Hard Difficulty Questions

Switch to batch inference and run predictions every 100 milliseconds.

Why wrong: Batch inference increases latency and is not suitable for real-time fraud detection.

Reduce the maximum tree depth and retrain the model.

Why wrong: Reducing tree depth may reduce latency but can also decrease model accuracy.

Apply post-training pruning to remove redundant trees.

Why wrong: Pruning reduces model size but may harm accuracy and is not the best way to reduce latency for a single prediction.

Store images indefinitely for audit

Why wrong: GDPR requires data minimization and limited retention; indefinite storage violates that.

Use celebrity recognition

Why wrong: Celebrity recognition does not avoid processing personal data of non-celebrities.

Use face detection only

Why wrong: Even face detection may process biometric data; GDPR requires explicit consent or anonymization.

Amazon DynamoDB Accelerator (DAX)

Why wrong: DAX is a caching service for DynamoDB, not relevant to security compliance.

AWS WAF

Why wrong: WAF protects web applications, but the primary threat protection for ML pipelines is more comprehensive via GuardDuty.

The action should be "bedrock:InvokeModelWithResponseStream".

Why wrong: InvokeModel is the correct action for synchronous invocation; InvokeModelWithResponseStream is for streaming.

The resource ARN is missing the account ID.

Why wrong: The account ID is present (123456789012).

The statement is missing a condition for the model ID.

Why wrong: Conditions are optional; the resource ARN already specifies the model ID.

Fine-tune the model on a dataset of post-2023 papers and deploy it.

Why wrong: Fine-tuning adds knowledge but does not remove pre-existing knowledge; the model may still fall back on older data.

Set the maxTokens to a low value to force the model to rely on recent context.

Why wrong: Token limits do not affect the model's internal knowledge; it can still use outdated information.

Include a system prompt instructing the model to ignore data before 2023.

Why wrong: System prompts are guidelines; models may still use pre-trained knowledge inadvertently.

Add dropout layers

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

Increase learning rate

Why wrong: Increasing learning rate may cause divergence, not reduce overfitting.

Use a different foundation model

Why wrong: Switching models doesn't address overfitting on current dataset.

Use distributed training across multiple GPUs

Why wrong: Distributed training reduces wall-clock time but still performs full fine-tuning.

Use prompt engineering instead of fine-tuning

Why wrong: Prompt engineering does not modify the model and may not achieve the desired task adaptation.

Full fine-tuning on the new dataset

Why wrong: Full fine-tuning is computationally expensive and risks forgetting original knowledge.

Use a smaller model

Why wrong: Model size does not directly correlate with harmfulness.

Fine-tune the model on a biased dataset

Why wrong: Fine-tuning on biased data would likely increase bias.

Disable all logging

Why wrong: Disabling logging reduces accountability.

The KnowledgeBaseId is not passed correctly

Why wrong: KnowledgeBaseId uses !Ref, which is correct.

The embedding model is not available in the region

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

The OpenSearch collection type should be SEARCH not VECTORSEARCH

Why wrong: VECTORSEARCH is correct for vector storage.

Increase model size to improve performance.

Why wrong: Model size should be chosen based on task; larger models may overfit or be costly.

Deploy the model to production immediately after fine-tuning.

Why wrong: Models should be evaluated and tested before production deployment.

Use an image captioning model and then analyze the caption text

Why wrong: Image captioning loses information needed for defect analysis.

Use a text-to-image generation model and analyze the generated image

Why wrong: Text-to-image models generate images, not answer questions.

Use a separate image analysis model and a text model, then combine outputs

Why wrong: This approach is complex and may not capture cross-modal relationships effectively.

Increase the service quota for Bedrock

Why wrong: Quotas are for usage limits, not model access.

Attach the AmazonBedrockFullAccess IAM policy

Why wrong: IAM policies grant API permissions, but model access is a separate approval.

Use a different AWS Region

Why wrong: The error is about model access, not region availability.

The ECS tasks are not configured to use a NAT gateway to reach the Bedrock endpoint

Why wrong: The VPC endpoint allows private connectivity without a NAT gateway.

The VPC endpoint policy does not grant the ECS task IAM role permission to invoke Bedrock

Why wrong: The policy allows all actions; the issue is network connectivity, not permissions.

The ECS task IAM role does not have the bedrock:InvokeModel permission

Why wrong: Missing permissions would result in an access denied error, not a timeout.

The training data is stored in the wrong S3 bucket

Why wrong: S3 permissions or path issues do not cause container start errors.

The container image does not have the correct entry point

Why wrong: Missing entry point would cause a different error.

The GPU drivers are outdated

Why wrong: Outdated GPU drivers would cause a different error, but SageMaker manages drivers.

Remove the sensitive attribute from the training data and retrain the model.

Why wrong: Removing the attribute may not eliminate bias because other features can act as proxies.

Collect more data from the under-represented demographic group and retrain the model.

Why wrong: Collecting more data is time-consuming and costly; the team should first analyze existing data for bias.

Adjust the model's decision threshold for the affected group after deployment.

Why wrong: Post-hoc adjustments may violate regulatory requirements and are less effective than addressing bias during training.

Fine-tune the model using SageMaker Ground Truth and increase the training epochs.

Why wrong: Fine-tuning addresses style but not immediate prompt adherence; Ground Truth is for labeling.

Increase the max token count and use a larger model variant.

Why wrong: Max tokens apply to text generation, not image generation.

Use a different foundation model and increase the image resolution.

Why wrong: Model switch might help, but resolution alone doesn't fix artifacts.

Fine-tune the model on a dataset of product descriptions that exemplify the desired tone.

Why wrong: Fine-tuning would modify the model, which the company wants to avoid.

Implement prompt chaining by breaking the task into multiple steps, each with its own prompt.

Why wrong: Prompt chaining can help with complex tasks but does not specifically address tone consistency.

Use retrieval-augmented generation (RAG) to pull example descriptions from a database and prepend them to the prompt.

Why wrong: RAG focuses on retrieving relevant information, not enforcing a specific tone.

Configure the SageMaker endpoint to use mutual TLS (mTLS) with client certificates

Why wrong: SageMaker endpoints do not support mTLS; VPC placement is the standard method.

Place the SageMaker endpoint behind an Amazon CloudFront distribution with an origin access identity

Why wrong: CloudFront does not enforce VPC origin; clients can access CloudFront from anywhere.

Use AWS STS to generate temporary credentials and require the client to sign requests with them

Why wrong: Signing requests does not restrict network origin; requests can come from anywhere.

JSON

Why wrong: JSON is also row-oriented and has overhead.

Image (JPEG/PNG)

Why wrong: Image formats are not suitable for tabular data.

CSV

Why wrong: CSV is row-oriented and less efficient for columnar access.

Audio feature extraction

Why wrong: Audio processing is not a built-in transform in Data Wrangler.

Image resizing and normalization

Why wrong: Image transformations are not natively supported; Data Wrangler focuses on tabular and text data.