模擬試験 1

Batch learning only says data is processed in batches. It does not specifically mean continuing an FM from a checkpoint to absorb new corpora while preserving prior knowledge. Trap: Daily refreshes may sound like batch jobs, but the exam is asking for the FM lifecycle strategy.

Static training leaves model weights fixed after a training run, so the model would not reliably reflect breaking news within 24 hours. Trap: A static model can be stable, but stability conflicts with the requirement for daily knowledge refresh.

Latent training is not the standard term for maintaining a foundation model with new data and checkpoints. Trap: The word latent appears in ML, but latent spaces are not a training strategy for this scenario.

Amazon EBS is block storage attached to compute resources such as EC2 instances, not an object store that Bedrock can directly reference for evaluation data. Trap: It stores data, but it is the wrong storage access pattern for a managed Bedrock evaluation dataset.

Amazon EFS is a managed network file system for workloads that mount file shares; it is not the object storage source expected by Bedrock evaluation. Trap: A JSONL file could sit on a file system, but the question specifically asks for object storage Bedrock can read directly.

AWS Snowcone is an edge data transfer and compute device, not a regional online object store for Bedrock evaluation jobs. Trap: Snowcone can move data into AWS, but Bedrock would still need the dataset in a service such as Amazon S3.

Temperature affects how random or deterministic the generated text is, but Bedrock is not billed by the chosen temperature value. Trap: It is an inference setting, so it sounds cost-related, but any cost impact would only be indirect if it changed output length.

The model is not being custom trained or fine-tuned, so the amount of training data is not part of this daily inference charge. Trap: Training data matters for model customization, but this is a managed inference workload.

There is no customer training job in the scenario, so training time cannot drive the charges. Trap: Training duration is a real ML cost dimension in training services, but not for this Bedrock inference pilot.

Amazon Q Developer is an AWS coding and development assistant. It is not a general service for building custom chatbots and generators with multiple foundation models. Trap: It uses generative AI, but its scope is developer productivity rather than managed multi-provider FM access.

Amazon Kendra is an enterprise search service. It can help retrieve content for a RAG architecture, but it does not provide managed generative model inference. Trap: The RAG clue makes search tempting, but the primary requirement is multi-provider foundation model access.

Amazon Comprehend performs natural language processing tasks such as entity detection, sentiment analysis, and PII detection. It is not a generative AI platform for foundation model inference. Trap: The chatbot is language-based, but Comprehend analyzes text instead of generating new responses.

SageMaker Data Wrangler helps prepare and transform data, but it is not the production feature repository that serves features online. Trap: It sounds relevant because features must be engineered, but the question asks how to version, share, and retrieve them across training and inference.

SageMaker Clarify is used for bias detection and explainability, not centralized feature storage or millisecond feature lookup. Trap: Feature attribution can sound like feature management, but Clarify explains model behavior rather than serving feature values.

SageMaker Model Cards document model details for governance, but they do not store engineered variables or support online inference retrieval. Trap: Versioned documentation is useful for governance, but it does not meet the offline and online feature-serving requirements.

Measurement bias would mean the features or labels were systematically recorded incorrectly. The scenario instead describes who was included in the training sample and in what proportions. Trap: The output is measured by zone, but the root issue is representativeness, not a faulty measurement process.

Observer bias comes from human observers or labelers influencing data through subjective judgment. The question does not describe biased reviewers; it describes an imbalanced dataset sample. Trap: A human process may have created the sample, but the named bias affecting the model output is the sampling imbalance.

Confirmation bias is a human tendency to favor evidence that supports an existing belief. The model problem shown here is a skewed training distribution for a subgroup. Trap: Someone might have oversampled because of a prior belief, but the exam asks for the bias type visible in the model data and output.

Crowd-sourced evaluation requires task design, privacy review, rater onboarding, calibration, and quality checks. That is not the least operational overhead for 12,000 healthcare summaries. Trap: Crowds can provide human judgment, but the question emphasizes consistency, speed, and low operations.

Human workers can be valuable for nuanced safety judgments, but they add instructions, training, inter-rater checks, and scheduling. That conflicts with the least-overhead requirement. Trap: Healthcare content feels like it deserves expert review, but this question asks for a scalable first comparison using consistent criteria.

RLHF is a model-improvement technique that uses human preference feedback for training or alignment. It is not a lightweight evaluation method for comparing toxicity rates across four existing models. Trap: RLHF is associated with safer LLM behavior, but it greatly increases workflow complexity when the task is measurement.

Principal components transform the original variables into latent components, so displayed weights would not map cleanly to prior fines, occupancy, or inspection age. Deep learning also reduces direct editability. Trap: PCA and deep learning can be useful for prediction, but they conflict with original-feature transparency.

k-NN predicts from distances to nearby examples and does not learn one global coefficient per feature. Changing feature scaling would not provide the requested simple weight dashboard. Trap: k-NN can rank records, but it is not an editable coefficient-based model.

A neural network has many interacting parameters rather than one stable, human-editable weight per input variable. Feature attribution is not the same as editable model coefficients. Trap: Neural networks can model complex patterns, but the requirement is interpretability and manual weight control.

Increasing generation steps mainly gives the model more refinement passes and usually increases latency; it is not the primary control for prompt adherence. Trap: More steps can improve image quality, so it sounds useful, but the question asks for stricter adherence without a large latency increase.

MASK_IMAGE_BLACK is a masking option for image editing workflows such as inpainting, not a way to make text-to-image generation follow a prompt more closely. Trap: It sounds like it controls image content, but there is no source image or mask in the scenario.

Prompt strength is typically used for image-to-image or inpainting control over how much to change an input image, not as the main text-to-image adherence parameter. Trap: The phrase sounds like stronger prompting, but CFG scale is the Stable Diffusion text-guidance control named in the scenario.