Sample questions
Microsoft Azure AI Fundamentals AI-900 practice questions
A company deploys an AI system to screen job applications. The system is a complex neural network that learns patterns from historical hiring data. A rejected candidate asks for an explanation, but the development team cannot describe how the decision was reached. Which Microsoft responsible AI principle is most directly violated?
Trap 1: Fairness
Fairness addresses bias and discrimination; while lack of transparency could hide bias, the core violation here is the inability to explain the decision.
Trap 2: Privacy and security
Privacy and security focus on data protection, not on providing explanations for decisions.
Trap 3: Reliability and safety
Reliability and safety ensure the system operates correctly and without harm, but they do not directly require explainability.
- A
Fairness
Why wrong: Fairness addresses bias and discrimination; while lack of transparency could hide bias, the core violation here is the inability to explain the decision.
- B
Transparency
Transparency requires that systems are explainable and that users are informed about how decisions are made.
- C
Privacy and security
Why wrong: Privacy and security focus on data protection, not on providing explanations for decisions.
- D
Reliability and safety
Why wrong: Reliability and safety ensure the system operates correctly and without harm, but they do not directly require explainability.
What is 'model versioning' and why is it essential in MLOps?
Trap 1: Updating the Python version used to run ML training scripts
Python version management is environment configuration — model versioning tracks trained model artefacts and their lineage.
Trap 2: Releasing new features of the Azure ML service as versioned API…
Azure service versioning is product development — model versioning is the practice of tracking your own trained model iterations.
Trap 3: Managing multiple versions of training data used by different model…
Data versioning is a related but separate practice — model versioning specifically tracks the trained model artefacts.
- A
Updating the Python version used to run ML training scripts
Why wrong: Python version management is environment configuration — model versioning tracks trained model artefacts and their lineage.
- B
Tracking each iteration of a trained model for rollback, A/B testing, auditing, and reproducibility
Model versioning enables safe updates (rollback when new version fails), experimentation (A/B test), and regulatory compliance (audit trail).
- C
Releasing new features of the Azure ML service as versioned API updates
Why wrong: Azure service versioning is product development — model versioning is the practice of tracking your own trained model iterations.
- D
Managing multiple versions of training data used by different model experiments
Why wrong: Data versioning is a related but separate practice — model versioning specifically tracks the trained model artefacts.
What is 'AI transparency' in Microsoft's Responsible AI principles?
Trap 1: Making AI model source code publicly available as open source
Open-source code is one form of transparency — AI transparency broadly covers disclosure of AI use, data, limitations, and decision rationale.
Trap 2: Reporting all AI project costs transparently in financial statements
Financial reporting is accounting — AI transparency is about understanding AI systems' behaviour, data, and limitations.
Trap 3: Making all training data publicly available for independent…
Training data disclosure is one transparency measure — AI transparency broadly applies to all aspects of how AI systems work and are used.
- A
Making AI model source code publicly available as open source
Why wrong: Open-source code is one form of transparency — AI transparency broadly covers disclosure of AI use, data, limitations, and decision rationale.
- B
Ensuring people understand when they're interacting with AI, how it works, and what its limitations are
Transparency covers AI disclosure, model documentation, explainability, and honest uncertainty communication — building informed trust.
- C
Reporting all AI project costs transparently in financial statements
Why wrong: Financial reporting is accounting — AI transparency is about understanding AI systems' behaviour, data, and limitations.
- D
Making all training data publicly available for independent researchers to audit
Why wrong: Training data disclosure is one transparency measure — AI transparency broadly applies to all aspects of how AI systems work and are used.
A company uses Azure OpenAI Service to generate marketing copy. They notice that sometimes the generated text contains repetitive phrases or gets stuck in loops. They want to reduce this behavior without changing the overall creativity of the model. Which parameter should they adjust?
Trap 1: Decrease the temperature parameter.
Decreasing temperature makes the model more deterministic and less random, but it does not specifically target repetition; it may reduce creativity and still produce loops.
Trap 2: Increase the presence_penalty parameter.
Increasing presence_penalty penalizes any token that has appeared at all, which can reduce topic repetition but frequency_penalty is more precise for repetitive phrases.
Trap 3: Decrease the top_p parameter.
Decreasing top_p makes the model consider only a smaller set of likely tokens, which can make outputs more focused but does not directly address repetition of already used tokens.
- A
Increase the frequency_penalty parameter.
Correct. Increasing frequency_penalty reduces the likelihood that the model will repeat the same tokens, making it effective against repetitive loops.
- B
Decrease the temperature parameter.
Why wrong: Decreasing temperature makes the model more deterministic and less random, but it does not specifically target repetition; it may reduce creativity and still produce loops.
- C
Increase the presence_penalty parameter.
Why wrong: Increasing presence_penalty penalizes any token that has appeared at all, which can reduce topic repetition but frequency_penalty is more precise for repetitive phrases.
- D
Decrease the top_p parameter.
Why wrong: Decreasing top_p makes the model consider only a smaller set of likely tokens, which can make outputs more focused but does not directly address repetition of already used tokens.
A data scientist is training a regression model to predict house prices using features like square footage, number of bedrooms, and location. After evaluating the model on a test set, the data scientist wants to select a metric that measures the average magnitude of prediction errors in the same units as the target variable (price). Which evaluation metric should the data scientist use?
Trap 1: Accuracy
Accuracy is a classification metric that measures the proportion of correct predictions, not suitable for continuous output.
Trap 2: F1 Score
F1 score is a classification metric that balances precision and recall, not used for regression.
Trap 3: Precision
Precision is a classification metric that measures the proportion of true positives among positive predictions, not applicable to regression.
- A
Root Mean Squared Error (RMSE)
RMSE measures the average magnitude of prediction errors in the original units, making it suitable for regression.
- B
Accuracy
Why wrong: Accuracy is a classification metric that measures the proportion of correct predictions, not suitable for continuous output.
- C
F1 Score
Why wrong: F1 score is a classification metric that balances precision and recall, not used for regression.
- D
Precision
Why wrong: Precision is a classification metric that measures the proportion of true positives among positive predictions, not applicable to regression.
A company uses Azure OpenAI Service to generate marketing copy. They want to ensure that the generated text does not contain offensive language or harmful stereotypes, even if the prompt inadvertently leads the model in that direction. Which Azure OpenAI feature should they configure to help prevent such outputs?
Trap 1: Prompt engineering
Prompt engineering involves crafting the input prompt to guide the model, but it cannot guarantee that the model will avoid generating harmful content on its own.
Trap 2: Fine-tuning
Fine-tuning adapts the model to a specific style or domain but does not automatically filter harmful content from the output.
Trap 3: Few-shot learning
Few-shot learning provides examples in the prompt to set context, but it does not provide a safety filter against offensive outputs.
- A
Content filtering
Content filtering applies safety rules to block offensive or harmful language in model outputs, regardless of the prompt's phrasing.
- B
Prompt engineering
Why wrong: Prompt engineering involves crafting the input prompt to guide the model, but it cannot guarantee that the model will avoid generating harmful content on its own.
- C
Fine-tuning
Why wrong: Fine-tuning adapts the model to a specific style or domain but does not automatically filter harmful content from the output.
- D
Few-shot learning
Why wrong: Few-shot learning provides examples in the prompt to set context, but it does not provide a safety filter against offensive outputs.
A medical research team wants to analyze MRI scans to identify and measure the precise boundaries of tumors. They need to assign each pixel in the image to a class (e.g., tumor, healthy tissue, background). Which Azure Computer Vision capability should they use?
Trap 1: Object Detection
Object detection identifies objects and their bounding boxes (rectangular regions), but does not assign a class to every pixel, which is required for precise boundary measurement.
Trap 2: Image Classification
Image classification assigns a label to the entire image, not to individual pixels. It cannot provide pixel-level boundaries for tumors.
Trap 3: Optical Character Recognition
OCR extracts text from images, not relevant for analyzing tumor boundaries in MRI scans.
- A
Object Detection
Why wrong: Object detection identifies objects and their bounding boxes (rectangular regions), but does not assign a class to every pixel, which is required for precise boundary measurement.
- B
Image Classification
Why wrong: Image classification assigns a label to the entire image, not to individual pixels. It cannot provide pixel-level boundaries for tumors.
- C
Semantic Segmentation
Semantic segmentation classifies every pixel, producing a detailed map of regions (e.g., tumor boundaries), which is exactly what the research team needs.
- D
Optical Character Recognition
Why wrong: OCR extracts text from images, not relevant for analyzing tumor boundaries in MRI scans.
A marketing team uses Azure OpenAI Service to generate marketing copy. They notice the generated text is often repetitive, using the same phrases and words multiple times. Which parameter should they increase to directly reduce this repetition?
Trap 1: Temperature
Temperature controls the randomness of token selection. Increasing it makes the model more creative but does not specifically penalize repetition.
Trap 2: Top-p
Top-p controls the cumulative probability of considered tokens; it influences diversity but is not specifically designed to penalize repetition.
Trap 3: Max tokens
Max tokens only sets the maximum length of the output; it does not affect whether the model repeats itself.
- A
Temperature
Why wrong: Temperature controls the randomness of token selection. Increasing it makes the model more creative but does not specifically penalize repetition.
- B
Frequency penalty
Correct. A higher frequency penalty reduces the likelihood of the model repeating the same tokens and phrases, directly addressing repetition.
- C
Top-p
Why wrong: Top-p controls the cumulative probability of considered tokens; it influences diversity but is not specifically designed to penalize repetition.
- D
Max tokens
Why wrong: Max tokens only sets the maximum length of the output; it does not affect whether the model repeats itself.
A data scientist trains a classification model to predict whether an email is spam or not. The model achieves 98% accuracy on the test set, but upon inspection, it classifies all emails as 'not spam' because the dataset has 95% non-spam emails. What is the most likely issue?
Trap 1: Overfitting
Overfitting occurs when a model learns training data too well and fails to generalize, typically resulting in a large gap between training and test accuracy. Here the accuracy is suspiciously high on the test set, but the model is trivial, not overfitted.
Trap 2: Underfitting
Underfitting happens when a model is too simple to capture patterns, leading to low accuracy on both training and test sets. That does not match the 98% accuracy.
Trap 3: Feature scaling error
Feature scaling issues can affect model convergence or performance, but they do not cause a model to output a constant class prediction unless combined with other factors.
- A
Overfitting
Why wrong: Overfitting occurs when a model learns training data too well and fails to generalize, typically resulting in a large gap between training and test accuracy. Here the accuracy is suspiciously high on the test set, but the model is trivial, not overfitted.
- B
Underfitting
Why wrong: Underfitting happens when a model is too simple to capture patterns, leading to low accuracy on both training and test sets. That does not match the 98% accuracy.
- C
Data imbalance
Data imbalance, where one class vastly outnumbers the other, can cause a model to predict the majority class exclusively. Accuracy is misleading in such cases; the model has not learned to identify spam.
- D
Feature scaling error
Why wrong: Feature scaling issues can affect model convergence or performance, but they do not cause a model to output a constant class prediction unless combined with other factors.
A quality assurance team at a software company uses Azure OpenAI Service to generate compliance reports. They need the model to produce the exact same output for a given prompt every time the API is called, to ensure reproducibility during testing. Which parameter should they set to achieve this deterministic behavior?
Trap 1: Set frequency penalty to 1
Frequency penalty reduces repetition of tokens but does not eliminate randomness; outputs may still vary.
Trap 2: Set top_p to 1
Top_p controls nucleus sampling; setting it to 1 includes all tokens, which still allows random choices and does not guarantee determinism.
Trap 3: Set max_tokens to the expected output length
Max_tokens limits the length of the output but does not affect the randomness of token selection.
- A
Set temperature to 0
Temperature controls randomness; setting it to 0 makes the model choose the most likely token every time, producing deterministic outputs.
- B
Set frequency penalty to 1
Why wrong: Frequency penalty reduces repetition of tokens but does not eliminate randomness; outputs may still vary.
- C
Set top_p to 1
Why wrong: Top_p controls nucleus sampling; setting it to 1 includes all tokens, which still allows random choices and does not guarantee determinism.
- D
Set max_tokens to the expected output length
Why wrong: Max_tokens limits the length of the output but does not affect the randomness of token selection.
A data scientist trains a binary classification model to detect fraudulent credit card transactions. The dataset contains 99.5% legitimate transactions and 0.5% fraudulent transactions. The model predicts every transaction as legitimate and achieves 99.5% accuracy on the test set. Which metric would best reveal that the model is failing to identify any fraudulent transactions?
Trap 1: Precision
Precision measures how many of the predicted fraudulent transactions are actually fraudulent. Since the model predicts zero fraud, precision is undefined (or 0/0), making it uninformative.
Trap 2: F1 score
The F1 score is the harmonic mean of precision and recall. With recall at 0, the F1 score would also be 0, but recall is the more direct indicator because it specifically measures the model's ability to find positive cases.
Trap 3: Mean Absolute Error (MAE)
MAE is a metric for regression tasks (predicting continuous values), not for binary classification, so it is not applicable here.
- A
Precision
Why wrong: Precision measures how many of the predicted fraudulent transactions are actually fraudulent. Since the model predicts zero fraud, precision is undefined (or 0/0), making it uninformative.
- B
Recall
Recall measures the fraction of actual fraudulent transactions that are correctly detected. The model catches none, so recall is 0, clearly showing the model's failure.
- C
F1 score
Why wrong: The F1 score is the harmonic mean of precision and recall. With recall at 0, the F1 score would also be 0, but recall is the more direct indicator because it specifically measures the model's ability to find positive cases.
- D
Mean Absolute Error (MAE)
Why wrong: MAE is a metric for regression tasks (predicting continuous values), not for binary classification, so it is not applicable here.
A developer uses Azure OpenAI Service to generate product reviews for an e-commerce site. The developer notices that the model often repeats the same phrases within the same review, making the output sound unnatural. Which parameter should the developer adjust to reduce this repetition?
Trap 1: Temperature
Temperature influences randomness in token selection, not specifically repetition.
Trap 2: Top_p
Top_p (nucleus sampling) controls the cumulative probability threshold for token selection, but does not directly penalize repetition.
Trap 3: Max_tokens
Max_tokens sets the maximum length of the output; it does not affect repetition within the generated text.
- A
Temperature
Why wrong: Temperature influences randomness in token selection, not specifically repetition.
- B
Top_p
Why wrong: Top_p (nucleus sampling) controls the cumulative probability threshold for token selection, but does not directly penalize repetition.
- C
Max_tokens
Why wrong: Max_tokens sets the maximum length of the output; it does not affect repetition within the generated text.
- D
Frequency_penalty
Frequency_penalty reduces the likelihood of repeating tokens that have already been used, directly addressing the repetition issue.
A data scientist trains a regression model to predict house prices. The model has a mean absolute error (MAE) of $5,000 on the test set. Which statement best interprets this metric?
Trap 1: The model is accurate 95% of the time.
Accuracy is a classification metric, not used for regression. MAE does not indicate a percentage of correct predictions.
Trap 2: The model's predictions are within $5,000 of the actual prices for…
This describes a quantile error (e.g., median absolute error), not the mean absolute error.
Trap 3: The square root of the average squared error is $5,000.
This describes Root Mean Squared Error (RMSE), which penalizes larger errors more heavily. MAE does not involve squaring.
- A
On average, the model's predictions are $5,000 away from the actual prices.
Correct. Mean Absolute Error (MAE) is the average absolute difference between predicted and actual values.
- B
The model is accurate 95% of the time.
Why wrong: Accuracy is a classification metric, not used for regression. MAE does not indicate a percentage of correct predictions.
- C
The model's predictions are within $5,000 of the actual prices for 50% of the houses.
Why wrong: This describes a quantile error (e.g., median absolute error), not the mean absolute error.
- D
The square root of the average squared error is $5,000.
Why wrong: This describes Root Mean Squared Error (RMSE), which penalizes larger errors more heavily. MAE does not involve squaring.
A data scientist trains a regression model to predict house prices using features like bedrooms, square footage, and location. The model achieves an R-squared of 0.95 on the test set. However, when deployed to predict prices in a new city with different property characteristics, the predictions are very inaccurate. Which concept best explains this poor performance?
Trap 1: Underfitting
Underfitting would cause poor performance on both training and test sets, which is not the case here.
Trap 2: High bias
High bias is associated with underfitting, where the model is too simple to capture patterns. This model shows high variance (overfitting).
Trap 3: Data drift
Data drift refers to a gradual change in data distribution over time, but the issue here is immediate poor generalization to a different city, not a temporal change.
- A
Overfitting
The model performs well on the original test set but fails on data from a different distribution (new city), which is a classic sign of overfitting.
- B
Underfitting
Why wrong: Underfitting would cause poor performance on both training and test sets, which is not the case here.
- C
High bias
Why wrong: High bias is associated with underfitting, where the model is too simple to capture patterns. This model shows high variance (overfitting).
- D
Data drift
Why wrong: Data drift refers to a gradual change in data distribution over time, but the issue here is immediate poor generalization to a different city, not a temporal change.
A logistics company receives thousands of handwritten shipping labels daily. They need an automated solution to extract the destination address, sender name, and package weight from these labels. Which prebuilt Azure Computer Vision capability should they use?
Trap 1: Object detection
Object detection identifies and locates objects in images, not for reading text.
Trap 2: Image classification
Image classification assigns a label to the entire image, not for extracting text.
Trap 3: Facial recognition
Facial recognition identifies or verifies individuals from their faces, irrelevant for text extraction.
- A
Optical Character Recognition (OCR)
OCR extracts text (including handwriting) from images, perfect for reading shipping labels.
- B
Object detection
Why wrong: Object detection identifies and locates objects in images, not for reading text.
- C
Image classification
Why wrong: Image classification assigns a label to the entire image, not for extracting text.
- D
Facial recognition
Why wrong: Facial recognition identifies or verifies individuals from their faces, irrelevant for text extraction.
A manufacturing company wants to use Azure Computer Vision to inspect products on an assembly line for defects. They have a labeled dataset with images of defective and non-defective products. They need to not only classify products as defective or not, but also identify the exact location of the defect (e.g., a crack) in the image. Which Azure Computer Vision capability should they use?
Trap 1: Custom Vision image classification
Image classification only labels the entire image (e.g., defective or not) and does not indicate where the defect is located.
Trap 2: Azure Face API
Face API is designed for detecting and analyzing human faces, not for general defect detection.
Trap 3: Optical Character Recognition (OCR)
OCR is used to extract printed or handwritten text from images, not for detecting defects.
- A
Custom Vision object detection
Correct. Object detection can be trained to identify and locate defects (e.g., cracks) within an image, providing both classification and location.
- B
Custom Vision image classification
Why wrong: Image classification only labels the entire image (e.g., defective or not) and does not indicate where the defect is located.
- C
Azure Face API
Why wrong: Face API is designed for detecting and analyzing human faces, not for general defect detection.
- D
Optical Character Recognition (OCR)
Why wrong: OCR is used to extract printed or handwritten text from images, not for detecting defects.
A developer wants to build a virtual assistant that can understand user intents such as 'Book a flight' or 'Check weather' and extract relevant entities like destination and date. The developer has a small set of labeled example utterances. Which Azure AI Language feature should the developer use?
Trap 1: Custom Question Answering
Custom Question Answering is designed to answer questions from a knowledge base, not to understand conversational intents and extract entities.
Trap 2: Custom Text Classification
Custom Text Classification assigns a single label to an entire document, but it does not extract entities or handle multiple intents in a conversational context.
Trap 3: Named Entity Recognition (NER)
NER extracts predefined or custom entities but does not identify intents. It is only part of the solution.
- A
Custom Question Answering
Why wrong: Custom Question Answering is designed to answer questions from a knowledge base, not to understand conversational intents and extract entities.
- B
Conversational Language Understanding (CLU)
CLU is specifically designed to extract intents and entities from conversational utterances. It can be trained with labeled examples to understand various user goals.
- C
Custom Text Classification
Why wrong: Custom Text Classification assigns a single label to an entire document, but it does not extract entities or handle multiple intents in a conversational context.
- D
Named Entity Recognition (NER)
Why wrong: NER extracts predefined or custom entities but does not identify intents. It is only part of the solution.
A museum wants to automatically transcribe handwritten labels on historical artifacts. The handwriting varies in style and may include numbers and special characters. Which Azure Computer Vision capability should they use?
Trap 1: Image captioning
Image captioning generates a natural language description of an image but does not extract text characters.
Trap 2: Facial recognition
Facial recognition identifies or verifies individuals from face images, not relevant for text extraction.
Trap 3: Object detection
Object detection finds and locates objects (e.g., artifacts) but does not read text.
- A
Image captioning
Why wrong: Image captioning generates a natural language description of an image but does not extract text characters.
- B
Optical Character Recognition (OCR)
OCR extracts text from images, including handwritten text, numbers, and special characters, which matches the requirement.
- C
Facial recognition
Why wrong: Facial recognition identifies or verifies individuals from face images, not relevant for text extraction.
- D
Object detection
Why wrong: Object detection finds and locates objects (e.g., artifacts) but does not read text.
What is 'conversation history' and why is it important in multi-turn chatbot interactions?
Trap 1: A database log of all chatbot conversations for compliance auditing
Audit logs are compliance records — conversation history in the context of LLMs is the prompt context sent to maintain dialogue coherence.
Trap 2: A summary of frequently asked questions generated from past user…
FAQ generation is analytics — conversation history is the real-time context window the model uses to maintain dialogue state.
Trap 3: The total number of messages a user has sent to a chatbot over…
Lifetime message counts are usage metrics — conversation history is the active context needed for coherent multi-turn dialogue.
- A
A database log of all chatbot conversations for compliance auditing
Why wrong: Audit logs are compliance records — conversation history in the context of LLMs is the prompt context sent to maintain dialogue coherence.
- B
The sequence of prior messages included in the prompt so the model can maintain context across turns
Because LLMs are stateless, conversation history must be sent with each request — enabling the model to understand references to previous turns.
- C
A summary of frequently asked questions generated from past user interactions
Why wrong: FAQ generation is analytics — conversation history is the real-time context window the model uses to maintain dialogue state.
- D
The total number of messages a user has sent to a chatbot over their lifetime
Why wrong: Lifetime message counts are usage metrics — conversation history is the active context needed for coherent multi-turn dialogue.
A developer is using Azure OpenAI Service to generate product descriptions. They want the output to be highly focused and deterministic, with less randomness. Which parameter should they decrease?
Trap 1: Max tokens
Max tokens controls the maximum length of the generated text, not the level of randomness.
Trap 2: Top-p
Top-p (nucleus sampling) also affects randomness, but temperature is the more direct control; decreasing temperature is the standard approach for focused outputs.
Trap 3: Frequency penalty
Frequency penalty reduces repetition of tokens, not the overall randomness or creativity.
- A
Temperature
Correct. Decreasing temperature reduces randomness, making the output more deterministic and focused.
- B
Max tokens
Why wrong: Max tokens controls the maximum length of the generated text, not the level of randomness.
- C
Top-p
Why wrong: Top-p (nucleus sampling) also affects randomness, but temperature is the more direct control; decreasing temperature is the standard approach for focused outputs.
- D
Frequency penalty
Why wrong: Frequency penalty reduces repetition of tokens, not the overall randomness or creativity.
A developer is using Azure OpenAI Service to classify customer support tickets into categories such as 'Billing', 'Technical Issue', and 'Account Management'. The developer provides three labeled examples for each category in the prompt to improve the model's accuracy. What technique is the developer applying?
Trap 1: Fine-tuning
Fine-tuning involves retraining the model on a custom dataset, which is not what the developer is doing here.
Trap 2: Prompt engineering
Prompt engineering is a broader field, but the specific technique of using examples in the prompt is called few-shot learning.
Trap 3: Retrieval-augmented generation
Retrieval-augmented generation (RAG) retrieves relevant external information to augment the model's response, not simply including examples in the prompt.
- A
Fine-tuning
Why wrong: Fine-tuning involves retraining the model on a custom dataset, which is not what the developer is doing here.
- B
Few-shot learning
By providing a few labeled examples in the prompt, the developer is using few-shot learning to guide the model's classification behavior without retraining.
- C
Prompt engineering
Why wrong: Prompt engineering is a broader field, but the specific technique of using examples in the prompt is called few-shot learning.
- D
Retrieval-augmented generation
Why wrong: Retrieval-augmented generation (RAG) retrieves relevant external information to augment the model's response, not simply including examples in the prompt.
A developer wants to use Azure OpenAI to generate text that follows a specific style, such as formal business letters. They provide three examples of the desired output format in the prompt and then ask the model to generate a new letter. Which technique is the developer using?
Trap 1: Zero-shot learning
Zero-shot learning does not use any examples in the prompt; the model relies solely on its pre-trained knowledge.
Trap 2: Fine-tuning
Fine-tuning requires retraining the model on a large labeled dataset, not just adding examples to a prompt.
Trap 3: Temperature scaling
Temperature scaling adjusts the randomness of the output, but does not help the model learn a specific format from examples.
- A
Zero-shot learning
Why wrong: Zero-shot learning does not use any examples in the prompt; the model relies solely on its pre-trained knowledge.
- B
Few-shot learning
Few-shot learning involves providing a few examples in the prompt to demonstrate the desired pattern, which the model then follows for new inputs.
- C
Fine-tuning
Why wrong: Fine-tuning requires retraining the model on a large labeled dataset, not just adding examples to a prompt.
- D
Temperature scaling
Why wrong: Temperature scaling adjusts the randomness of the output, but does not help the model learn a specific format from examples.
What is the Azure AI Vision background removal feature used for?
Trap 1: Blurring the background to create depth of field effects
Blurring is a separate photo editing effect — background removal completely removes the background from images.
Trap 2: Identifying what type of background (indoor/outdoor) is in an image
Background categorization is image classification — background removal physically separates subject from background.
Trap 3: Replacing backgrounds in video calls
Video call background replacement uses similar technology but in real-time video — AI Vision background removal processes static images.
- A
Blurring the background to create depth of field effects
Why wrong: Blurring is a separate photo editing effect — background removal completely removes the background from images.
- B
Automatically separating foreground subjects from the background in images
Background removal isolates the main subject from its background, useful for product photography and visual content creation.
- C
Identifying what type of background (indoor/outdoor) is in an image
Why wrong: Background categorization is image classification — background removal physically separates subject from background.
- D
Replacing backgrounds in video calls
Why wrong: Video call background replacement uses similar technology but in real-time video — AI Vision background removal processes static images.
A marketing team uses Azure OpenAI Service to generate ad copy. They notice the model sometimes uses offensive language. Which Azure OpenAI feature should they use to automatically block such content?
Trap 1: Setting the temperature parameter to 0.0
Temperature controls randomness, not content safety. Lowering it makes output more deterministic but does not filter offensive content.
Trap 2: Using the frequency_penalty parameter
Frequency penalty reduces repetition but does not block offensive language.
Trap 3: Configuring the max_tokens parameter
Max_tokens controls the length of the generated output, not its content safety.
- A
Setting the temperature parameter to 0.0
Why wrong: Temperature controls randomness, not content safety. Lowering it makes output more deterministic but does not filter offensive content.
- B
Using the frequency_penalty parameter
Why wrong: Frequency penalty reduces repetition but does not block offensive language.
- C
Enabling content filtering
Content filtering is a built-in safety feature designed to detect and block harmful or offensive content in both prompts and completions.
- D
Configuring the max_tokens parameter
Why wrong: Max_tokens controls the length of the generated output, not its content safety.
Question Discussion
Share a tip, memory trick, or ask about the reasoning behind this question. Do not post real exam questions, leaked content, braindumps, or copyrighted exam material. Comments are moderated and may be removed without notice.
Sign in to join the discussion.