ML Model Development — AWS Certified Machine Learning Engineer MLA-C01 practice questions

To reduce overfitting in gradient boosting, tune which trio first?

1. Depth, eta, and subsample
2. Larger learning rate always
3. More rounds and depth only
4. Only higher tree depth values

Answer: A — Depth, eta, and subsample

Max_depth, eta, and subsample are the primary regularization knobs; more rounds and depth alone overfit further.

For a long-running one-off scoring of many gigabytes, MOST fit is which?

1. SageMaker Batch Transform
2. Serverless inference endpoint
3. Multi-container endpoint
4. Real-time endpoint always

Answer: A — SageMaker Batch Transform

Batch Transform is purpose-built for large offline scoring; the others carry payload limits or idle cost.

A tokenizer in NLP MOST directly does which action?

1. Trains model weights on data
2. Renders visualization plots
3. Splits text into token IDs
4. Encrypts model payload blobs

Answer: C — Splits text into token IDs

Tokenizers produce subword or word IDs; training and encryption are separate concerns.

A multi-model endpoint (MME) is PREFERRED under which condition?

1. Many similar small shared models
2. Each model has its own endpoint
3. Only one active model exists
4. Only offline batch scoring runs

Answer: A — Many similar small shared models

MME hosts many similar-framework models behind one container; the others do not benefit from MME.

High train accuracy with low validation accuracy MOST OFTEN implies which cause?

1. Overfitting to training set
2. Model is too simple overall
3. Learning rate is too low
4. Regularization is too high

Answer: A — Overfitting to training set

This divergence is classic overfit; the others describe underfit or convergence issues.

Fine-tuning large LLMs cheaply is BEST done with which method?

1. Full-weight retrain from scratch
2. Random init and full retrain
3. LoRA or adapter fine-tuning
4. Only prompt-tuning short strings

Answer: C — LoRA or adapter fine-tuning

LoRA and adapters train few extra parameters; full retrain is expensive; prompt-tuning alone is limited.

The bias-variance tradeoff MOST DIRECTLY describes which relationship?

1. Compute versus memory cost
2. Batch versus online serving
3. Underfit versus overfit balance
4. GPU versus CPU cost gap

Answer: C — Underfit versus overfit balance

Bias-variance is the underfit-versus-overfit balance; the others are infra concepts.

For tabular regression MOST fit SageMaker built-in options include which?

1. Only BlazingText only supervised
2. Only Object2Vec pair encoder
3. XGBoost or Linear Learner
4. Only K-Means clustering variant

Answer: C — XGBoost or Linear Learner

XGBoost and Linear Learner are standard tabular built-ins; the others fit text, pairs, or clustering.

For a large deep-learning training job, MOST fit instance family is which?

1. p4 or p5 GPU instances
2. m5 general-purpose CPU
3. r5 memory-optimized CPU
4. t3 burstable CPU family

Answer: A — p4 or p5 GPU instances

P-family GPUs are built for deep learning; CPU-only families are too slow.

Encrypt SageMaker model artifacts at rest using which mechanism?

1. Local instance disk only
2. Plaintext S3 object uploads
3. KMS with customer-managed key
4. Only IAM role attachment

Answer: C — KMS with customer-managed key

SSE-KMS with a CMK encrypts artifacts; IAM controls access but not encryption; plaintext is unencrypted.

RNN / LSTM best for?

Sequential data (text, time series). LSTM handles long-term dependencies better than vanilla RNN.

Amazon Transcribe?

Speech-to-text. Supports streaming + batch, custom vocabulary, speaker diarization.

SageMaker Hyperparameter Tuning?

Automated search (Bayesian, random, grid, Hyperband) over hyperparameter ranges.

SageMaker for image classification?

Built-in Image Classification (ResNet) or Object Detection. Or BYO with PyTorch/TF.