Describe features of generative AI workloads on Azure practice questions

B. Azure Computer Vision

Why wrong: Azure Computer Vision analyzes and extracts information from images and videos, not text generation.

C. Language Understanding (LUIS)

Why wrong: LUIS is a conversational AI service for understanding user intent and extracting entities from text, not for generating new text.

D. Azure Machine Learning

Why wrong: Azure Machine Learning is a comprehensive platform for building, training, and deploying custom models, but it would require significant effort to build a generative text model, whereas Azure OpenAI Service provides a pre-built solution.

B. GPT-3 for text and Custom Vision for images

Why wrong: Custom Vision is used for image classification and object detection, not for generating new images from text.

C. BERT for text and OCR for images

Why wrong: BERT is a language model for understanding (not generation), and OCR extracts text from images, not image generation.

D. Language Understanding (LUIS) and Face API

Why wrong: LUIS extracts intents and entities from user input, and Face API analyzes faces in images; neither generates text or images.

Model deployment

Why wrong: Model deployment is about hosting the model, not filtering content.

Token limit

Why wrong: Token limit controls the length of output, not content safety.

Prompt engineering

Why wrong: Prompt engineering helps guide output but doesn't guarantee blocking inappropriate content; content filters are needed for that.

Fine-tuning the model with a custom dataset

Why wrong: Fine-tuning trains the model on specific data to improve performance on a task. However, it does not inherently filter out objectionable content; the model might still generate harmful text if the fine-tuning data contains such patterns.

Increasing the token limit per response

Why wrong: Token limits control the maximum length of the text generated (e.g., 200 tokens). They do not prevent the model from producing harmful content; they only limit how long the output can be.

Using prompt engineering techniques

Why wrong: Prompt engineering (e.g., adding instructions like 'do not use bad language') can help guide the model, but it is not a guaranteed safety mechanism. The model may still generate inappropriate content if the instruction is not perfectly followed. Content filtering is the official safeguard.

Increase the temperature parameter to 1.0 to force more creative responses

Why wrong: Increasing temperature makes the model more creative and random, which would increase the likelihood of generating incorrect or irrelevant information, not reduce it.

Fine-tune the model on the knowledge base documents using supervised learning

Why wrong: Fine-tuning teaches the model new facts but does not guarantee that it will only answer from those facts; it can still hallucinate, especially when asked questions outside the training data.

Use prompt engineering with a system message that tells the model to never make up facts

Why wrong: While helpful, a simple instruction is not reliable. Without actual document context, the model may still invent answers to fill gaps in knowledge.

Max tokens

Why wrong: Max tokens limits the total number of tokens in the output, affecting length, not creativity or diversity.

Top probability

Why wrong: Top probability (nucleus sampling) sets a cumulative probability threshold to filter token choices; lowering it can reduce diversity, raising it has limited effect on creativity.

Frequency penalty

Why wrong: Frequency penalty reduces the tendency to repeat the same words or phrases, which can increase diversity, but temperature is more directly associated with overall creativity and randomness.

Completions API

Why wrong: The Completions API is designed for single-turn prompts and does not manage conversation history natively.

Embeddings API

Why wrong: The Embeddings API converts text into numerical vectors for semantic similarity, not for generating conversational responses.

Fine-tuning

Why wrong: Fine-tuning is a training process to adapt a model to specific data, not an endpoint used during inference for conversation.

Temperature parameter adjustment

Why wrong: Adjusting temperature controls randomness and creativity but does not force the model to adopt a specific style.

Prompt engineering with detailed instructions

Why wrong: While instructions can guide the model, they are less reliable for consistent style adaptation compared to fine-tuning.

Content filtering

Why wrong: Content filtering blocks harmful or offensive text but does not influence brand voice.

Frequency penalty

Why wrong: Frequency penalty discourages the model from repeating the same words, which can reduce repetition but does not directly increase overall randomness.

Top_p (nucleus sampling)

Why wrong: Top_p controls the cumulative probability of token choices; lowering it can make output more focused, increasing it can add diversity but it's not the primary parameter for randomness.

Presence penalty

Why wrong: Presence penalty encourages the model to talk about new topics, which can add diversity in content but does not primarily control randomness in token selection.

Increase the top_p parameter

Why wrong: Top_p controls nucleus sampling and can also increase diversity, but temperature is the primary parameter for adjusting randomness and creativity.

Increase the frequency_penalty parameter

Why wrong: Frequency penalty reduces the likelihood of repeating the same tokens, which helps reduce repetition but does not directly increase creativity or variety.

Decrease the max_tokens parameter

Why wrong: Max_tokens limits the length of the generated response; decreasing it would make outputs shorter, not more creative.

Fine-tune the base model on the example descriptions

Why wrong: Fine-tuning does adapt the model, but it requires retraining and is more heavy-weight; for a small set of examples, few-shot is more practical and cost-effective.

Increase the temperature parameter to the maximum value

Why wrong: Increasing temperature makes the output more random and creative, which is the opposite of matching a specific brand voice consistently.

Train a new model using Azure Machine Learning

Why wrong: This would require building and training a custom model from scratch, which is unnecessary when Azure OpenAI can be guided via prompts.

The Chat Completions API

Why wrong: The Chat Completions API processes prompts and generates responses but does not provide an upfront token count estimation.

The Embeddings API

Why wrong: The Embeddings API converts text into vectors and consumes tokens, but it does not return a token estimate before the call.

The Content Filter configuration

Why wrong: Content filters are designed to block harmful content and do not provide any token estimation functionality.

Max tokens

Why wrong: Max tokens controls the maximum length of the generated text, not the level of randomness.

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.

Frequency penalty

Why wrong: Frequency penalty reduces repetition of tokens, not the overall randomness or creativity.

Temperature

Why wrong: Temperature controls randomness; lowering it makes output more deterministic but does not specifically reduce repetition.

Max tokens

Why wrong: Max tokens sets the maximum length of the generated text; it does not affect repetition.

Top P

Why wrong: Top P (nucleus sampling) controls the cumulative probability threshold for token selection but does not specifically penalize repetition.

Temperature

Why wrong: Temperature controls the randomness of token selection. Increasing it makes the model more creative but does not specifically penalize repetition.

Top-p

Why wrong: Top-p controls the cumulative probability of considered tokens; it influences diversity but is not specifically designed to penalize repetition.

Max tokens

Why wrong: Max tokens only sets the maximum length of the output; it does not affect whether the model repeats itself.

Set frequency penalty to 1

Why wrong: Frequency penalty reduces repetition of tokens but does not eliminate randomness; outputs may still vary.

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.

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.

temperature

Why wrong: Temperature controls the randomness of the output; lowering it makes the output more deterministic, but it does not affect the length of the generated text.

top_p

Why wrong: Top_p (nucleus sampling) controls the diversity of the output by considering only the most probable tokens, but it does not limit the number of tokens generated.

frequency_penalty

Why wrong: Frequency_penalty applies a penalty to tokens that have already appeared in the text, reducing repetition but not controlling the length of the generated content.

Temperature

Why wrong: Temperature controls the randomness of the output, not the length.

top_p

Why wrong: top_p (nucleus sampling) influences diversity by selecting from a cumulative probability, but does not set a length limit.

frequency_penalty

Why wrong: frequency_penalty reduces repetition of token sequences, but does not cap the total length.

frequency_penalty

Why wrong: Increasing frequency_penalty reduces the likelihood of repeating the same tokens, which decreases repetition but does not necessarily make the output more creative or diverse in word choices.

presence_penalty

Why wrong: Increasing presence_penalty encourages the model to talk about new topics by penalizing tokens that have already appeared, but it does not directly increase creativity or unexpected word choices.

top_p

Why wrong: top_p (nucleus sampling) is used to control the cumulative probability of tokens considered. While it can affect diversity, increasing temperature is more directly associated with creativity and unexpected choices.

Top P

Why wrong: Top P (nucleus sampling) also affects diversity, but temperature is the more direct parameter for increasing randomness and variability.

Frequency penalty

Why wrong: Frequency penalty reduces the repetition of words and phrases, but it does not directly increase overall creativity or diversity; it targets repetition.

Max tokens

Why wrong: Max tokens limits the length of the output, which does not affect the diversity of the generated content.