ML Model Development
Free AWS Certified Machine Learning Engineer - Associate practice — 6 questions on ML Model Development, with explanations. No sign-up. Full 12-question mixed test →
A team already ran a SageMaker hyperparameter tuning job for an XGBoost model using last year's customer churn dataset. They now have this year's updated churn dataset (different distribution and volume) and want to tune a new XGBoost model, using the completed tuning job's knowledge to converge faster. Which warm start configuration should they use?
TRANSFER_LEARNING warm start is designed for cases where the input dataset, hyperparameter ranges, or objective differs from the parent job, letting SageMaker AMT reuse prior evaluations as a prior while adapting to the new data.
A data science team is training a 20-billion-parameter transformer model on SageMaker. The model itself does not fit into a single GPU's memory even with a batch size of 1. Which SageMaker distributed training approach should they use?
SMP partitions the model itself (layers/tensors) across multiple GPUs so a model too large for one device's memory can be trained, which is exactly the scenario described.
You want SageMaker Debugger to automatically stop a training job as soon as it detects the training loss continuing to decrease while validation loss stops improving (overfitting), with the least amount of custom code. Which configuration accomplishes this?
SageMaker Debugger ships a built-in 'overfit' rule that monitors the gap between training and validation loss; attaching it with a StopTraining action lets SageMaker automatically terminate the job with no custom log parsing.
Before training a credit-approval model, a team wants to measure whether the training dataset itself has a disproportionate representation of approved loans between two demographic groups, prior to any model training. Which type of SageMaker Clarify metric should they compute?
CI and DPL are pretraining bias metrics that examine the raw dataset's label and group distributions before any model exists, exactly matching the requirement to check the data itself.
Which statement about the SageMaker built-in DeepAR forecasting algorithm is correct?
DeepAR is specifically designed to train one global RNN-based model over a set of related time series, and it produces probabilistic forecasts, providing multiple quantile estimates (e.g., p10, p50, p90) rather than a single value.
A team is training a large model using SageMaker Managed Spot Training. The job has been interrupted twice by spot reclamation, and each time training restarted from epoch 0, wasting hours of GPU time. What should they configure so the job resumes from the last completed epoch instead of restarting?
Managed Spot Training resumes correctly only when the training script itself saves checkpoints to the configured checkpoint_local_path (synced to checkpoint_s3_uri) and reloads them on restart — this is the mechanism that actually preserves progress across interruptions.
Ready for the real thing?
The full course: two full-length practice tests, video lessons for every exam domain, hands-on labs and detailed explanations.
$109.99 $34.99 with code FREETEST33 — valid through September 7.
Get my $34.99 deal →