights, gradients and optimizer states are still stored, so memory use does not drop.
• ✗ D. Pre-training from scratch: it needs vastly more data and compute than one GPU and a set of tickets, and it would discard the language ability of the existing model.
Question 2 of 5
A model performs well overall but poorly for one customer segment. What should the team do before deployment?
Show answer
✅ Answer: B
Difficulty: advanced · Topic: Amazon SageMaker Clarify
Why aggregate metrics are not enough
An overall metric is an average across all records. If one customer segment is small, its poor results barely move the average.
Example: 95% overall accuracy can hide 70% accuracy for a segment that makes up 5% of the data.
Segment-level evaluation
• Slice evaluation results by segment (region, age band, product line) and compute metrics for each slice.
• Investigate causes: too few training examples for the segment, missing or poor-quality features, labels that behave differently, or a distribution shift.
• Fix with more or better data for that segment, reweighting or resampling, new features, or a model that handles the segment better, then re-evaluate.
SageMaker Clarify can compute post-training bias metrics that compare model performance across groups, such as differences in accuracy or recall, to support this analysis.
Why each option works or fails
• ✗ A. Deploy anyway: the overall target is met only because the segment's poor performance is averaged away; those customers would still receive bad predictions in production.
• ✓ B. Evaluate per segment and fix the gaps (correct): measuring metrics for each segment exposes where the model fails, and addressing the underlying data or model causes improves it for those customers before they are affected.
• ✗ C. Raise that segment's threshold: a segment-specific threshold only shifts the trade-off between false positives and negatives; it does not repair the model's weak signal, and treating a group differently without analysis can create fairness and compliance issues.
• ✗ D. Remove the segment from evaluation: this inflates the reported score while leaving the model just as bad for those customers, hiding the problem rather than solving it.
Question 3 of 5
A team can use a simple interpretable model with slightly lower accuracy or a complex model with higher accuracy. Regulators require explanations of individual decisions. Which choice fits?
Show answer
✅ Answer: C
Difficulty: advanced
Interpretability versus accuracy
• Interpretable models such as linear or logistic regression, small decision trees or scorecards expose their logic directly: each feature's weight or split explains a decision.
• Complex models such as large gradient-boosted ensembles or deep networks are often more accurate but act as "black boxes".
When regulations require explanations for individual decisions (for example why a loan was denied), a team can:
• Prefer the interpretable model if the accuracy loss is small and acceptable, or
• Use the complex model together with robust post-hoc explanation tools, such as SHAP values from SageMaker Clarify, validated for consistency, documented, and reviewed as part of model governance.
Either way, the explanation requirement is treated as a hard constraint, not a trade-off against a few points of accuracy.
Why each option works or fails
• ✗ A. Most parameters: more parameters do not guarantee higher accuracy (they can overfit), and choosing by size ignores the explainability requirement.
• ✗ B. Avoid ML entirely: many models can be explained, either inherently or with tools like SHAP, so abandoning ML is unnecessary.
• ✓ C. (correct): Interpretable model or complex model with robust explanations: both paths meet the regulatory need to explain individual decisions, and the choice depends on how much accuracy the simpler model gives up.
• ✗ D. Complex model with no explanations: this directly violates the regulatory requirement, regardless of any accuracy gain.
Question 4 of 5
A spam filter must rarely flag legitimate emails as spam. Which metric matters most?
Show answer
✅ Answer: B
Difficulty: advanced
Which error hurts: false positives or false negatives
In a spam filter, "positive" means "spam".
• A false positive is a legitimate email sent to the spam folder: the user may miss an important message.
• A false negative is spam that reaches the inbox: annoying but less harmful.
Precision for the spam class = TP / (TP + FP). High precision means that when the filter says "spam", it is almost always right, so few legitimate emails are flagged.
Recall for the spam class = TP / (TP + FN) measures how much spam is caught.
Example: flagging 100 emails with 98 real spam gives 98% precision; the 2 wrongly flagged emails are the errors the business wants to minimize. Raising the decision threshold usually increases precision at the cost of recall.
Why each option works or fails
• ✗ A. Recall for spam: recall rewards catching as much spam as possible; a filter can reach high recall by flagging aggressively, which increases the false positives the business wants to avoid.
• ✓ B. Precision for the spam class (correct): it falls whenever legitimate emails are wrongly flagged, so it is the metric that directly tracks the requirement to rarely misclassify good mail.
• ✗ C. Mean absolute error: MAE measures the average absolute difference for continuous numeric predictions (regression), not correct versus incorrect class labels.
• ✗ D. R-squared: R-squared measures how much variance a regression model explains; it has no meaning for spam or not-spam decisions.
Question 5 of 5
A factory needs to know which of 12 defect types appear anywhere in an image (several can appear at once), without locating them. Which approach fits?
Show answer
✅ Answer: D
Difficulty: advanced · Topic: SageMaker Image Classification
Multi-class vs multi-label classification
• Multi-class classification picks exactly one class per image: the defect is either type 3 or type 7.
• Multi-label classification gives an independent yes/no for every class, so one image can be tagged with type 3 and type 7 at the same time. Technically the output layer uses a sigmoid per class instead of one softmax over all classes, and training uses one label vector per image, for example [0,0,1,0,0,0,1,0,0,0,0,0].
Choosing the task by the question asked:
• "Which defects are present?" is classification.
• "Where are they?" is object detection (boxes).
• "Which pixels?" is semantic segmentation.
The SageMaker built-in image classification algorithm supports a multi-label mode, and Rekognition Custom Labels also supports multi-label image classification.
Why each option works or fails
• ✗ A. Object detection: it requires bounding-box annotations for every defect and outputs locations, which the factory explicitly does not need; it costs more to label and train.
• ✗ B. Semantic segmentation: it needs pixel-level masks and predicts a class for every pixel, far more detail and labeling effort than a presence check needs.
• ✗ C. BlazingText on file names: file names say nothing about what is visible in the image, and BlazingText is a text algorithm that never sees pixels.
• ✓ D. Multi-label image classification (correct): it reports, for each of the 12 defect types, whether it appears anywhere in the image, allows several at once and needs no location labels.
🎯 Takeaways
• When full fine-tuning does not fit, train small adapters on a frozen (optionally quantized) base with LoRA or QLoRA.
• Good overall numbers can hide segment failures; evaluate per segment and fix the data or model, rather than deploying anyway or manipulating thresholds and test sets.
• When regulators require individual explanations, choose an interpretable model or pair a complex one with robust explanation tooling; accuracy never replaces compliance.
• When false positives are the costly mistake, optimize precision; when missed positives are costly, optimize recall.
• "Which types are present, possibly several, without location" is multi-label classification; boxes and masks answer "where", which costs more to label.
🔗 More from this exam
• All MLA-C02 practice sets
• Every exam on this blog
📚 Beyond this set
More MLA-C02 questions: https://hyunhp.github.io/cert-bank/#exam=MLA-C02
Study guide: https://hyunhp.github.io/cert-bank/study.html#exam=MLA-C02
These are original practice questions, not real exam questions, and cert-bank is not affiliated with any certification vendor.
Original practice questions written for this blog, not exam dumps.
Comments
Post a Comment