Amazon AWS Certified Machine Learning Engineer - Associate MLA-C01 Practice Test 112
Question 1110
During a controlled replay of the data-preparation workflow in a separate ML staging account, the following condition is observed: New event producers change a numerical source field to strings while legacy objects remain numeric, causing the shared Glue transformation to fail. The dataset contract must remain reproducible for a backfill. What should the ML engineer inspect FIRST to distinguish this failure from unrelated issues?
- Compare online Feature Store record values and offline event-time keyed feature history for the same entity and model snapshot.
- Inspect group sample sizes, label-positive fractions, and the appropriate SageMaker Clarify pretraining bias report.
- Compare source schemas by partition, Glue Data Catalog types, and failed cast rows for the evolving field.
- Check job names, bookmark settings, run history, source object timestamps, and the previously committed checkpoint.
Explanation
Correct Answer
C:Compare source schemas by partition, Glue Data Catalog types, and failed cast rows for the evolving field. This directly tests glue schema evolution with mixed input types in the reported AWS ML workflow. A catalog entry does not automatically reconcile incompatible field types across all source partitions. It distinguishes the problem before choosing a change. The related targeted response is: Define explicit compatible conversions and schema contracts, isolate incompatible versions, and validate before publishing features. Reported signal: New event producers change a numerical source field to strings while legacy objects remain numeric, causing the shared Glue transformation to fail. Operational setting: During a controlled replay of the data-preparation workflow in a separate ML staging account.
Incorrect Answers
A:Compare online Feature Store record values and offline event-time keyed feature history for the same entity and model snapshot. This is a possible diagnostic for sagemaker feature store online offline consistency, but it does not directly test the glue schema evolution with mixed input types failure reported here. Prioritize: Compare source schemas by partition, Glue Data Catalog types, and failed cast rows for the evolving field. The relevant case is: New event producers change a numerical source field to strings while legacy objects remain numeric, causing the shared Glue transformation to fail. Operational setting: During a controlled replay of the data-preparation workflow in a separate ML staging account.
B:Inspect group sample sizes, label-positive fractions, and the appropriate SageMaker Clarify pretraining bias report. This is a possible diagnostic for clarify label imbalance measurement, but it does not directly test the glue schema evolution with mixed input types failure reported here. Prioritize: Compare source schemas by partition, Glue Data Catalog types, and failed cast rows for the evolving field. The relevant case is: New event producers change a numerical source field to strings while legacy objects remain numeric, causing the shared Glue transformation to fail. Operational setting: During a controlled replay of the data-preparation workflow in a separate ML staging account.
D:Check job names, bookmark settings, run history, source object timestamps, and the previously committed checkpoint. This is a possible diagnostic for glue job bookmark state error, but it does not directly test the glue schema evolution with mixed input types failure reported here. Prioritize: Compare source schemas by partition, Glue Data Catalog types, and failed cast rows for the evolving field. The relevant case is: New event producers change a numerical source field to strings while legacy objects remain numeric, causing the shared Glue transformation to fail. Operational setting: During a controlled replay of the data-preparation workflow in a separate ML staging account.
Question 1111
During a production MLOps reliability and security review, the following condition is observed: SageMaker Model Monitor model-quality jobs cannot match later labels to predictions because inference requests lack stable inference IDs. The owner needs an auditable resolution without broadening privileges. Which measured outcome would confirm that the correction worked as intended?
- Measured training cost drops while interrupted test jobs resume from checkpoints within the delivery window.
- Repeated requests for one test user reach the same assigned variant and measured outcomes remain comparable.
- Cost reports attribute the majority of in-scope ML spend to the correct team without altering access controls.
- The monitor joins each known labeled request to the corresponding earlier prediction and computes expected metrics.
Explanation
Correct Answer
D:The monitor joins each known labeled request to the corresponding earlier prediction and computes expected metrics. This is the acceptance evidence for the described model monitor ground truth association scenario. Model quality analysis requires reliable association between predictions and subsequent truth records. The corrective implementation it validates is: Capture stable prediction identifiers and ingest verified labels with matching keys for model-quality analysis. Reported signal: SageMaker Model Monitor model-quality jobs cannot match later labels to predictions because inference requests lack stable inference IDs. Operational setting: During a production MLOps reliability and security review.
Incorrect Answers
A:Measured training cost drops while interrupted test jobs resume from checkpoints within the delivery window. This could validate spot training cost versus interruption recovery, but it does not establish that model monitor ground truth association is resolved for this workload. Confirm instead: The monitor joins each known labeled request to the corresponding earlier prediction and computes expected metrics. Observed problem: SageMaker Model Monitor model-quality jobs cannot match later labels to predictions because inference requests lack stable inference IDs. Operational setting: During a production MLOps reliability and security review.
B:Repeated requests for one test user reach the same assigned variant and measured outcomes remain comparable. This could validate champion challenger a/b assignment consistency, but it does not establish that model monitor ground truth association is resolved for this workload. Confirm instead: The monitor joins each known labeled request to the corresponding earlier prediction and computes expected metrics. Observed problem: SageMaker Model Monitor model-quality jobs cannot match later labels to predictions because inference requests lack stable inference IDs. Operational setting: During a production MLOps reliability and security review.
C:Cost reports attribute the majority of in-scope ML spend to the correct team without altering access controls. This could validate tag-based ml training cost attribution, but it does not establish that model monitor ground truth association is resolved for this workload. Confirm instead: The monitor joins each known labeled request to the corresponding earlier prediction and computes expected metrics. Observed problem: SageMaker Model Monitor model-quality jobs cannot match later labels to predictions because inference requests lack stable inference IDs. Operational setting: During a production MLOps reliability and security review.
Question 1112
While reproducing a training incident in an independent SageMaker experiment, the following condition is observed: A tabular model shows one post-outcome administrative flag as its strongest feature in SageMaker Clarify despite implausibly high offline scores. The experiment must be compared fairly against its prior baseline. What should the ML engineer inspect FIRST to distinguish this failure from unrelated issues?
- Inspect Clarify attribution results, feature timestamp semantics, and whether the flag is created after the target event.
- Measure class prevalence, precision-recall curve, ROC curve, and alert volume at the intended operating threshold.
- Inspect target types, business loss function, sample labels, and whether the required output is continuous or categorical.
- Inspect interruption events, checkpoint directory mapping, restoration logic, and persistent S3 checkpoint artifacts.
Explanation
Correct Answer
A:Inspect Clarify attribution results, feature timestamp semantics, and whether the flag is created after the target event. This directly tests clarify feature attribution against target leakage in the reported AWS ML workflow. Attribution can expose suspicious shortcuts; a high importance value is not proof a feature is valid or causal. It distinguishes the problem before choosing a change. The related targeted response is: Remove label-derived or post-outcome features, rebuild a point-in-time split, and recompute explanations and holdout metrics. Reported signal: A tabular model shows one post-outcome administrative flag as its strongest feature in SageMaker Clarify despite implausibly high offline scores. Operational setting: While reproducing a training incident in an independent SageMaker experiment.
Incorrect Answers
B:Measure class prevalence, precision-recall curve, ROC curve, and alert volume at the intended operating threshold. This is a possible diagnostic for roc-auc versus pr-auc for rare events, but it does not directly test the clarify feature attribution against target leakage failure reported here. Prioritize: Inspect Clarify attribution results, feature timestamp semantics, and whether the flag is created after the target event. The relevant case is: A tabular model shows one post-outcome administrative flag as its strongest feature in SageMaker Clarify despite implausibly high offline scores. Operational setting: While reproducing a training incident in an independent SageMaker experiment.
C:Inspect target types, business loss function, sample labels, and whether the required output is continuous or categorical. This is a possible diagnostic for regression versus classification problem formulation, but it does not directly test the clarify feature attribution against target leakage failure reported here. Prioritize: Inspect Clarify attribution results, feature timestamp semantics, and whether the flag is created after the target event. The relevant case is: A tabular model shows one post-outcome administrative flag as its strongest feature in SageMaker Clarify despite implausibly high offline scores. Operational setting: While reproducing a training incident in an independent SageMaker experiment.
D:Inspect interruption events, checkpoint directory mapping, restoration logic, and persistent S3 checkpoint artifacts. This is a possible diagnostic for managed spot training checkpoint recovery, but it does not directly test the clarify feature attribution against target leakage failure reported here. Prioritize: Inspect Clarify attribution results, feature timestamp semantics, and whether the flag is created after the target event. The relevant case is: A tabular model shows one post-outcome administrative flag as its strongest feature in SageMaker Clarify despite implausibly high offline scores. Operational setting: While reproducing a training incident in an independent SageMaker experiment.
Question 1113
During an approved rollout of an AWS ML inference workload, the following condition is observed: A custom image works when run locally but SageMaker cannot invoke it because the container lacks an expected health route and invocation handler. The team must meet its documented availability and rollback constraints. Which measured outcome would confirm that the correction worked as intended?
- Cold-start p95 improves during planned demand windows while capacity cost remains measured.
- The production endpoint matches the reviewed model package and auditor can trace its evaluation and approver.
- Endpoint health becomes Ready and inference returns the specified payload type for a known test request.
- Distinct requested model IDs return correct predictions with acceptable cold load latency and lower aggregate idle cost.
Explanation
Correct Answer
C:Endpoint health becomes Ready and inference returns the specified payload type for a known test request. This is the acceptance evidence for the described bring-your-own-inference-container contract scenario. A Docker image is not automatically a SageMaker-compatible inference server without the runtime HTTP contract. The corrective implementation it validates is: Implement the SageMaker inference container /ping and /invocations expectations or use a supported serving toolkit. Reported signal: A custom image works when run locally but SageMaker cannot invoke it because the container lacks an expected health route and invocation handler. Operational setting: During an approved rollout of an AWS ML inference workload.
Incorrect Answers
A:Cold-start p95 improves during planned demand windows while capacity cost remains measured. This could validate sagemaker endpoint provisioned concurrency tradeoff, but it does not establish that bring-your-own-inference-container contract is resolved for this workload. Confirm instead: Endpoint health becomes Ready and inference returns the specified payload type for a known test request. Observed problem: A custom image works when run locally but SageMaker cannot invoke it because the container lacks an expected health route and invocation handler. Operational setting: During an approved rollout of an AWS ML inference workload.
B:The production endpoint matches the reviewed model package and auditor can trace its evaluation and approver. This could validate model registry version promotion provenance, but it does not establish that bring-your-own-inference-container contract is resolved for this workload. Confirm instead: Endpoint health becomes Ready and inference returns the specified payload type for a known test request. Observed problem: A custom image works when run locally but SageMaker cannot invoke it because the container lacks an expected health route and invocation handler. Operational setting: During an approved rollout of an AWS ML inference workload.
D:Distinct requested model IDs return correct predictions with acceptable cold load latency and lower aggregate idle cost. This could validate multi-model endpoint hosting tradeoff, but it does not establish that bring-your-own-inference-container contract is resolved for this workload. Confirm instead: Endpoint health becomes Ready and inference returns the specified payload type for a known test request. Observed problem: A custom image works when run locally but SageMaker cannot invoke it because the container lacks an expected health route and invocation handler. Operational setting: During an approved rollout of an AWS ML inference workload.
Question 1114
At a production AWS ML data lake supporting customer-facing predictions, the following condition is observed: A missing-value imputer calculates its median on the complete labeled dataset before train/test splitting, inflating apparent validation performance. Engineers must avoid changing unrelated source partitions. Which observation most directly identifies the root cause of the reported behavior?
- Inspect preprocessing fit order, imputation statistics, dataset split boundaries, and pipeline serialization.
- Inspect sampler placement, duplicate neighbor IDs, fold membership, and class-conditioned metric changes.
- Compare vocabulary hashes, normalization steps, token IDs for test strings, and serialized preprocessing versions.
- Inspect feature availability timestamps and whether each label time follows the last feature observation.
Explanation
Correct Answer
A:Inspect preprocessing fit order, imputation statistics, dataset split boundaries, and pipeline serialization. This directly tests median imputation fitted on training split in the reported AWS ML workflow. Preprocessing parameters learned from held-out records leak information into model selection. It distinguishes the problem before choosing a change. The related targeted response is: Split first; fit the imputer only on the training partition and apply those learned statistics to validation and test data. Reported signal: A missing-value imputer calculates its median on the complete labeled dataset before train/test splitting, inflating apparent validation performance. Operational setting: At a production AWS ML data lake supporting customer-facing predictions.
Incorrect Answers
B:Inspect sampler placement, duplicate neighbor IDs, fold membership, and class-conditioned metric changes. This is a possible diagnostic for class imbalance handled inside cross-validation, but it does not directly test the median imputation fitted on training split failure reported here. Prioritize: Inspect preprocessing fit order, imputation statistics, dataset split boundaries, and pipeline serialization. The relevant case is: A missing-value imputer calculates its median on the complete labeled dataset before train/test splitting, inflating apparent validation performance. Operational setting: At a production AWS ML data lake supporting customer-facing predictions.
C:Compare vocabulary hashes, normalization steps, token IDs for test strings, and serialized preprocessing versions. This is a possible diagnostic for text tokenization mismatch in serving, but it does not directly test the median imputation fitted on training split failure reported here. Prioritize: Inspect preprocessing fit order, imputation statistics, dataset split boundaries, and pipeline serialization. The relevant case is: A missing-value imputer calculates its median on the complete labeled dataset before train/test splitting, inflating apparent validation performance. Operational setting: At a production AWS ML data lake supporting customer-facing predictions.
D:Inspect feature availability timestamps and whether each label time follows the last feature observation. This is a possible diagnostic for chronological split for forecast evaluation, but it does not directly test the median imputation fitted on training split failure reported here. Prioritize: Inspect preprocessing fit order, imputation statistics, dataset split boundaries, and pipeline serialization. The relevant case is: A missing-value imputer calculates its median on the complete labeled dataset before train/test splitting, inflating apparent validation performance. Operational setting: At a production AWS ML data lake supporting customer-facing predictions.
Question 1115
At a production AWS ML data lake supporting customer-facing predictions, the following condition is observed: An offline churn dataset uses the latest cancellation-status feature for all historical rows, exposing outcomes not known when earlier predictions would run. Engineers must avoid changing unrelated source partitions. Select TWO: the appropriate diagnostic check and the targeted implementation change.
- Perform point-in-time correct joins using the most recent eligible pre-prediction feature record, not the globally latest value.
- Apply resampling only inside each training fold and evaluate untouched validation folds using precision and recall metrics.
- Inspect feature event timestamps, target prediction cutoffs, and join predicates in the historical training query.
- Define point-in-time offline joins, freshness SLAs, and aligned feature version logic for online retrieval.
- Check job names, bookmark settings, run history, source object timestamps, and the previously committed checkpoint.
Explanation
Correct Answers
A:Perform point-in-time correct joins using the most recent eligible pre-prediction feature record, not the globally latest value. This is the appropriate controlled remedy for point-in-time feature data leakage. A feature created after the prediction event cannot legitimately inform that earlier prediction. Verify the cause with: Inspect feature event timestamps, target prediction cutoffs, and join predicates in the historical training query. Then confirm: For every training row the joined feature timestamp is at or before its prediction timestamp and no label-derived value leaks. This is one part of a select-two decision; the matching evidence and corrective action are both required. Reported signal: An offline churn dataset uses the latest cancellation-status feature for all historical rows, exposing outcomes not known when earlier predictions would run. Operational setting: At a production AWS ML data lake supporting customer-facing predictions.
C:Inspect feature event timestamps, target prediction cutoffs, and join predicates in the historical training query. This directly tests point-in-time feature data leakage in the reported AWS ML workflow. A feature created after the prediction event cannot legitimately inform that earlier prediction. It distinguishes the problem before choosing a change. The related targeted response is: Perform point-in-time correct joins using the most recent eligible pre-prediction feature record, not the globally latest value. This is one part of a select-two decision; the matching evidence and corrective action are both required. Reported signal: An offline churn dataset uses the latest cancellation-status feature for all historical rows, exposing outcomes not known when earlier predictions would run. Operational setting: At a production AWS ML data lake supporting customer-facing predictions.
Incorrect Answers
B:Apply resampling only inside each training fold and evaluate untouched validation folds using precision and recall metrics. This selection concerns class imbalance handled inside cross-validation rather than the required diagnostic-and-remedy pair for point-in-time feature data leakage. Use evidence: Inspect feature event timestamps, target prediction cutoffs, and join predicates in the historical training query. and the matching correction: Perform point-in-time correct joins using the most recent eligible pre-prediction feature record, not the globally latest value. Signal: An offline churn dataset uses the latest cancellation-status feature for all historical rows, exposing outcomes not known when earlier predictions would run. Operational setting: At a production AWS ML data lake supporting customer-facing predictions.
D:Define point-in-time offline joins, freshness SLAs, and aligned feature version logic for online retrieval. This selection concerns sagemaker feature store online offline consistency rather than the required diagnostic-and-remedy pair for point-in-time feature data leakage. Use evidence: Inspect feature event timestamps, target prediction cutoffs, and join predicates in the historical training query. and the matching correction: Perform point-in-time correct joins using the most recent eligible pre-prediction feature record, not the globally latest value. Signal: An offline churn dataset uses the latest cancellation-status feature for all historical rows, exposing outcomes not known when earlier predictions would run. Operational setting: At a production AWS ML data lake supporting customer-facing predictions.
E:Check job names, bookmark settings, run history, source object timestamps, and the previously committed checkpoint. This selection concerns glue job bookmark state error rather than the required diagnostic-and-remedy pair for point-in-time feature data leakage. Use evidence: Inspect feature event timestamps, target prediction cutoffs, and join predicates in the historical training query. and the matching correction: Perform point-in-time correct joins using the most recent eligible pre-prediction feature record, not the globally latest value. Signal: An offline churn dataset uses the latest cancellation-status feature for all historical rows, exposing outcomes not known when earlier predictions would run. Operational setting: At a production AWS ML data lake supporting customer-facing predictions.
Question 1116
As a regional data-pipeline rollout is validated against approved AWS storage and streams, the following condition is observed: An overnight ML export runs expensive full-table queries on a production RDS database and causes application transaction latency spikes. The data team must preserve event-time meaning and lineage. What postchange observation provides the strongest acceptance evidence?
- Identical inputs create identical token IDs in train and serving paths, including punctuation and unknown words.
- Application p95 latency returns to target while exported records meet a measured completeness and staleness objective.
- A held-out split uses the training-fitted median and independent evaluation no longer incorporates held-out distribution values.
- Training read time and object operations decline while the resulting row count and checksums match.
Explanation
Correct Answer
B:Application p95 latency returns to target while exported records meet a measured completeness and staleness objective. This is the acceptance evidence for the described rds extraction overload into ml scenario. Extract jobs contend with transactional queries; moving or bounding reads can isolate production workloads. The corrective implementation it validates is: Use controlled incremental extraction or a suitable read replica and throttle export jobs within the approved data freshness target. Reported signal: An overnight ML export runs expensive full-table queries on a production RDS database and causes application transaction latency spikes. Operational setting: As a regional data-pipeline rollout is validated against approved AWS storage and streams.
Incorrect Answers
A:Identical inputs create identical token IDs in train and serving paths, including punctuation and unknown words. This could validate text tokenization mismatch in serving, but it does not establish that rds extraction overload into ml is resolved for this workload. Confirm instead: Application p95 latency returns to target while exported records meet a measured completeness and staleness objective. Observed problem: An overnight ML export runs expensive full-table queries on a production RDS database and causes application transaction latency spikes. Operational setting: As a regional data-pipeline rollout is validated against approved AWS storage and streams.
C:A held-out split uses the training-fitted median and independent evaluation no longer incorporates held-out distribution values. This could validate median imputation fitted on training split, but it does not establish that rds extraction overload into ml is resolved for this workload. Confirm instead: Application p95 latency returns to target while exported records meet a measured completeness and staleness objective. Observed problem: An overnight ML export runs expensive full-table queries on a production RDS database and causes application transaction latency spikes. Operational setting: As a regional data-pipeline rollout is validated against approved AWS storage and streams.
D:Training read time and object operations decline while the resulting row count and checksums match. This could validate columnar file compaction for small objects, but it does not establish that rds extraction overload into ml is resolved for this workload. Confirm instead: Application p95 latency returns to target while exported records meet a measured completeness and staleness objective. Observed problem: An overnight ML export runs expensive full-table queries on a production RDS database and causes application transaction latency spikes. Operational setting: As a regional data-pipeline rollout is validated against approved AWS storage and streams.
Question 1117
As a regional data-pipeline rollout is validated against approved AWS storage and streams, the following condition is observed: A training preparation scan reads every month of event history even though the feature job requests one recent date, causing high bytes scanned and slow ingestion. The data team must preserve event-time meaning and lineage. Which TWO complementary choices provide the decisive diagnosis AND corrective response? Select TWO.
- Measure label distribution differences with an appropriate pretraining bias metric and improve data representation before training.
- Write partitioned columnar Parquet data by event date, register accurate partitions in the catalog, and filter on partition keys.
- Compare the query predicates with S3 object prefixes, partition metadata, and query plans to confirm partitions are not being pruned.
- Inspect S3 replication rules, region-specific bucket policies, Glue job destinations, KMS keys, and applicable residency controls.
- Make the ingestion handler idempotent using a stable object/version key and conditional ledger writes.
Explanation
Correct Answers
B:Write partitioned columnar Parquet data by event date, register accurate partitions in the catalog, and filter on partition keys. This is the appropriate controlled remedy for partition pruning in s3 parquet training input. Partition-aware columnar scans reduce irrelevant I/O; naming a date column without partition metadata does not create pruning. Verify the cause with: Compare the query predicates with S3 object prefixes, partition metadata, and query plans to confirm partitions are not being pruned. Then confirm: An equivalent recent-date query reads only its selected partitions while returning the same labeled rows. This is one part of a select-two decision; the matching evidence and corrective action are both required. Reported signal: A training preparation scan reads every month of event history even though the feature job requests one recent date, causing high bytes scanned and slow ingestion. Operational setting: As a regional data-pipeline rollout is validated against approved AWS storage and streams.
C:Compare the query predicates with S3 object prefixes, partition metadata, and query plans to confirm partitions are not being pruned. This directly tests partition pruning in s3 parquet training input in the reported AWS ML workflow. Partition-aware columnar scans reduce irrelevant I/O; naming a date column without partition metadata does not create pruning. It distinguishes the problem before choosing a change. The related targeted response is: Write partitioned columnar Parquet data by event date, register accurate partitions in the catalog, and filter on partition keys. This is one part of a select-two decision; the matching evidence and corrective action are both required. Reported signal: A training preparation scan reads every month of event history even though the feature job requests one recent date, causing high bytes scanned and slow ingestion. Operational setting: As a regional data-pipeline rollout is validated against approved AWS storage and streams.
Incorrect Answers
A:Measure label distribution differences with an appropriate pretraining bias metric and improve data representation before training. This selection concerns clarify label imbalance measurement rather than the required diagnostic-and-remedy pair for partition pruning in s3 parquet training input. Use evidence: Compare the query predicates with S3 object prefixes, partition metadata, and query plans to confirm partitions are not being pruned. and the matching correction: Write partitioned columnar Parquet data by event date, register accurate partitions in the catalog, and filter on partition keys. Signal: A training preparation scan reads every month of event history even though the feature job requests one recent date, causing high bytes scanned and slow ingestion. Operational setting: As a regional data-pipeline rollout is validated against approved AWS storage and streams.
D:Inspect S3 replication rules, region-specific bucket policies, Glue job destinations, KMS keys, and applicable residency controls. This selection concerns model training data residency boundary rather than the required diagnostic-and-remedy pair for partition pruning in s3 parquet training input. Use evidence: Compare the query predicates with S3 object prefixes, partition metadata, and query plans to confirm partitions are not being pruned. and the matching correction: Write partitioned columnar Parquet data by event date, register accurate partitions in the catalog, and filter on partition keys. Signal: A training preparation scan reads every month of event history even though the feature job requests one recent date, causing high bytes scanned and slow ingestion. Operational setting: As a regional data-pipeline rollout is validated against approved AWS storage and streams.
E:Make the ingestion handler idempotent using a stable object/version key and conditional ledger writes. This selection concerns s3 object event at-least-once processing rather than the required diagnostic-and-remedy pair for partition pruning in s3 parquet training input. Use evidence: Compare the query predicates with S3 object prefixes, partition metadata, and query plans to confirm partitions are not being pruned. and the matching correction: Write partitioned columnar Parquet data by event date, register accurate partitions in the catalog, and filter on partition keys. Signal: A training preparation scan reads every month of event history even though the feature job requests one recent date, causing high bytes scanned and slow ingestion. Operational setting: As a regional data-pipeline rollout is validated against approved AWS storage and streams.
Question 1118
While reproducing a training incident in an independent SageMaker experiment, the following condition is observed: A retail analytics group needs to discover coherent groups of customers but has no verified segment labels for supervised classification. The experiment must be compared fairly against its prior baseline. Which observation most directly identifies the root cause of the reported behavior?
- Inspect Clarify attribution results, feature timestamp semantics, and whether the flag is created after the target event.
- Check whether trustworthy target labels exist and compare feature distance scaling and cluster stability on a sample.
- Inspect target types, business loss function, sample labels, and whether the required output is continuous or categorical.
- Inspect Docker entrypoint, training logs, image architecture, input mount points, and required container contract.
Explanation
Correct Answer
B:Check whether trustworthy target labels exist and compare feature distance scaling and cluster stability on a sample. This directly tests clustering for unlabeled customer segments in the reported AWS ML workflow. Unsupervised grouping does not require labeled segment classes and should not be judged by an unavailable label accuracy metric. It distinguishes the problem before choosing a change. The related targeted response is: Use an unsupervised clustering approach such as k-means after appropriate feature preparation and assess whether groups are meaningful. Reported signal: A retail analytics group needs to discover coherent groups of customers but has no verified segment labels for supervised classification. Operational setting: While reproducing a training incident in an independent SageMaker experiment.
Incorrect Answers
A:Inspect Clarify attribution results, feature timestamp semantics, and whether the flag is created after the target event. This is a possible diagnostic for clarify feature attribution against target leakage, but it does not directly test the clustering for unlabeled customer segments failure reported here. Prioritize: Check whether trustworthy target labels exist and compare feature distance scaling and cluster stability on a sample. The relevant case is: A retail analytics group needs to discover coherent groups of customers but has no verified segment labels for supervised classification. Operational setting: While reproducing a training incident in an independent SageMaker experiment.
C:Inspect target types, business loss function, sample labels, and whether the required output is continuous or categorical. This is a possible diagnostic for regression versus classification problem formulation, but it does not directly test the clustering for unlabeled customer segments failure reported here. Prioritize: Check whether trustworthy target labels exist and compare feature distance scaling and cluster stability on a sample. The relevant case is: A retail analytics group needs to discover coherent groups of customers but has no verified segment labels for supervised classification. Operational setting: While reproducing a training incident in an independent SageMaker experiment.
D:Inspect Docker entrypoint, training logs, image architecture, input mount points, and required container contract. This is a possible diagnostic for training image compatibility for byoc, but it does not directly test the clustering for unlabeled customer segments failure reported here. Prioritize: Check whether trustworthy target labels exist and compare feature distance scaling and cluster stability on a sample. The relevant case is: A retail analytics group needs to discover coherent groups of customers but has no verified segment labels for supervised classification. Operational setting: While reproducing a training incident in an independent SageMaker experiment.
Question 1119
At the model-governance gate before promotion of a new ML artifact, the following condition is observed: A fraud model has structured transaction attributes and labeled outcomes, but engineers start with a large image neural network despite limited compute and few features. The team needs a defensible technical decision rather than an untracked workaround. Which measured outcome would confirm that the correction worked as intended?
- The tabular baseline meets recall, cost, and latency constraints on a representative held-out dataset.
- The chosen version meets task metrics and license conditions and its inference contract is reproducible.
- Fraud recall meets the agreed target with documented review-volume and false-positive tradeoffs.
- Training completes within GPU memory limits and reaches comparable optimizer behavior and accuracy.
Explanation
Correct Answer
A:The tabular baseline meets recall, cost, and latency constraints on a representative held-out dataset. This is the acceptance evidence for the described xgboost tabular classification baseline scenario. Gradient-boosted trees are a practical strong baseline for structured tabular features without needing an image model architecture. The corrective implementation it validates is: Train a gradient-boosted tree baseline such as SageMaker XGBoost and compare it against more complex candidates. Reported signal: A fraud model has structured transaction attributes and labeled outcomes, but engineers start with a large image neural network despite limited compute and few features. Operational setting: At the model-governance gate before promotion of a new ML artifact.
Incorrect Answers
B:The chosen version meets task metrics and license conditions and its inference contract is reproducible. This could validate sagemaker jumpstart template provenance, but it does not establish that xgboost tabular classification baseline is resolved for this workload. Confirm instead: The tabular baseline meets recall, cost, and latency constraints on a representative held-out dataset. Observed problem: A fraud model has structured transaction attributes and labeled outcomes, but engineers start with a large image neural network despite limited compute and few features. Operational setting: At the model-governance gate before promotion of a new ML artifact.
C:Fraud recall meets the agreed target with documented review-volume and false-positive tradeoffs. This could validate precision versus recall fraud threshold, but it does not establish that xgboost tabular classification baseline is resolved for this workload. Confirm instead: The tabular baseline meets recall, cost, and latency constraints on a representative held-out dataset. Observed problem: A fraud model has structured transaction attributes and labeled outcomes, but engineers start with a large image neural network despite limited compute and few features. Operational setting: At the model-governance gate before promotion of a new ML artifact.
D:Training completes within GPU memory limits and reaches comparable optimizer behavior and accuracy. This could validate training batch size causes gpu memory exhaustion, but it does not establish that xgboost tabular classification baseline is resolved for this workload. Confirm instead: The tabular baseline meets recall, cost, and latency constraints on a representative held-out dataset. Observed problem: A fraud model has structured transaction attributes and labeled outcomes, but engineers start with a large image neural network despite limited compute and few features. Operational setting: At the model-governance gate before promotion of a new ML artifact.
Question 1120
At the model-governance gate before promotion of a new ML artifact, the following condition is observed: A pretrained image encoder loses useful general features after fine-tuning on a small labeled dataset using an excessively high learning rate. The team needs a defensible technical decision rather than an untracked workaround. What should the ML engineer inspect FIRST to distinguish this failure from unrelated issues?
- Inspect training job InputDataConfig, environment variables such as SM_CHANNEL_TRAINING, and container paths.
- Inspect target types, business loss function, sample labels, and whether the required output is continuous or categorical.
- Compare pretrained validation metrics, fine-tuning learning curves, and layer-wise parameter updates.
- Check whether trustworthy target labels exist and compare feature distance scaling and cluster stability on a sample.
Explanation
Correct Answer
C:Compare pretrained validation metrics, fine-tuning learning curves, and layer-wise parameter updates. This directly tests fine-tuning pretrained model learning-rate stability in the reported AWS ML workflow. Large updates on small data can overwrite pretrained features; careful transfer learning limits destructive forgetting. It distinguishes the problem before choosing a change. The related targeted response is: Use a smaller learning rate, optionally freeze early layers, and assess staged unfreezing against a held-out set. Reported signal: A pretrained image encoder loses useful general features after fine-tuning on a small labeled dataset using an excessively high learning rate. Operational setting: At the model-governance gate before promotion of a new ML artifact.
Incorrect Answers
A:Inspect training job InputDataConfig, environment variables such as SM_CHANNEL_TRAINING, and container paths. This is a possible diagnostic for sagemaker script-mode data channel mismatch, but it does not directly test the fine-tuning pretrained model learning-rate stability failure reported here. Prioritize: Compare pretrained validation metrics, fine-tuning learning curves, and layer-wise parameter updates. The relevant case is: A pretrained image encoder loses useful general features after fine-tuning on a small labeled dataset using an excessively high learning rate. Operational setting: At the model-governance gate before promotion of a new ML artifact.
B:Inspect target types, business loss function, sample labels, and whether the required output is continuous or categorical. This is a possible diagnostic for regression versus classification problem formulation, but it does not directly test the fine-tuning pretrained model learning-rate stability failure reported here. Prioritize: Compare pretrained validation metrics, fine-tuning learning curves, and layer-wise parameter updates. The relevant case is: A pretrained image encoder loses useful general features after fine-tuning on a small labeled dataset using an excessively high learning rate. Operational setting: At the model-governance gate before promotion of a new ML artifact.
D:Check whether trustworthy target labels exist and compare feature distance scaling and cluster stability on a sample. This is a possible diagnostic for clustering for unlabeled customer segments, but it does not directly test the fine-tuning pretrained model learning-rate stability failure reported here. Prioritize: Compare pretrained validation metrics, fine-tuning learning curves, and layer-wise parameter updates. The relevant case is: A pretrained image encoder loses useful general features after fine-tuning on a small labeled dataset using an excessively high learning rate. Operational setting: At the model-governance gate before promotion of a new ML artifact.
