Applications of Foundation Models practice questions

Enable data encryption in transit and at rest.

Why it fails: Encryption protects data confidentiality but does not prevent the model provider from using the data for training.

Use a different foundation model from a different provider.

Why it fails: Simply switching models does not guarantee that data will not be used for training unless the provider offers similar controls.

Configure a VPC endpoint for Amazon Bedrock.

Why it fails: VPC endpoints keep traffic within the AWS network but do not affect data usage for training.

The AWS CLI is missing the --endpoint-url parameter.

Why it fails: The command succeeded; endpoint is not the issue.

The max_tokens_to_sample is too low.

Why it fails: The stop_reason is stop_sequence, not max_tokens.

The model does not support this use case.

Why it fails: Claude v2 supports explanation tasks.

Amazon SageMaker Ground Truth for data labeling

Why it fails: Ground Truth is for creating labeled datasets, not for retrieval-augmented generation.

Amazon DynamoDB as the knowledge base store

Why it fails: DynamoDB is a key-value and document database, not designed for vector similarity search.

Amazon Kendra for indexing the knowledge base

Why it fails: Kendra is a search service but not the vector store used directly by Bedrock Knowledge Bases.

Fine-tune the model using a custom dataset and deploy as a custom model

Why it fails: Fine-tuned models may still be subject to data retention policies.

Use the on-demand model through Amazon Bedrock

Why it fails: On-demand may not provide the required data isolation for HIPAA.

Use a third-party model hosted outside of AWS

Why it fails: Third-party models may not have HIPAA-compliant data handling.

Use chain-of-thought prompting to step through the code

Why it fails: Chain-of-thought focuses on reasoning, not security.

Provide few-shot examples of secure code

Why it fails: Few-shot examples help but may not be sufficient for all cases.

Set max_tokens to a low value to limit output

Why it fails: Token limit does not affect code security.

Fine-tune the model on a dataset of valid SQL queries

Why it fails: Fine-tuning improves accuracy but does not guarantee safety.

Use a separate model to validate the SQL after generation

Why it fails: Post-generation validation adds latency and may not catch all issues.

Limit the max_tokens to 50 to reduce complexity

Why it fails: Token limit does not ensure SQL validity.

Reduce the temperature parameter to 0.1

Why it fails: Lower temperature makes output more deterministic but does not fix factual errors.

Increase the max_tokens to allow longer descriptions

Why it fails: Longer output does not improve accuracy.

Use human reviewers to correct errors after generation

Why it fails: Human review is a workaround, not a model improvement.

Train a custom model using Amazon SageMaker

Why it fails: SageMaker requires more effort than using Comprehend.

Fine-tune a large language model on labeled emails

Why it fails: Fine-tuning is resource-intensive for simple classification.

Use Amazon Lex with a classifier intent

Why it fails: Lex is designed for chatbots, not document classification.

Accuracy

Why it fails: Accuracy is for classification, not generation.

Perplexity

Why it fails: Perplexity measures how well the model predicts text, not quality.

BLEU score

Why it fails: BLEU is designed for translation, not general text generation.

Perform supervised fine-tuning on a curated dataset of safe responses.

Why it fails: Supervised fine-tuning alone may not sufficiently reduce toxicity without explicit reward signals.

Use prompt engineering to instruct the model to avoid toxic language.

Why it fails: Prompt engineering is a runtime technique, not a training/fine-tuning approach.

Implement adversarial validation by testing against toxic inputs.

Why it fails: Adversarial validation is a testing technique, not a training method to adjust model weights.

The API request is missing a required parameter such as 'prompt'.

Why it fails: The error specifically mentions the 'maxTokens' parameter, not a missing required parameter.

The AWS region does not support the AI21 Labs model.

Why it fails: The error is a client-side 400 error, not a regional availability issue.

The value of 'maxTokens' exceeds the model's maximum limit.

Why it fails: The error states the parameter is not supported, not that the value is out of range.

Implement a post-processing step to replace out-of-stock recommendations with in-stock alternatives.

Why it fails: Post-processing is reactive and does not prevent the model from generating out-of-stock recommendations in the first place.

Fine-tune the model on a dataset of past successful subject lines that only include in-stock products.

Why it fails: Fine-tuning is more expensive and time-consuming than using a system prompt, and the company wants to avoid retraining.

Use a retrieval-augmented generation (RAG) approach to retrieve a list of in-stock products and include it in the prompt.

Why it fails: RAG can help but does not force the model to only use the retrieved list; it may still generate out-of-stock items.

Use a pre-built sentiment analysis API

Why it fails: A pre-built sentiment analysis API may not capture domain-specific nuances in customer feedback and is less effective than fine-tuning with a small labeled dataset.

Fine-tune a foundation model on their dataset

Why it fails: It presents fine-tuning as a viable approach, but the question has a duplicate option. The correct answer is option C, which represents the same approach but is correctly marked as correct. In a properly constructed exam, only one fine-tuning option would appear, and it would be the correct answer.

Use zero-shot classification

Why it fails: Zero-shot classification does not use any labeled data and is less accurate than fine-tuning when a small labeled dataset is available.

AWS Lambda

Why it fails: Lambda is for running code, not a model API.

Amazon Bedrock

Why it fails: This is a duplicate of the correct option A. Although Amazon Bedrock provides the capability, only option A is the designated correct answer. Selecting this duplicate would be incorrect.

Amazon SageMaker Studio

Why it fails: Studio is an IDE, not a model API.

Add dropout layers

Why it fails: Dropout is regularization but not commonly added to fine-tuned models.

Increase learning rate

Why it fails: Raising the learning rate makes updates larger, which typically worsens the training-versus-validation gap or destabilises convergence. It is tempting because a higher rate can escape plateaus, but the symptom here is overfitting, addressed by regularisation or more data, not by amplifying gradient steps.

Use a different foundation model

Why it fails: High training accuracy with poor validation accuracy indicates overfitting, which regularisation, dropout or more data addresses; swapping the foundation model discards the fine-tuning work without touching the cause. Changing models suits a task mismatch, where the architecture or pretraining domain cannot represent the problem at all.

Supported output modalities

Why it fails: Output modalities describe whether a model returns text, images or audio; for a text generation task the output is already fixed as text, so this factor does not discriminate between candidates. It matters when selecting a multimodal model for tasks requiring image or speech output alongside text.

Availability of automatic scaling

Why it fails: Automatic scaling concerns the inference infrastructure hosting the model, not the model's own suitability for text generation. It becomes a genuine selection factor when comparing managed endpoints or deployment options on throughput and elasticity, rather than comparing foundation models themselves.

Guaranteed output accuracy

Why it fails: Incorrect. Bedrock does not guarantee output accuracy; model outputs are probabilistic and should be validated.

The KnowledgeBaseId is not passed correctly

Why it fails: KnowledgeBaseId uses !Ref, which is correct.

The embedding model is not available in the region

Why it fails: The embedding model arn is region-specific and assigned via Sub.

The OpenSearch collection type should be SEARCH not VECTORSEARCH

Why it fails: VECTORSEARCH is correct for vector storage.

SageMaker Managed Spot Training

Why it fails: Spot Training is a cost-optimization feature for training jobs, not a capability for fine-tuning itself.

SageMaker Canvas

Why it fails: Canvas is a no-code ML tool for building models, not for fine-tuning foundation models.

SageMaker Autopilot

Why it fails: Autopilot automates model selection and training but is not designed for fine-tuning existing foundation models.

Amazon GuardDuty

Why it fails: GuardDuty monitors CloudTrail, VPC flow logs and DNS for malicious activity or compromised credentials; it never inspects generated image content. It tempts because it is a security service, and would be correct for detecting anomalous API calls or cryptomining against the Bedrock workload.

AWS WAF

Why it fails: AWS WAF filters inbound HTTP requests against web ACL rules at the edge, so it cannot assess images already produced by Bedrock. It tempts because it enforces content-style rules, and would be correct for blocking malicious request patterns reaching a public web application.

Amazon Comprehend

Why it fails: Comprehend performs natural language processing on text, so it cannot evaluate generated image pixels for policy violations. It is tempting because it detects harmful content in text, and would be correct for screening prompt or caption text rather than the image output itself.