I85PREP
⌘ K
watsonx GenAI Engineer study materials
watsonx GenAI Engineer study resources

watsonx GenAI Engineer practice questions

One question at a time. Build your watsonx GenAI Engineer understanding.

First 50 of 200 questions. Each one lists the choices, the correct answer, and an explanation. The checker at the top of this page saves progress in this browser when it is available.

Continue in the practice tool above to load more questions, 50 at a time.

Question 1watsonx platform fundamentals and architecture

In the watsonx platform, which component is responsible for storing and serving deployed foundation models so applications can call them through an inference endpoint?

  1. The watsonx.data Presto engine
  2. The watsonx.ai model inference service Correct answer.
  3. The watsonx.governance fact sheet registry
  4. The watsonx Orchestrate skill builder

Explanation

watsonx.ai provides the model inference service that hosts and serves foundation models behind REST endpoints so applications can send prompts and receive completions. Governance, data query, and orchestration components serve different purposes in the platform.

Question 2watsonx platform fundamentals and architecture

A team wants a single control plane to manage prompts, model assets, and deployment spaces across projects. Which watsonx capability best matches this need?

  1. watsonx.ai prompt lab and deployment spaces Correct answer.
  2. watsonx.governance risk register
  3. watsonx.data metastore
  4. watsonx Code Assistant

Explanation

watsonx.ai bundles the prompt lab, projects, model asset management, and deployment spaces into one workbench for generative AI development. The other options address data cataloging, governance tracking, or code assistance rather than model and prompt lifecycle management.

Question 3watsonx platform fundamentals and architecture

Which statement accurately describes the relationship between watsonx.ai and watsonx.governance in a typical deployment?

  1. watsonx.governance replaces the watsonx.ai inference service for production traffic
  2. watsonx.governance executes the model weights while watsonx.ai only stores them
  3. watsonx.ai requires watsonx.data to store prompt templates before any model can run
  4. watsonx.governance consumes model metadata and evaluation results produced for models managed in watsonx.ai Correct answer.

Explanation

Governance relies on metadata, fact sheets, and evaluation outputs generated around models managed in watsonx.ai, so the two work together rather than substituting for one another. Inference remains an watsonx.ai responsibility, and prompt storage does not require watsonx.data.

Question 4watsonx platform fundamentals and architecture

What is the primary purpose of a watsonx.ai project?

  1. A collaborative workspace to manage data assets, prompts, and model experiments Correct answer.
  2. A network gateway that routes inference requests between regions
  3. A container registry that stores model images for deployment
  4. A billing account that meters token consumption per department

Explanation

A watsonx.ai project is a collaborative workspace where teams associate data assets, prompts, notebooks, and model experiments. Billing, networking, and image storage are handled by other platform services.

Question 5watsonx platform fundamentals and architecture

A deployment space in watsonx.ai is best described as which of the following?

  1. A repository of governance fact sheets for regulatory review
  2. An environment where trained or tuned model assets are deployed and exposed as online endpoints Correct answer.
  3. A sandbox for editing notebook cells before they are committed to Git
  4. A catalog of tables and schemas used by SQL queries

Explanation

Deployment spaces provide the runtime environment where model assets are promoted and served as online endpoints for applications. Notebook editing, governance records, and table catalogs belong to other parts of the platform.

Question 6watsonx platform fundamentals and architecture

An engineer needs to promote a tuned model from a development project to a production endpoint. Which sequence reflects the intended watsonx.ai workflow?

  1. Save the model asset in the project, promote it to a deployment space, then create an online deployment Correct answer.
  2. Register the model in the governance risk register, then call the register directly for inference
  3. Create the online deployment first, then attach the project notebook, then save the model asset
  4. Export the model to watsonx.data, query it with Presto, then expose the query as an endpoint

Explanation

The platform workflow is to persist the model as an asset in the project, promote it into a deployment space, and then create an online deployment that yields a scoring endpoint. The other sequences invert steps or route through components that do not serve inference.

Question 7watsonx platform fundamentals and architecture

Which watsonx.ai capability lets a user compare outputs from several foundation models side by side using the same prompt?

  1. The governance model inventory
  2. The data virtualization view
  3. The prompt lab with model comparison Correct answer.
  4. The deployment space scoring tab

Explanation

The prompt lab supports experimenting with prompts and comparing responses from multiple foundation models against the same input. Deployment spaces serve already-deployed models, while governance and data views do not run prompt comparisons.

Question 8watsonx platform fundamentals and architecture

What is the role of a foundation model in the watsonx.ai architecture?

  1. A vector index that stores embeddings for retrieval
  2. A scheduler that allocates GPU capacity across tenants
  3. A large pretrained model that can be prompted or tuned for many downstream tasks Correct answer.
  4. A rules engine that validates prompt syntax before submission

Explanation

Foundation models are large pretrained models that serve as the base for prompting and tuning across many tasks. Syntax validation, embedding storage, and capacity scheduling are separate infrastructure functions.

Question 9watsonx platform fundamentals and architecture

A team wants to adapt a foundation model to its proprietary support transcripts without hosting the base weights themselves. Which watsonx.ai approach fits best?

  1. Store the transcripts in a governance fact sheet and prompt the fact sheet
  2. Rewrite the base model weights manually in a notebook
  3. Replace the inference service with a Presto query over the transcripts
  4. Tune the model with the tuning studio using the team's dataset and deploy the tuned asset Correct answer.

Explanation

The tuning studio in watsonx.ai lets teams train an adapted model on their own data and deploy the resulting tuned asset through a deployment space. Manual weight editing, SQL substitution, and fact sheet prompting are not supported adaptation paths.

Question 10watsonx platform fundamentals and architecture

Which statement best explains why watsonx separates projects from deployment spaces?

  1. Projects are billed per token, while deployment spaces are billed per storage gigabyte
  2. Projects support collaborative experimentation, while deployment spaces isolate and govern runtime serving Correct answer.
  3. Projects run only on GPU hardware, while deployment spaces run only on CPU hardware
  4. Projects are shared across all accounts, while deployment spaces are private to one user

Explanation

The separation lets teams experiment freely in projects and then promote vetted assets into deployment spaces that control runtime access and serving. Billing, hardware type, and account visibility are not the basis of this separation.

Question 11watsonx platform fundamentals and architecture

What does watsonx.data primarily provide within the watsonx platform?

  1. A prompt engineering environment for foundation models
  2. A lakehouse that queries open data formats across sources through a shared engine Correct answer.
  3. An orchestration engine for business skills
  4. A governance dashboard for model risk

Explanation

watsonx.data is the data lakehouse layer that lets users query data in open formats across multiple sources using a shared query engine. Prompt engineering, governance, and orchestration are handled by other watsonx components.

Question 12watsonx platform fundamentals and architecture

A RAG application built on watsonx.ai needs to retrieve relevant passages before calling the model. Which architectural element is required to support that retrieval step?

  1. A Presto catalog of the model's hyperparameters
  2. A vector store holding embeddings of the source documents Correct answer.
  3. A deployment space that stores the prompt templates
  4. A fact sheet describing the model's training data

Explanation

Retrieval-augmented generation depends on a vector store that indexes document embeddings so relevant passages can be found at query time. Fact sheets, hyperparameter catalogs, and prompt template storage do not perform similarity search.

Question 13watsonx platform fundamentals and architecture

Which watsonx component is designed to build and run AI-driven business automations that combine skills and assistants?

  1. watsonx.governance
  2. watsonx Orchestrate Correct answer.
  3. watsonx.ai tuning studio
  4. watsonx.data

Explanation

watsonx Orchestrate is the component for composing skills and assistants into business automations. The other components address data querying, governance, and model tuning respectively.

Question 14watsonx platform fundamentals and architecture

An architect must explain how the watsonx platform enforces tenant isolation for model endpoints. Which mechanism is central to that isolation?

  1. Relying on the foundation model to reject foreign requests
  2. IBM Cloud identity and access management scoping resources to accounts and spaces Correct answer.
  3. Storing all tenants' data in one shared deployment space
  4. Embedding tenant identifiers inside each prompt string

Explanation

Isolation is enforced through identity and access management, which scopes resources such as projects and deployment spaces to accounts and authorized users. Prompt-level identifiers, model-side rejection, and shared spaces do not provide reliable isolation.

Question 15watsonx platform fundamentals and architecture

Which statement correctly distinguishes the watsonx.ai inference API from the watsonx.ai tuning API?

  1. The inference API sends prompts to a served model, while the tuning API trains an adapted model on a dataset Correct answer.
  2. The inference API manages deployment spaces, while the tuning API manages projects
  3. The inference API is used only for embeddings, while the tuning API is used only for text generation
  4. The inference API trains model weights, while the tuning API only returns token counts

Explanation

The inference API handles prompt submission and response generation against a served model, whereas the tuning API submits training jobs that produce adapted model assets. Deployment and project management, as well as embeddings versus generation, are not the distinction.

Question 16watsonx platform fundamentals and architecture

What is the purpose of a model fact sheet in watsonx.governance?

  1. To cache prompt responses for faster inference
  2. To define the network routes used by the inference endpoint
  3. To store the model's binary weights for redeployment
  4. To document a model's intended use, training data characteristics, and evaluation results Correct answer.

Explanation

A fact sheet captures governance-relevant documentation such as intended use, data characteristics, and evaluation outcomes for a model. It does not store weights, cache responses, or configure networking.

Question 17watsonx platform fundamentals and architecture

A team deploys a model to a deployment space and later needs to update the served version without changing the endpoint URL used by applications. What watsonx.ai capability supports this?

  1. Updating the deployment within the same space so the endpoint reference remains stable Correct answer.
  2. Moving the model into a governance fact sheet
  3. Creating a new account for each model version
  4. Converting the model into a Presto view

Explanation

Deployments within a space can be revised so applications keep calling a stable endpoint while the underlying served asset changes. New accounts, fact sheets, and SQL views do not provide versioned serving behind a stable URL.

Question 18watsonx platform fundamentals and architecture

Which pairing of watsonx component and primary function is correct?

  1. watsonx.data and serving foundation model prompts
  2. watsonx.ai and querying open table formats with SQL
  3. watsonx Orchestrate and tuning foundation model weights
  4. watsonx.governance and tracking model risk throughout the lifecycle Correct answer.

Explanation

watsonx.governance tracks model risk, compliance, and lifecycle documentation. The other pairings swap functions that belong to watsonx.ai, watsonx.data, and Orchestrate respectively.

Question 19watsonx platform fundamentals and architecture

What does the watsonx.ai foundation model catalog provide to users?

  1. A directory of Presto connectors for external databases
  2. A curated list of available models with their capabilities and supported tasks Correct answer.
  3. A list of orchestration skills available to assistants
  4. A registry of governance approvals for each deployment

Explanation

The model catalog presents the foundation models available in the platform along with their capabilities and supported tasks so users can choose appropriately. Governance approvals, database connectors, and orchestration skills are cataloged elsewhere.

Question 20watsonx platform fundamentals and architecture

A solution must ground model responses in the customer's own documents while keeping the base model unchanged. Which watsonx.ai pattern satisfies this requirement?

  1. Retrieval-augmented generation that injects retrieved passages into the prompt Correct answer.
  2. Converting documents into governance fact sheets
  3. Direct weight editing inside a deployment space
  4. Full fine-tuning that overwrites the base model weights

Explanation

RAG grounds responses by retrieving relevant passages and placing them in the prompt context, leaving the base model untouched. Fine-tuning changes weights, weight editing is unsupported, and fact sheets are documentation rather than grounding context.

Question 21watsonx platform fundamentals and architecture

Which consideration most directly affects the choice between a smaller and a larger foundation model in watsonx.ai for a high-volume application?

  1. The number of governance fact sheets attached to the model
  2. The number of skills registered in Orchestrate
  3. The storage format used by the watsonx.data catalog
  4. The trade-off between latency and cost per call versus response quality Correct answer.

Explanation

Model size affects latency, token cost, and output quality, so high-volume applications often balance these factors when selecting a model. Governance artifacts, data catalog formats, and orchestration skills do not drive that selection.

Question 22watsonx platform fundamentals and architecture

What is the function of an watsonx.ai deployment's online scoring endpoint?

  1. It renders the governance dashboard for the model
  2. It accepts requests at runtime and returns model predictions or generated text Correct answer.
  3. It stores the training dataset used to tune the model
  4. It schedules nightly batch jobs for data ingestion

Explanation

An online scoring endpoint receives runtime requests and returns model outputs such as predictions or generated text. Datasets, dashboards, and ingestion schedules are handled by other services.

Question 23watsonx platform fundamentals and architecture

A team wants prompts and model configurations to be reusable across several applications. Which watsonx.ai construct supports this most directly?

  1. Prompt templates stored as project assets and referenced by deployments Correct answer.
  2. Presto views created for each application
  3. Governance fact sheets copied into each application
  4. Hard-coded prompt strings embedded in each application binary

Explanation

Prompt templates saved as project assets can be versioned and reused across applications and deployments. Hard-coded strings, fact sheets, and SQL views do not provide reusable, managed prompt configuration.

Question 24watsonx platform fundamentals and architecture

Which statement about the watsonx platform's use of open-source foundation models is accurate?

  1. Open-source models bypass governance because their weights are public
  2. Only proprietary IBM models can ever appear in the catalog
  3. Open-source models cannot be tuned inside watsonx.ai
  4. The catalog can include open-source models alongside IBM-provided models, subject to licensing terms Correct answer.

Explanation

The catalog can offer both open-source and IBM-provided foundation models, and users must respect the applicable license terms. Open-source models can still be tuned, and governance obligations continue to apply regardless of weight availability.

Question 25Foundation models and prompt engineering

In watsonx, which parameter of a text generation request controls the randomness of token selection, where a value of 0 makes the output nearly deterministic?

  1. max_new_tokens
  2. decoding_method
  3. stop_sequences
  4. temperature Correct answer.

Explanation

Temperature scales the logits before sampling, so a low value sharpens the distribution toward the most likely tokens and 0 approximates greedy decoding. max_new_tokens only bounds output length, stop_sequences only truncate output, and decoding_method selects greedy versus sampling but does not itself set randomness.

Question 26Foundation models and prompt engineering

A prompt instructs a foundation model to classify support tickets into exactly three categories but the model sometimes returns free-form commentary. Which prompt engineering change most directly constrains the output format?

  1. Add the phrase "think step by step" before the categories
  2. Add a few labeled input-output examples showing only the category label as output Correct answer.
  3. Increase the temperature to encourage variety
  4. Remove the category definitions to shorten the prompt

Explanation

Few-shot examples demonstrate the exact input-to-output mapping, so the model imitates the terse label-only response pattern. Raising temperature increases deviation from the pattern, and removing definitions reduces the information the model needs to choose a category correctly.

Question 27Foundation models and prompt engineering

Which statement best describes the role of a foundation model's context window in prompt design?

  1. It is the maximum number of tokens the model can attend to across input and generated output Correct answer.
  2. It is the number of parameters stored in the model's largest transformer layer
  3. It is the number of concurrent API requests a watsonx deployment can serve
  4. It is the fixed number of few-shot examples the model requires to learn a task

Explanation

The context window bounds the total token budget the model can condition on, covering the prompt plus the completion. Exceeding it forces truncation or rejection, so prompt designers must budget tokens for instructions, examples, retrieved context, and output.

Question 28Foundation models and prompt engineering

You are building a retrieval-augmented prompt where retrieved passages sometimes contradict the user's question. Which instruction placement and wording best reduces hallucinated answers?

  1. Place the passages first and add a closing instruction to answer only from the provided context and say when it is insufficient Correct answer.
  2. Set temperature to 1.0 so the model explores alternative answers
  3. Place the instruction in the middle of the passages to save tokens
  4. Rely on the model's pretrained knowledge and omit the passages entirely

Explanation

Instructions placed after long context are followed more reliably, and explicitly authorizing an "insufficient information" response prevents the model from inventing facts. Grounding the answer in supplied passages is the core control for retrieval-augmented generation.

Question 29Foundation models and prompt engineering

What is the primary purpose of a system prompt in a chat-style foundation model request?

  1. To set persistent behavioral guidance and persona that conditions all subsequent turns Correct answer.
  2. To store the conversation history so the model can recall earlier messages
  3. To override the model's tokenizer and vocabulary at inference time
  4. To specify the billable rate for each generated token

Explanation

A system prompt establishes durable instructions such as role, tone, and constraints that apply across the dialogue. Conversation history is carried in the message list, and tokenization or billing are set by the platform, not by prompt text.

Question 30Foundation models and prompt engineering

A team observes that increasing temperature from 0.2 to 0.9 makes outputs more varied but also more factually inconsistent. Which explanation is most accurate?

  1. Higher temperature reduces the number of output tokens, truncating facts
  2. Higher temperature retrains the model on recent data, introducing errors
  3. Higher temperature flattens the probability distribution so lower-probability tokens are sampled more often Correct answer.
  4. Higher temperature increases the model's context window and admits noisy tokens

Explanation

Sampling with temperature divides logits by the temperature value, so larger values flatten the distribution and give unlikely tokens more chance of selection. This diversity is useful for creative tasks but harmful when factual precision matters.

Question 31Foundation models and prompt engineering

Which technique asks a model to produce intermediate reasoning steps before a final answer and typically improves performance on multi-step arithmetic word problems?

  1. Chain-of-thought prompting Correct answer.
  2. Zero-shot classification
  3. Top-k sampling
  4. Prompt templating with variables

Explanation

Chain-of-thought prompting elicits intermediate reasoning that guides the model toward the final result on tasks requiring multiple steps. It is most effective when examples or explicit instructions show the reasoning format.

Question 32Foundation models and prompt engineering

When writing a prompt template for watsonx Prompt Lab that will be reused across many inputs, what is the recommended way to inject per-request data?

  1. Append all possible inputs into the system prompt
  2. Use input variables or placeholders that are filled at runtime Correct answer.
  3. Encode the data as stop sequences
  4. Hard-code each customer's data into a separate saved prompt

Explanation

Prompt templates use variables such as {input} that are substituted at invocation, enabling one tested prompt to serve many requests. Hard-coding values defeats reuse and multiplies maintenance, while stop sequences only terminate generation.

Question 33Foundation models and prompt engineering

A model given a long document answers a question about a detail near the beginning incorrectly, though the detail is present. Which prompt adjustment is most likely to help?

  1. Increase max_new_tokens substantially
  2. Remove the document and rely on the model's memory
  3. Restate the question and key instructions after the document so they remain salient Correct answer.
  4. Increase temperature to force more exploration

Explanation

Models can lose track of instructions buried before very long context, a position-sensitivity effect. Repeating the question and constraints after the document keeps the task salient at generation time.

Question 34Foundation models and prompt engineering

What does the top_p (nucleus) sampling parameter control?

  1. The smallest set of tokens whose cumulative probability reaches p, from which the next token is sampled Correct answer.
  2. The penalty applied to tokens that already appeared in the output
  3. The maximum number of tokens the model may generate
  4. The exact number of highest-probability tokens retained before sampling

Explanation

Nucleus sampling truncates the distribution to the minimal token set whose cumulative probability exceeds p, then samples from it. This adapts the candidate pool size to how concentrated the distribution is at each step.

Question 35Foundation models and prompt engineering

Which prompt is the strongest example of role-based prompting for a summarization task in watsonx?

  1. "{report}"
  2. "You are a financial analyst. Summarize the following earnings report in three bullet points for an executive audience: {report}" Correct answer.
  3. "Please output the report verbatim without changes."
  4. "Summarize this."

Explanation

Role-based prompting assigns a persona, audience, format, and length, all of which constrain the model's output usefully. The other prompts either lack task specification or contradict the summarization goal.

Question 36Foundation models and prompt engineering

A developer wants deterministic, repeatable outputs for a regression test of a prompt. Which configuration best supports this?

  1. Set temperature to 1 and top_p to 1
  2. Set a high repetition penalty and high temperature
  3. Rely on random seed absence and average several runs
  4. Set temperature to 0 and use greedy decoding Correct answer.

Explanation

Greedy decoding with temperature 0 selects the highest-probability token at each step, removing sampling randomness. Even then, floating-point and infrastructure differences can cause minor variation, so tests should tolerate small drift.

Question 37Foundation models and prompt engineering

Which statement about zero-shot versus few-shot prompting is correct?

  1. Zero-shot prompting always outperforms few-shot prompting on classification tasks
  2. Few-shot prompting supplies labeled examples in the prompt, while zero-shot relies only on instructions Correct answer.
  3. Few-shot prompting requires fine-tuning the model's weights on the examples
  4. Zero-shot prompting cannot be used with instruction-tuned models

Explanation

Few-shot prompting conditions the model on in-context examples without changing weights, whereas zero-shot relies on instructions alone. Examples often improve format adherence and accuracy, though they consume context tokens.

Question 38Foundation models and prompt engineering

A prompt asks a model to extract dates from contracts, but it returns dates in inconsistent formats. Which approach most reliably standardizes the output?

  1. Ask the model to be careful about formatting
  2. Add more narrative context about the contract
  3. Specify the exact output format with an example, such as YYYY-MM-DD, and request JSON Correct answer.
  4. Lower max_new_tokens so only short answers fit

Explanation

Explicit format specifications plus a concrete example anchor the model to a canonical representation, and requesting JSON makes the structure machine-checkable. Vague exhortations and length limits do not define a format.

Question 39Foundation models and prompt engineering

What is the main risk of including a very large number of few-shot examples in a single prompt?

  1. Preventing the use of a system prompt
  2. Consuming much of the context window and increasing latency and cost per request Correct answer.
  3. Disabling the model's tokenizer
  4. Permanently altering the model's weights

Explanation

Every example occupies input tokens, so many examples crowd out room for the actual task and raise cost and latency. Selecting a small, diverse set of high-quality examples is usually better than maximizing count.

Question 40Foundation models and prompt engineering

Which prompt engineering practice best mitigates prompt injection when a model processes untrusted user-supplied text?

  1. Concatenate untrusted text directly into the system prompt
  2. Clearly delimit untrusted content and instruct the model to treat it as data, not instructions Correct answer.
  3. Increase temperature so the model ignores malicious instructions
  4. Rely solely on the model's safety training with no delimiting

Explanation

Delimiters and explicit instructions to treat embedded text as data reduce the chance that content overrides the developer's instructions. Defense in depth also includes output validation, since prompting alone cannot fully prevent injection.

Question 41Foundation models and prompt engineering

In watsonx text generation, what does a stop sequence do?

  1. It prevents the model from generating any punctuation
  2. It sets the maximum number of tokens per request
  3. It halts generation when the specified string is produced and the string is typically excluded from output Correct answer.
  4. It deletes the string from the prompt before sending

Explanation

Stop sequences act as terminators: when the model emits the string, decoding ends, which is useful for cutting off at delimiters like a newline or a role tag. Length limits are controlled separately by token-count parameters.

Question 42Foundation models and prompt engineering

A summarization prompt produces fluent summaries that omit key figures present in the source. Which change is most appropriate?

  1. Raise temperature to make the summary more creative
  2. Ask the model to translate the document before summarizing
  3. Instruct the model to preserve all numeric values and include a checklist item requiring figures Correct answer.
  4. Shorten the prompt by removing the source document

Explanation

Explicitly requiring preservation of numeric facts and giving a checklist directs attention to details that abstractive summarization tends to drop. Faithfulness constraints are the standard remedy for omission.

Question 43Foundation models and prompt engineering

Which statement accurately distinguishes fine-tuning from prompt engineering on a foundation model?

  1. Both methods update weights but differ only in learning rate
  2. Fine-tuning updates model weights using labeled data, while prompt engineering changes only the input without weight updates Correct answer.
  3. Fine-tuning changes only the prompt template, while prompt engineering updates weights
  4. Prompt engineering permanently stores new knowledge in the model's parameters

Explanation

Fine-tuning performs additional training so parameters adapt to a task, whereas prompt engineering steers a frozen model through instructions and examples. Prompting is faster and cheaper to iterate but does not persist knowledge in weights.

Question 44Foundation models and prompt engineering

What is the effect of the repetition penalty parameter during generation?

  1. It forces the model to repeat key phrases for emphasis
  2. It reduces the likelihood of tokens that have already appeared in the generated text Correct answer.
  3. It caps the number of distinct tokens in the vocabulary
  4. It increases the context window size

Explanation

A repetition penalty down-weights previously generated tokens to discourage loops and verbatim repetition. Setting it too high can harm tasks where legitimate repetition is required, such as code or structured lists.

Question 45Foundation models and prompt engineering

A team wants a model to answer questions strictly in the style of their brand voice across many sessions. Which design is most appropriate?

  1. Rely on the model inferring brand voice from the question alone
  2. Set temperature to maximum so the voice varies naturally
  3. Encode brand voice rules and examples in a reusable system prompt applied to every request Correct answer.
  4. Include the brand rules only in the first user message of each session

Explanation

A reusable system prompt applies consistent behavioral guidance to every request in a session and across sessions when reused. Brand voice is a durable instruction rather than a one-off user turn.

Question 46Foundation models and prompt engineering

Which of these is the best example of an instruction that reduces ambiguity in a generation prompt?

  1. "Do your best with the content"
  2. "Be good and thorough"
  3. "Return exactly five bullet points, each under 15 words, covering only risks mentioned in the text" Correct answer.
  4. "Write something useful about the text"

Explanation

Effective instructions specify countable constraints such as count, length, and scope, which are verifiable and reduce interpretation variance. Vague exhortations give the model no measurable target.

Question 47Foundation models and prompt engineering

When a chain-of-thought prompt is used with a reasoning-capable model in production, what is a common way to avoid exposing intermediate reasoning to end users?

  1. Remove all instructions from the prompt
  2. Set top_k to 1 so reasoning is skipped
  3. Disable the system prompt entirely
  4. Request reasoning and final answer separately, then return only the final-answer field Correct answer.

Explanation

Separating the reasoning trace from the final answer lets applications log or discard the reasoning while presenting only the conclusion. This also supports output validation on a structured final field.

Question 48Foundation models and prompt engineering

Which scenario is the best fit for using a higher temperature setting?

  1. Normalizing addresses to a canonical postal format
  2. Extracting invoice totals into a fixed accounting schema
  3. Classifying a support ticket into one of four predefined queues
  4. Brainstorming marketing slogans where diverse creative options are desired Correct answer.

Explanation

Creative ideation benefits from sampling diversity, so a moderate-to-high temperature is appropriate. Extraction, classification, and normalization require consistency and accuracy, favoring low temperature.

Question 49Foundation models and prompt engineering

A prompt includes the instruction "If the answer is not in the provided context, respond with 'Not found'." What failure mode is this primarily designed to address?

  1. Hallucination of unsupported answers Correct answer.
  2. Loss of the system prompt
  3. Excessive output length
  4. Slow tokenization of the input

Explanation

Giving the model an explicit abstention path reduces the tendency to fabricate answers when grounding evidence is missing. It is a standard guardrail in retrieval-augmented generation pipelines.

Question 50Foundation models and prompt engineering

In a watsonx text generation request, which parameter pair most directly governs how much the next-token distribution is narrowed before sampling?

  1. min_new_tokens and truncate_input_tokens
  2. stop sequences and max_new_tokens
  3. temperature and top_p Correct answer.
  4. decoding_method and repetition_penalty

Explanation

Temperature rescales the logits and top_p truncates the cumulative probability mass that remains eligible for sampling, so together they control how sharply the distribution is narrowed. Length controls such as max_new_tokens and stop sequences affect when generation ends, not how the next token is chosen.