AIF-C01 practice questions and answers

Correct answer: A, B — It is pre-trained on a very large, broad corpus · It can be adapted to many downstream tasks without retraining from scratch

Foundation models are defined by broad pre-training at scale and by adaptability: prompting, RAG, or light fine-tuning steers one base model to many tasks. They notably do not need labelled data for every task, they generate text, images, and other modalities rather than only numbers, and they typically require accelerated hardware.

Correct answer: B — The model is overfitting and has memorised the training data

A large gap between strong training performance and weak test performance is the signature of overfitting: the model has memorised noise rather than learning a general pattern. Underfitting would show poor scores on both sets. A matching distribution would produce similar scores, and a low learning rate slows training rather than creating this gap.

Correct answer: B — Time series forecasting

Predicting a numeric quantity at future timestamps from historical observations ordered in time is exactly time series forecasting. Binary classification predicts one of two categories, anomaly detection flags unusual points rather than projecting future values, and topic modelling groups documents by subject.

Correct answer: B — Telling users they are interacting with an AI system and being clear about its limitations

Transparency is about the user understanding that AI is involved and what it can and cannot be trusted to do. Publishing weights is a licensing decision, private logging is an internal audit control, and TLS is a confidentiality measure rather than a disclosure practice.

Correct answer: B — In the IAM permissions and action group definitions backing the agent

A prompt is guidance, not a security boundary, so a destructive capability the agent must never have should simply not be granted in its IAM role or action group. Temperature and token limits are generation settings and provide no authorisation control.

Correct answer: A — Adjust the chunking strategy and add overlap between chunks

Broken chunks are an ingestion problem, so tuning chunk size and adding overlap so ideas are not split across boundaries fixes the root cause. Temperature does not repair truncated context, a bigger context window still receives the same damaged chunks, and disabling retrieval removes the grounding the application needs.

Correct answer: A, B — Ground answers in retrieved source documents · Instruct the model to answer only from the provided context and say when it does not know

Grounding in retrieved sources and instructing the model to abstain when the context is silent are the two standard mitigations. Higher temperature increases invention, stripping the system prompt removes the guardrail instructions, and concurrency is a scaling concern with no effect on factuality.

Correct answer: A — Model Evaluation jobs, with automatic or human review

Model Evaluation runs your prompt dataset against candidate models and scores the results automatically or through a human workforce, which is exactly a model bake-off. Provisioned Throughput reserves capacity, Guardrails filter content, and knowledge bases provide retrieval.

Correct answer: B — Customer content is not used to train the underlying base models, and data is encrypted in transit and at rest

AWS states that customer content submitted to Bedrock is not used to train the underlying base models, and the service encrypts data in transit and at rest. The other options misstate both the training policy and the platform's built-in encryption.

Correct answer: A, B — Audit the dataset for representation gaps across relevant groups · Measure model performance separately for each group rather than only in aggregate

Bias mitigation starts with knowing the data's composition and continues with disaggregated evaluation so a strong aggregate score cannot hide a weak subgroup. Longer training on skewed data entrenches the skew, removing validation destroys the evidence, and temperature is unrelated.

Correct answer: A — Prompt caching of the shared prefix

When a large block of context repeats across requests, caching that prefix avoids paying full input token price to reprocess it every time. Temperature does not affect cost, extra examples add tokens, and a longer context model does nothing to remove the repeated payload.

Correct answer: A — Bedrock Agents

Agents let a model plan multi-step work and invoke defined action groups backed by APIs or Lambda functions, then compose an answer from the results. Model Evaluation scores model quality, Provisioned Throughput reserves inference capacity, and Guardrails filter unwanted content.

Correct answer: B — Using a trained model to produce a prediction from new input

Inference is the serving step: the trained model receives unseen input and returns a prediction. Adjusting weights from labelled examples is training, choosing features is feature engineering, and splitting the dataset is a data preparation step that happens before training.

Correct answer: A, B — Grant the agent the minimum IAM permissions its legitimate tasks require · Validate and constrain tool inputs and outputs rather than trusting model-generated arguments

Prompt injection is contained by limiting blast radius through least privilege and by treating model-produced tool arguments as untrusted input that must be validated. Trusting instructions embedded in retrieved content is the attack itself, administrator access maximises damage, and disabling logging removes your detection.

Correct answer: B — Recall on the fraud class

Recall on the minority class measures how many actual fraud cases the model caught, and an always-negative model has a recall of zero. Accuracy would read a misleading 99.9% because the classes are so imbalanced. Mean squared error and R-squared are regression metrics and do not apply to a classification output.

Correct answer: A — Amazon SageMaker Model Cards

Model Cards capture intended use, risk rating, training details, and evaluation results in one governance artefact reviewers can sign off on. CloudWatch shows operational metrics, Cost Explorer analyses spend, and S3 Inventory lists objects.

Correct answer: A — Amazon SageMaker Clarify

Clarify computes pre-training and post-training bias metrics across sensitive attributes and also produces explainability reports. Macie discovers sensitive data in S3, Config tracks resource configuration compliance, and Inspector scans workloads for software vulnerabilities.

Correct answer: A — Whether the retrieved chunks actually contain the answer

Retrieval quality is judged on whether the right passages came back, which you can check independently of what the model then wrote. Fluency reflects the generator, latency is a performance metric, and system prompt length is a configuration detail.

Correct answer: A — Fairness

Systematically different outcomes for comparable candidates across groups is a fairness and bias problem, and it calls for bias measurement and mitigation before deployment. Latency, elasticity, and durability are performance and infrastructure properties with no bearing on this.

Correct answer: B — Revise the generation prompt to instruct the model to use all supplied context and increase the output token limit

If retrieval is working, the gap is in generation, so the fix is a prompt that tells the model to synthesise across all provided passages plus enough output budget to include them. Cutting retrieval to one result removes information, rebuilding with random embeddings destroys search quality, and maximum temperature makes output less faithful.

Correct answer: A — Supervised learning, classification

The label exists (defaulted or not) so the training is supervised, and the target is a discrete category, which makes it classification. Regression would apply if the target were a continuous number such as the loss amount. Clustering has no label at all, and reinforcement learning has no labelled history to learn from in this setup.

Correct answer: B — Unsupervised learning

Clustering customers with no labelled target is the definition of unsupervised learning: the algorithm finds structure in unlabelled data. Supervised learning needs labelled examples of the answer, which do not exist here. Reinforcement learning learns from reward signals in an environment, and transfer learning is a technique for reusing a trained model, not a learning paradigm for this task.

Correct answer: B — A numeric vector that represents the meaning of text so similar text sits close together

An embedding maps content into a vector space where semantic similarity becomes geometric closeness, which is what makes vector search possible. It is not a retrievable copy of the source text, not the system prompt, and not a model variant.

Correct answer: A — Amazon Macie

Macie uses managed data identifiers to discover and classify sensitive data such as PII in S3 and reports findings. GuardDuty detects threat activity, Secrets Manager stores credentials, and EventBridge routes events.

Correct answer: B — Fine-tuning adapts a pre-trained model to a narrower task using a labelled dataset

Fine-tuning continues training a pre-trained model on task-specific labelled examples so it specialises in a narrower domain or format. Prompts still matter afterwards, fine-tuning carries training and hosting costs that prompt engineering does not, and no training technique eliminates hallucination.

Correct answer: A — Text-to-image generation

Producing an image from a written description is text-to-image generation, offered by image foundation models. Named entity recognition pulls entities out of text, diarisation separates speakers in audio, and OCR reads text out of an existing image rather than creating one.

Correct answer: A — Service control policies in AWS Organizations combined with IAM policies naming approved model resources

Guardrails that are actually enforced come from policy: SCPs set the organisational ceiling and IAM policies restrict invocation to approved model ARNs. A wiki and self-reporting are documentation rather than controls, and temperature has nothing to do with model authorisation.

Correct answer: A — Amazon Polly

Polly is the text-to-speech service and turns generated text into audio. Transcribe goes the other direction from speech to text, Lex builds conversational bot flows with intents and slots, and Translate converts text between languages.

Correct answer: B — Human review or a model-based evaluator scoring relevance and faithfulness

Summaries have many acceptable phrasings, so scoring needs judgement about relevance and faithfulness, delivered either by human reviewers or an evaluator model. Exact match fails valid paraphrases, mean squared error against length is meaningless here, and token count measures verbosity rather than quality.

Correct answer: B — Few-shot prompting

Supplying a handful of demonstrations inside the prompt is few-shot prompting, and it steers format and style without changing model weights. Zero-shot supplies no examples, continued pre-training updates the model on new corpora, and RLHF is a training-time alignment method.

Correct answer: B — Use a knowledge base with RAG and return source attributions with the response

A knowledge base returns the retrieved chunks alongside the generated answer, so the application can show exactly which article backed each claim. Temperature affects creativity, a shorter output limit truncates answers, and model size does not create citations.

Correct answer: A — The volume of data and accelerated compute hours required

Pre-training consumes enormous corpora and thousands of accelerator hours, which is why almost no organisation does it and why fine-tuning or prompting is the normal path. API key counts, bucket policy size, and console Region have no meaningful bearing on training cost.

Correct answer: B — Hallucination

A hallucination is fluent, confident output that is factually wrong or invented, which is what a fabricated citation is. Overfitting describes a training failure measured on held-out data, data drift is a change in input distribution over time, and gradient explosion is a numerical training problem.

Correct answer: A — Maintaining lineage of datasets, model versions, and approvals for each deployment

Accountability requires being able to say which data and which model version produced a decision and who approved it, which is what lineage and approval records give you. Unreviewed deploys, deleted evaluations, and shared root credentials each destroy that trail.

Correct answer: A — Amazon SageMaker Canvas

SageMaker Canvas is the no-code visual surface aimed at analysts who want predictions from tabular data without writing training code. EMR runs big data frameworks, Glue DataBrew is a visual data preparation tool that cleans data rather than producing predictions, and Athena runs SQL queries over S3.

Correct answer: B — "Classify the complaint into exactly one of billing, delivery, or quality. Reply with only the label."

A prompt that names the allowed labels and constrains the output format gives the model no room to add prose, which is what downstream parsing needs. The other prompts are open-ended, so the model will return free-form text of unpredictable shape.

Correct answer: A — Amazon SageMaker AI

SageMaker AI is the managed platform for the whole custom ML lifecycle: notebooks, training jobs, tuning, and hosted inference endpoints. Rekognition and Comprehend are pre-trained AI services for images and text that you consume through an API rather than train yourself, and Q Business is a generative AI assistant over enterprise content.

Correct answer: A, B — The user sees the first tokens sooner, which improves perceived responsiveness · Long answers do not appear frozen while the model is still generating

Streaming is a user experience optimisation: time to first token drops and long answers render progressively. It does not change how many output tokens are produced or billed, it does not replace content filtering, and it has no effect on answer quality.

Correct answer: A — Input tokens are billed, so longer context raises per-request cost and latency

Foundation model pricing counts input tokens as well as output tokens, so stuffing the window costs money on every call and adds processing latency. The model does not retain the context between calls, streaming still works with long prompts, and no special networking is required.

Correct answer: A — Amazon Bedrock

Bedrock is the managed service that exposes foundation models from Amazon and several third-party providers behind one serverless API, with no infrastructure to run. Ground Truth is a data labelling service, Kendra is enterprise search, and Deep Learning AMIs are machine images you would manage yourself on EC2.

Correct answer: B — It stores embeddings and returns the chunks most semantically similar to the query

The vector store holds embeddings of the document chunks and performs a nearest-neighbour search so the most relevant passages can be injected into the prompt. It does not hold model weights, it matches by meaning rather than exact string, and it performs no speech processing.

Correct answer: B — Batch transform

Batch transform spins up compute, scores the whole dataset, and shuts down, which suits a large periodic job with no online latency need. A continuously running real-time endpoint bills around the clock for a weekly job, provisioned concurrency reserves capacity you would not use, and edge deployment addresses on-device inference rather than bulk scoring.

Correct answer: B — Computing an exact payroll total that must reconcile to the cent

Exact arithmetic that must reconcile belongs in deterministic code or a database, not a probabilistic text generator. Drafting, summarising, and rewriting are all language tasks where a generative model with human review is a sound fit.

Correct answer: A, B — The task needs a highly specific output style that long prompts fail to enforce reliably · A large, high-quality labelled dataset for the task already exists

Fine-tuning earns its cost when behaviour must be baked in and you have enough labelled data to teach it. Rapidly changing facts belong in retrieval rather than weights, a team avoiding training cost should stay with prompting, and ten examples is a few-shot prompt rather than a training set.

Correct answer: B — Provisioned Throughput

Provisioned Throughput reserves model units for a committed term, which is what gives predictable throughput and latency for steady high-volume traffic. On-demand bills per token without a capacity guarantee, batch inference is for offline jobs, and evaluation jobs measure quality rather than serve traffic.

Correct answer: B — Pick the smallest model that meets the accuracy bar on your evaluation set

Model selection is an evaluation exercise: measure candidates on your own data and take the cheapest one that clears the bar, because simple tasks rarely need frontier capability. Defaulting to the largest, longest-context, or newest model spends money without evidence that the task requires it.

Correct answer: B — To tune hyperparameters and detect overfitting before final testing

A validation set gives an unbiased signal during development so you can compare hyperparameter settings and stop before the model starts memorising the training set. Holding data back reduces rather than increases training volume, it does nothing to label data, and it has no bearing on instance cost.

Correct answer: A — An interface VPC endpoint powered by AWS PrivateLink

A PrivateLink interface endpoint keeps API calls on the AWS network with a private IP inside the VPC, which is the standard answer for no-internet requirements. An internet gateway and a NAT gateway both route traffic out to the internet, and a public ALB exposes an ingress rather than securing egress.

Correct answer: A — Feature attribution such as SHAP values

Feature attribution methods like SHAP quantify how much each input pushed the prediction one way or the other, which is the standard explainability tool for tabular models. Batch size, multi-AZ deployment, and dataset compression affect training or infrastructure, not interpretability.

Correct answer: A — Data drift, as real-world input has shifted away from the training distribution

A static model degrading over time on unchanged code is the classic signature of drift in the incoming data or the relationship it models, which is why monitoring and periodic retraining matter. Instance deprecation, bucket versioning, and extra IAM permissions do not change prediction quality.

Correct answer: B — The output becomes more varied and creative

Temperature flattens the probability distribution over the next token, so higher values make the model more likely to pick less probable words and produce varied output. Lower temperature is what makes output deterministic, output length is controlled by a separate max tokens parameter, and retrieval depth is a retrieval setting, not a sampling one.

Correct answer: A — Amazon Q Business

Amazon Q Business is the managed enterprise assistant with connectors to common content systems and permission-aware retrieval so users only see what they are entitled to. SageMaker Studio is an ML development environment, Personalize builds recommendations, and Glue is an ETL service.

Correct answer: A — Chain-of-thought prompting

Chain-of-thought prompting asks the model to work through intermediate steps, which measurably improves multi-step reasoning. Negative prompting steers away from unwanted content, truncation cuts text, and beam width affects decoding search rather than reasoning structure.

Correct answer: A — AWS IAM identity-based policies

Access to Bedrock model invocation is authorised through IAM policies that name the model resource and the InvokeModel actions. CloudFront signed URLs protect content distribution, Shield Advanced mitigates DDoS, and Route 53 health checks monitor endpoints.

Correct answer: A — AWS CloudTrail

CloudTrail is the API audit log across AWS services and captures the identity, time, source IP, and parameters of each management and data event it supports. Trusted Advisor gives best-practice checks, QuickSight visualises business data, and Batch runs compute jobs.

Correct answer: A — Amazon Comprehend

Comprehend is the managed natural language processing service and returns entities, key phrases, sentiment, and language detection from raw text with no training required. Textract extracts text and structure from scanned documents, Transcribe converts speech to text, and Polly converts text to speech.

Correct answer: A — AWS Artifact

AWS Artifact is the self-service portal for compliance reports and agreements, which auditors typically ask for. Organizations manages multi-account structure, Detective investigates security findings, and Systems Manager operates instances and configuration.

Correct answer: A — A customer managed AWS KMS key

A customer managed KMS key gives the company control over rotation, policy, and revocation, which is the requirement here. An AWS owned key offers no such control, base64 is an encoding rather than encryption, and a permissions boundary limits IAM privileges without encrypting anything.

Correct answer: B — Retrieval Augmented Generation over an indexed document store

RAG retrieves the relevant passages at request time and passes them to the model as context, so refreshing the index is all that a weekly document change requires. Weekly fine-tuning is expensive and slow for content that changes constantly, pre-training from scratch is orders of magnitude more costly, and temperature has nothing to do with factual grounding.

Correct answer: B — Applying a fixed, well-documented tax rule that never changes

A deterministic rule that is fully specified and stable should be implemented as ordinary code, where it is cheaper, exact, and auditable. The other three involve patterns learned from large volumes of historical data that would be impractical to express as hand-written rules, which is where machine learning earns its cost.

Correct answer: A — A clinician reviews and approves every recommendation before it reaches a patient

Human-in-the-loop review before a high-stakes action is exactly what human oversight means in a safety-critical domain. Bigger instances, caching, and multi-AZ deployment improve performance and availability but add no human judgement to the decision.

Correct answer: B — A chunk of text, often a word fragment, that the model reads and generates

Language models operate over tokens, which are subword chunks produced by a tokenizer, and both input and output length are billed in tokens. The word also means an auth credential elsewhere in computing, but in an LLM context it refers to text units, not credentials, model weights, or memory.

Correct answer: A — Guardrails with sensitive information filters

Guardrails apply configurable filters, including sensitive information detection and redaction, to both prompts and responses at request time. Provisioned Throughput is a capacity feature, Model Evaluation is an offline scoring tool, and custom model import brings your own weights without adding content controls.

Correct answer: A, B — Machine learning is a subset of artificial intelligence · Deep learning is a subset of machine learning that uses multi-layer neural networks

The nesting runs AI to ML to deep learning: ML is one way of building AI, and deep learning is the branch of ML built on multi-layer neural networks. AI is the outermost set, not a subset of deep learning, the fields are directly related, and many ML algorithms such as decision trees and linear regression use no neural network at all.

Correct answer: A, B — The model may produce factually incorrect statements presented confidently · The model may reproduce biased language present in its training data

Hallucinated facts and inherited bias are the two headline risks that responsible AI controls are designed to mitigate. Blanket refusal is not a general behaviour, foundation models integrate with APIs routinely through agents and tools, and data handling depends on the service configuration rather than being an automatic deletion guarantee.