Hard — AIF-C01 Practice Quiz

Question 1

A financial services company has a central ML platform team in one AWS account that trains and approves models, and several business-unit teams in separate AWS accounts that deploy those models. Compliance requires that only approved model versions be deployable, that model metadata and approval status be centrally tracked, and that teams cannot deploy unapproved models. Which approach best meets these governance requirements?

Accepted Answer

Use a shared Amazon SageMaker Model Registry with model package groups, register models with an approval status, and grant business-unit accounts read/deploy access only to Approved model versions via resource policies and IAM.

Answer

Copy approved model artifacts into an S3 bucket in each business-unit account and let each team deploy directly from their local copy without central tracking.

Answer

Give every business-unit account full IAM access to the central training account so they can pick whichever model they want to deploy.

Answer

Email the approved model artifact files to each business-unit team and rely on a manual checklist to confirm they deploy the correct version.

Question 2

A consumer-facing company operates in a region where privacy regulations grant customers the 'right to be forgotten.' A customer has formally requested that all of their personal data be deleted, including any data used to train the company's machine learning models. The compliance team asks the ML team how to best respond to this request for data used in model training. What is the most appropriate action?

Accepted Answer

Remove the customer's records from the training datasets and retrain or fine-tune the affected models on the updated data, documenting the process for audit purposes

Answer

Delete only the customer's records from the production inference logs, since the trained model no longer stores individual records

Answer

Apply a Bedrock Guardrail to filter out the customer's name from all future model responses

Answer

Rotate the KMS key used to encrypt the training data bucket so the customer's data can no longer be decrypted

Question 3

A retail company runs a customer-facing chatbot on a large foundation model that produces high-quality responses but has high per-request cost and slow response times during peak traffic. The team wants to reduce inference cost and latency while preserving as much response quality as possible for their specific chatbot use case. Which approach best meets these goals?

Accepted Answer

Use model distillation to train a smaller student model using the large model's outputs, then deploy the smaller model for inference

Answer

Increase the temperature parameter on the large model to speed up response generation

Answer

Add more few-shot examples to every prompt so the large model returns answers faster

Answer

Enable provisioned throughput on the large model to permanently lower its per-token price

Question 4

A healthcare software company is building an assistant that must reliably interpret and generate highly specialized oncology terminology and follow domain-specific clinical writing conventions. The underlying knowledge (drug names, protocols) is stable and rarely changes. Testing shows a general foundation model misuses terminology and produces text in the wrong clinical style even when relevant reference documents are supplied in the prompt. Which approach best addresses this problem?

Accepted Answer

Fine-tune the foundation model on a curated dataset of oncology text and clinical writing examples

Answer

Implement a Retrieval Augmented Generation pipeline that injects reference documents into each prompt

Answer

Increase the temperature parameter so the model produces more varied terminology

Answer

Add more reference documents to the knowledge base to expand retrieval coverage

Question 5

A payments company is building a fraud detection model. Fraudulent transactions make up only 0.5% of all transactions in the training dataset. During evaluation, the team notices the model reports 99.5% accuracy but almost never flags real fraud. What is the most likely explanation for this outcome?

Accepted Answer

The dataset is highly imbalanced, so overall accuracy is a misleading metric and the model simply predicts the majority (non-fraud) class.

Answer

The model is overfitting because it was trained for too many epochs on the fraud class.

Answer

The model needs to be switched from supervised learning to unsupervised clustering to detect fraud.

Answer

The inference endpoint is under-provisioned, causing it to drop fraud predictions under load.

Question 6

A healthcare company runs an on-premises data processing application that needs to call Amazon SageMaker inference endpoints in AWS. The security team requires that no long-term AWS access keys be stored on the on-premises servers, and that all access use temporary credentials tied to the servers' existing X.509 certificates issued by the company's private certificate authority. Which approach best meets these requirements?

Accepted Answer

Configure IAM Roles Anywhere with a trust anchor referencing the company's certificate authority, so the servers exchange their X.509 certificates for temporary IAM role credentials.

Answer

Create an IAM user for the application and store its access key and secret key in a configuration file on each on-premises server.

Answer

Embed a hardcoded IAM role ARN in the application and rely on the SageMaker endpoint's resource policy to authorize the on-premises servers by IP address.

Answer

Generate long-term access keys for a shared IAM role and rotate them manually every 90 days using a scheduled script.

Question 7

A team builds a RAG application using Amazon Bedrock. They generated and stored document embeddings in a vector database using one embedding model. Later, they switched the query-time embedding model to a different, newer model to improve results, but kept the previously indexed document vectors unchanged. Users now report that retrieved passages are irrelevant to their questions. What is the most likely cause?

Accepted Answer

The query embeddings and stored document embeddings were produced by different models, so they occupy incompatible vector spaces and cannot be compared meaningfully.

Answer

The temperature parameter on the generation model was set too high, causing random retrieval.

Answer

The context window of the foundation model is too small to hold the retrieved passages.

Answer

The vector database does not support enough dimensions for the newer model's output.

Question 8

A company has deployed a RAG-based customer support assistant using a foundation model and a vector store of product documentation. Users complain that the assistant sometimes gives accurate-sounding but incorrect answers. An engineer suspects the retrieval step is returning irrelevant document chunks. Which evaluation approach BEST helps confirm whether the retrieval component is the source of the problem?

Accepted Answer

Measure the relevance of the retrieved document chunks against the user query (retrieval precision/recall) before the model generates its answer

Answer

Increase the model's temperature parameter to encourage more diverse responses

Answer

Calculate the token throughput and average latency of the endpoint under load

Answer

Fine-tune the foundation model on the full product documentation corpus

Question 9

A media company wants a foundation-model application that generates news summaries in the exact editorial writing style used by their senior editors. They have 5,000 high-quality example summaries written in this style, and the style rarely changes. The underlying news content changes daily and comes from a separate live feed that is already injected into the prompt at request time. The team wants the lowest possible per-request latency and does not want to inject style examples into every prompt. Which approach best meets these requirements?

Accepted Answer

Fine-tune the foundation model on the 5,000 example summaries so the editorial style is learned into the model weights.

Answer

Build a Retrieval Augmented Generation pipeline that retrieves style examples from a vector store and adds them to each prompt.

Answer

Include all 5,000 example summaries in the system prompt as few-shot examples for every request.

Answer

Use continued pre-training on the daily live news feed so the model always has the latest content.

Question 10

A financial services team uses a small foundation model with a RAG pipeline to answer questions about their internal policies. Retrieval is accurate and the correct policy documents are consistently injected into the prompt, but users complain that the model still fails to correctly combine multiple policy rules to reach a conclusion (e.g., determining eligibility that depends on three separate conditions). What is the most likely cause, and what should the team do first?

Accepted Answer

The model lacks sufficient reasoning capability for multi-step logic; evaluate a larger, more capable foundation model.

Answer

Retrieval quality is the problem; increase the top-k value to inject more documents into the prompt.

Answer

The knowledge base is stale; set up more frequent re-indexing of the policy documents.

Answer

The embeddings model is mismatched; switch to a different embeddings model to improve semantic search.

Question 11

A bank deployed a loan-approval ML model six months ago. The team validated fairness across demographic groups before launch. Recently, customer complaints suggest approval rates have shifted unfavorably for certain groups, even though the model code has not changed. Which responsible-AI practice most directly addresses this situation?

Accepted Answer

Continuously monitor the deployed model for data drift and fairness degradation over time

Answer

Publish a model card documenting the model's intended use and training data

Answer

Increase the model's temperature setting to introduce more variability in outputs

Answer

Add a Bedrock guardrail to block requests containing personally identifiable information