I85PREP
⌘ K
C1000-200 study materials
C1000-200 study resources

C1000-200 practice questions

One question at a time. Build your C1000-200 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 1Unverified - official content domains not published on the cited page

In IBM watsonx.ai, which component is used to store and manage prompt templates, model configurations, and deployment assets in a governed, versioned way?

  1. The deployment runtime log stream
  2. The Prompt Lab session history
  3. The watsonx.ai project asset repository Correct answer.
  4. The notebook kernel state

Explanation

Watsonx.ai projects act as the governed workspace where prompts, model configurations, notebooks, and deployments are saved as versioned assets. Session history and kernel state are ephemeral and are not governed assets. Runtime logs record execution output rather than asset definitions.

Question 2Unverified - official content domains not published on the cited page

A team needs a foundation model that can both summarize long documents and generate embeddings for a retrieval pipeline. Which two model families in watsonx.ai are typically paired for this pattern?

  1. A translation model and a tabular regression model
  2. A text generation model and an embedding model Correct answer.
  3. A time-series model and a classification model
  4. A speech-to-text model and a vision model

Explanation

Retrieval-augmented generation pairs an embedding model, which vectorizes documents for similarity search, with a text generation model that produces the summary from retrieved context. The other pairings do not address vector search plus natural-language summarization. Speech, vision, time-series, and tabular models serve different modalities.

Question 3Unverified - official content domains not published on the cited page

What is the primary purpose of the prompt template in a watsonx.ai Prompt Lab experiment?

  1. To compile the model into a container image
  2. To encrypt the model weights at rest
  3. To define the structured input text with placeholders that are filled at run time Correct answer.
  4. To set the billing plan for the deployment

Explanation

A prompt template contains the instruction and example text plus variable placeholders that are substituted with real inputs each time the prompt is run. It does not perform containerization, billing configuration, or encryption. Those tasks belong to deployment, account management, and platform security respectively.

Question 4Unverified - official content domains not published on the cited page

Which parameter in watsonx.ai text generation controls the randomness of token selection, where a value of 0 makes output nearly deterministic?

  1. Max tokens
  2. Stop sequences
  3. Repetition penalty
  4. Temperature Correct answer.

Explanation

Temperature scales the probability distribution over the next token; at 0 the model effectively always picks the most likely token, giving near-deterministic output. Max tokens limits length, stop sequences end generation, and repetition penalty discourages repeated tokens. None of those three directly controls sampling randomness.

Question 5Unverified - official content domains not published on the cited page

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

  1. The Prompt Lab compare mode Correct answer.
  2. The model risk report
  3. The notebook debugger
  4. The data refinery

Explanation

Prompt Lab provides a comparison view where the same prompt is sent to multiple models so outputs can be evaluated side by side. The data refinery is for data preparation, the model risk report supports governance documentation, and the notebook debugger helps troubleshoot code. Only the compare mode addresses model selection by output.

Question 6Unverified - official content domains not published on the cited page

A developer wants to call a deployed watsonx.ai foundation model from an external application. Which mechanism is the standard way to authenticate those REST API calls?

  1. A shared notebook password
  2. A database connection string
  3. An IBM Cloud API key exchanged for an IAM bearer token Correct answer.
  4. A public anonymous access flag

Explanation

Watsonx.ai REST endpoints are protected by IBM Cloud Identity and Access Management, so callers exchange an API key for a short-lived bearer token and pass it in the Authorization header. Notebook passwords, anonymous access, and database connection strings are not valid IAM credentials. This keeps programmatic access auditable and scoped by policy.

Question 7Unverified - official content domains not published on the cited page

In watsonx.ai governance, what does a model card primarily document?

  1. The source code of the inference container
  2. The monthly cost of inference calls
  3. The network topology of the deployment cluster
  4. The intended use, training data details, and known limitations of a model Correct answer.

Explanation

A model card is a governance artifact describing a model's intended purpose, training data characteristics, evaluation results, and known limitations. It does not report billing, network topology, or container source code. Model cards help reviewers judge whether a model is appropriate for a given use case.

Question 8Unverified - official content domains not published on the cited page

Which watsonx.ai feature enables a foundation model to answer questions over an organization's private documents without retraining the model?

  1. Increasing the max tokens parameter
  2. Switching the deployment to a larger hardware profile
  3. Retrieval-augmented generation with a vector index Correct answer.
  4. Full fine-tuning of all model weights

Explanation

Retrieval-augmented generation stores private documents as embeddings in a vector index, retrieves relevant passages at query time, and inserts them into the prompt. This grounds answers in private content without changing model weights. Fine-tuning, token limits, and hardware profiles do not supply document knowledge at query time.

Question 9Unverified - official content domains not published on the cited page

A prompt produces fluent but off-topic answers. Which change is most likely to improve relevance while keeping the model unchanged?

  1. Remove all stop sequences
  2. Raise temperature to the maximum
  3. Add explicit task instructions and few-shot examples to the prompt Correct answer.
  4. Reduce max tokens to a very small number

Explanation

Clear instructions plus a few worked examples steer the model toward the intended task format and content. Truncating output, maximizing randomness, or removing stop sequences does not improve topical focus. Prompt refinement is the lowest-cost intervention before considering fine-tuning.

Question 10Unverified - official content domains not published on the cited page

An application needs consistent JSON output from a watsonx.ai text generation model. Which combination of techniques is most appropriate?

  1. Increase max tokens and disable stop sequences
  2. Set temperature to 1.5 and rely on free-form prose
  3. Provide a JSON schema in the prompt, use few-shot examples, and lower temperature Correct answer.
  4. Use an embedding model instead of a generation model

Explanation

Specifying the target schema, showing valid examples, and lowering temperature all reduce formatting variance so the model reliably emits parseable JSON. High temperature and free-form prose increase structural drift. Embedding models return vectors, not JSON documents.

Question 11Unverified - official content domains not published on the cited page

A team deploys a watsonx.ai model and observes high latency under load. Which action best addresses throughput without changing the model?

  1. Disable IAM authentication
  2. Shorten the prompt template by one word
  3. Increase the deployment's serving capacity or replica count Correct answer.
  4. Raise the temperature parameter

Explanation

Throughput under concurrent load is governed by serving capacity, so scaling replicas or the hardware profile distributes requests across more resources. Temperature, trivial prompt edits, and authentication changes do not add serving capacity. Disabling authentication would also be a security violation.

Question 12Unverified - official content domains not published on the cited page

When preparing training data for a tuning job in watsonx.ai, why must records be split into training and validation sets?

  1. To measure generalization on unseen examples and detect overfitting Correct answer.
  2. To reduce the storage cost of the data asset
  3. To convert the data into embeddings
  4. To satisfy the model's token limit

Explanation

A held-out validation split estimates how the tuned model performs on data it has not seen, revealing overfitting to the training examples. Storage cost, token limits, and embedding conversion are unrelated to the purpose of splitting. Without a validation split, quality gains cannot be trusted.

Question 13Unverified - official content domains not published on the cited page

A prompt includes several long few-shot examples but the model's answer is cut off mid-sentence. Which parameter should be adjusted first?

  1. Set temperature to zero
  2. Increase stop sequences
  3. Increase max tokens so the full response can be generated Correct answer.
  4. Decrease max tokens to force brevity

Explanation

Truncated output indicates the generation hit the max tokens ceiling, so raising that limit allows the response to complete. Lowering it worsens truncation, stop sequences end output earlier, and temperature does not control length. Note that prompt and completion tokens share the model's context window.

Question 14Unverified - official content domains not published on the cited page

Which statement best describes the difference between prompt tuning and full fine-tuning in watsonx.ai?

  1. Prompt tuning updates all model weights, while fine-tuning updates none
  2. Prompt tuning requires no training data at all
  3. Both methods leave the base model completely unchanged
  4. Prompt tuning learns a small set of soft prompt parameters while the base model stays frozen; fine-tuning updates model weights Correct answer.

Explanation

Prompt tuning optimizes a compact set of learned prompt vectors against a frozen base model, making it cheaper to store and switch. Full fine-tuning adjusts the model's weights and produces a larger, separate model artifact. Both approaches still require task-specific training examples.

Question 15Unverified - official content domains not published on the cited page

A retrieval pipeline returns irrelevant passages, causing poor answers. Which step should be tuned first to improve retrieval quality?

  1. The max tokens of the generation model
  2. The embedding model choice and the chunking strategy for documents Correct answer.
  3. The temperature of the generation model
  4. The deployment replica count

Explanation

Retrieval quality depends primarily on how documents are chunked and how well the embedding model represents their meaning. Generation parameters and replica counts operate after retrieval and cannot fix irrelevant context. Improving chunk size and embedding choice directly raises the relevance of retrieved passages.

Question 16Unverified - official content domains not published on the cited page

In watsonx.governance, which artifact tracks a model's evaluation metrics, training details, and approvals across its lifecycle?

  1. The API gateway access log
  2. The notebook output cell
  3. The model inventory entry with its factsheet Correct answer.
  4. The object storage bucket policy

Explanation

The model inventory records each model and its factsheet, which captures evaluation metrics, training details, and approval status for audit. Notebook outputs, gateway logs, and bucket policies are operational artifacts and do not provide lifecycle governance records. The inventory is the system of record for model risk review.

Question 17Unverified - official content domains not published on the cited page

A team wants reproducible experiments when calling a watsonx.ai generation model. Which setting most directly supports deterministic output?

  1. Adding many stop sequences
  2. Setting temperature to 1 and increasing max tokens
  3. Using a larger hardware profile
  4. Setting temperature to 0 and fixing the random seed where supported Correct answer.

Explanation

Deterministic decoding requires minimizing sampling randomness, so temperature near 0 combined with a fixed seed where the API supports it gives repeatable results. High temperature, stop sequences, and hardware size do not remove sampling variance. Reproducibility also requires storing the exact prompt and parameters as a project asset.

Question 18Unverified - official content domains not published on the cited page

Which practice best mitigates prompt injection risk in a watsonx.ai application that processes user-supplied text?

  1. Increase the model's temperature to make attacks less predictable
  2. Remove all system instructions so the model is neutral
  3. Allow the model to call any tool without restriction
  4. Treat retrieved and user content as untrusted, isolate it from instructions, and validate outputs Correct answer.

Explanation

Prompt injection succeeds when untrusted content is interpreted as instructions, so separating data from instructions, delimiting untrusted text, and validating outputs reduces risk. Raising temperature adds noise but not security, removing instructions weakens control, and unrestricted tool access expands the attack surface. Defense also includes least-privilege tool permissions.

Question 19Unverified - official content domains not published on the cited page

What is the effect of the top-p (nucleus) sampling parameter in watsonx.ai text generation?

  1. It limits sampling to the smallest set of tokens whose cumulative probability reaches p Correct answer.
  2. It penalizes tokens that already appeared in the text
  3. It sets the maximum number of output tokens to p
  4. It defines the number of few-shot examples in the prompt

Explanation

Top-p sampling restricts the candidate pool to the most probable tokens whose combined probability reaches the threshold p, trimming the long tail. It is distinct from max tokens, repetition penalty, and prompt example count. Lowering p generally makes output more focused and less diverse.

Question 20Unverified - official content domains not published on the cited page

A regulated team must explain why a foundation model produced a particular recommendation. Which combination of watsonx capabilities best supports this?

  1. Deploying more replicas and enabling autoscaling
  2. Compressing the model weights for faster loading
  3. Increasing temperature and max tokens
  4. Prompt logging, retrieval source citations, and governance factsheet documentation Correct answer.

Explanation

Explainability requires capturing the exact prompt and parameters, citing the retrieved sources that grounded the answer, and documenting model facts in governance artifacts. Scaling, temperature changes, and weight compression affect performance but not traceability. Together these records let reviewers reconstruct how an answer was produced.

Question 21Unverified - official content domains not published on the cited page

Which scenario is the strongest candidate for fine-tuning rather than prompt engineering in watsonx.ai?

  1. Thousands of examples requiring a consistent domain-specific style and classification behavior Correct answer.
  2. A one-off summary of a single short email
  3. A quick prototype to test whether a model can answer a question
  4. A task where the model must cite private documents at query time

Explanation

Fine-tuning pays off when many labeled examples encode a stable style or decision boundary that prompting struggles to enforce. One-off tasks and quick prototypes are cheaper with prompting, and citing private documents is a retrieval problem. Fine-tuning also requires enough high-quality examples to avoid overfitting.

Question 22Unverified - official content domains not published on the cited page

A watsonx.ai deployment must serve bursty traffic while controlling cost. Which design is most appropriate?

  1. Run a single replica and queue all requests
  2. Use autoscaling with a sensible minimum and maximum replica range Correct answer.
  3. Disable monitoring to reduce overhead
  4. Provision permanently for peak load with maximum replicas

Explanation

Autoscaling adds replicas during demand spikes and scales back when traffic falls, balancing latency against cost. Permanent peak provisioning wastes money, a single replica creates a bottleneck and outage risk, and disabling monitoring removes the signals needed to tune scaling. A bounded replica range guards against runaway spend.

Question 23Unverified - official content domains not published on the cited page

An evaluation shows a tuned model scores highly on its training set but poorly on new data. Which conclusion is best supported?

  1. The model is overfitting and needs more diverse data or regularization Correct answer.
  2. The deployment needs more replicas
  3. The prompt template should be shortened
  4. The model needs a higher inference temperature

Explanation

A large gap between training and held-out performance is the classic signature of overfitting, where the model memorizes training examples instead of generalizing. More diverse data, stronger regularization, or fewer epochs address this. Inference temperature, replica count, and prompt length do not change learned weights.

Question 24Unverified - official content domains not published on the cited page

A team must choose between a large model with strong reasoning and a smaller model with lower latency for a high-volume classification task. Which evaluation approach is most defensible?

  1. Pick the largest model because size always correlates with accuracy
  2. Pick the smaller model because it is cheaper regardless of accuracy
  3. Deploy both without measurement and let users decide
  4. Benchmark both on a representative labeled sample for accuracy and latency, then weigh cost against quality thresholds Correct answer.

Explanation

Model selection should be evidence-based: measure task accuracy on a representative labeled set alongside latency, then choose the model that meets quality thresholds at acceptable cost. Assuming size or cost alone decides ignores task-specific performance. Undocumented deployment also prevents accountability and later tuning.

Question 25Unverified - official content domains not published on the cited page

A RAG application answers correctly when documents are short but fails when relevant text spans several long documents. Which diagnosis is most plausible?

  1. The deployment has too many replicas
  2. The embedding chunks are too small or too few, so cross-document context is fragmented or missing from the prompt Correct answer.
  3. The model's max tokens setting is irrelevant to any failure
  4. The temperature is too low for the model to reason

Explanation

When answers require synthesizing evidence across long documents, small chunks and a limited top-k can omit necessary context from the prompt. Adjusting chunk overlap, chunk size, and retrieval depth addresses this. Low temperature aids consistency, replica count affects capacity, and max tokens genuinely constrains available context.

Question 26Unverified - official content domains not published on the cited page

In watsonx.ai, which object must typically exist before you can start a prompt-tuning experiment from the Prompt Lab?

  1. A deployed Python function
  2. A relational database connection
  3. A Git repository connected to the project
  4. A foundation model available in the selected project or space Correct answer.

Explanation

Prompt tuning and other tuning experiments in watsonx.ai operate against a foundation model that the project or deployment space can access, so the model must be available first. Supporting assets such as storage or Git are useful but are not the prerequisite for launching a tuning experiment.

Question 27Unverified - official content domains not published on the cited page

A watsonx.ai user wants the same prompt to produce varied creative marketing copy across several runs. Which decoding setting should be raised?

  1. Max tokens
  2. Repetition penalty removal
  3. Stop sequences
  4. Temperature Correct answer.

Explanation

Temperature controls the randomness of token selection; raising it makes sampling less deterministic and increases variation between runs. Max tokens only bounds output length, and stop sequences only terminate generation at chosen strings.

Question 28Unverified - official content domains not published on the cited page

Which watsonx.ai capability is designed to let a foundation model use external tools such as a weather API during a conversation?

  1. Agent tool calling Correct answer.
  2. Prompt templates
  3. Model cards
  4. Tokenizer configuration

Explanation

Agent-style tool calling lets the model decide when to invoke an external function or API and incorporate the returned result into its response. Prompt templates and tokenizer settings shape text but do not execute external tools.

Question 29Unverified - official content domains not published on the cited page

A team wants to compare two prompt variants fairly across many test inputs in watsonx.ai. Which practice best supports this?

  1. Shorten only the second variant to save tokens
  2. Evaluate each variant on a different dataset
  3. Change one variable at a time and evaluate both variants on the same fixed input set Correct answer.
  4. Increase temperature for both variants between runs

Explanation

A controlled comparison requires holding the inputs and decoding settings constant while changing only the prompt under test. Using different datasets or varying temperature confounds the results and makes conclusions unreliable.

Question 30Unverified - official content domains not published on the cited page

In watsonx.ai, what is the primary role of a project?

  1. A billing account limited to one model
  2. A network gateway for API traffic
  3. A collaboration workspace that holds assets such as models, prompts, and notebooks Correct answer.
  4. A runtime container that replaces deployment spaces

Explanation

Projects in watsonx.ai are collaborative workspaces where teams organize data, notebooks, prompts, and model assets. Deployment spaces serve a different purpose, handling serving and inference endpoints.

Question 31Unverified - official content domains not published on the cited page

Which task is a foundation model in watsonx.ai most directly suited to perform without additional training?

  1. Guaranteeing regulatory compliance of a loan decision
  2. Replacing a transactional database engine
  3. Providing deterministic arithmetic for financial ledgers
  4. Summarizing a supplied support transcript into key points Correct answer.

Explanation

Foundation models are strong at natural-language transformation tasks such as summarization when given the source text in the prompt. They cannot guarantee compliance outcomes, replace database engines, or provide exact deterministic arithmetic.

Question 32Unverified - official content domains not published on the cited page

A developer needs to send prompts to a deployed watsonx.ai model from a script. Which credential approach is appropriate?

  1. Authenticate with an IBM Cloud API key to obtain a bearer token for the request Correct answer.
  2. Publish the API key in a public repository
  3. Disable authentication on the endpoint
  4. Embed a personal password directly in the script

Explanation

Watsonx.ai REST calls are authenticated by exchanging an IBM Cloud API key for a bearer token, which is then sent with the request. Hard-coding passwords or exposing keys creates security failures and is never appropriate.

Question 33Unverified - official content domains not published on the cited page

Which watsonx.ai asset stores a reusable prompt with variables that can be filled at runtime?

  1. Service instance
  2. Model card
  3. Deployment space
  4. Prompt template Correct answer.

Explanation

A prompt template holds a reusable prompt structure with placeholders that are populated with input values at invocation time. Deployment spaces, model cards, and service instances serve unrelated purposes.

Question 34Unverified - official content domains not published on the cited page

A team wants a watsonx.ai model to reliably extract dates from invoices in a fixed schema. Which approach is most appropriate?

  1. Provide a clear instruction plus a schema example and constrain output format Correct answer.
  2. Raise temperature to its maximum value
  3. Remove all formatting instructions from the prompt
  4. Rely on the model to choose any output layout

Explanation

Structured extraction improves when the prompt states the task, shows the expected schema, and constrains the output format. High temperature and absent formatting guidance increase variability and make parsing unreliable.

Question 35Unverified - official content domains not published on the cited page

When a watsonx.ai tuning job is configured, which parameter most directly controls how many times the model sees the full training dataset?

  1. Max output tokens
  2. Learning rate warmup
  3. Batch size
  4. Number of epochs Correct answer.

Explanation

Epochs define how many full passes the training process makes over the dataset. Batch size controls samples per step and max output tokens affects inference length, not training passes.

Question 36Unverified - official content domains not published on the cited page

A RAG pipeline in watsonx.ai returns chunks, but answers miss information split across chunk boundaries. Which change best addresses this?

  1. Reduce the number of retrieved chunks to one
  2. Shorten the user question before retrieval
  3. Lower the model temperature to zero
  4. Add overlapping windows between adjacent chunks during ingestion Correct answer.

Explanation

Overlapping chunk windows preserve context that would otherwise be split across boundaries, helping retrieval capture complete ideas. Retrieving fewer chunks or changing temperature does not recover text lost at boundaries.

Question 37Unverified - official content domains not published on the cited page

Which watsonx.governance practice best supports auditing who approved a model for production?

  1. Store the approval decision only in a private chat message
  2. Record approval decisions and reviewers in the model's lifecycle record Correct answer.
  3. Delete evaluation results after approval
  4. Rely on the developer's memory of the review

Explanation

Auditability requires that approvals, reviewers, and supporting evaluations be captured in the governed lifecycle record. Informal channels or deleted evidence cannot support a later audit.

Question 38Unverified - official content domains not published on the cited page

A team notices a watsonx.ai model answers questions about recent events incorrectly. Which explanation is most consistent with how foundation models work?

  1. The deployment space is blocking factual queries
  2. The model's temperature is too low to recall facts
  3. The tokenizer has removed all dates from the prompt
  4. The model's knowledge is limited to its training data and lacks the recent information Correct answer.

Explanation

A foundation model's parametric knowledge is bounded by its training cutoff, so events after that point are unknown unless supplied through retrieval or tools. Temperature, tokenization, and deployment spaces do not explain missing recent facts.

Question 39Unverified - official content domains not published on the cited page

Which prompt engineering technique is most appropriate when a task requires the model to reason through multiple arithmetic steps before answering?

  1. Ask the model to show intermediate reasoning steps before the final answer Correct answer.
  2. Increase max tokens without changing the prompt
  3. Remove all examples from the prompt
  4. Ask only for the final number with no context

Explanation

Eliciting intermediate reasoning steps helps models handle multi-step problems more reliably than demanding only a final answer. Increasing length or removing examples does not supply the reasoning structure.

Question 40Unverified - official content domains not published on the cited page

A team must ensure a watsonx.ai application does not retain sensitive user text in prompts logged for debugging. Which control is most direct?

  1. Use a larger model for all requests
  2. Disable evaluation of the application
  3. Redact or mask sensitive fields before prompts are written to logs Correct answer.
  4. Increase the model temperature

Explanation

Redaction or masking at the logging boundary prevents sensitive content from persisting in debug records. Model size, temperature, and evaluation settings do not control what is written to logs.

Question 41Unverified - official content domains not published on the cited page

In watsonx.ai, which factor most directly determines the maximum context a prompt can include?

  1. The model's context window limit Correct answer.
  2. The number of deployment spaces in the account
  3. The number of users in the project
  4. The size of the project's Git repository

Explanation

Each foundation model has a context window that bounds the combined prompt and completion tokens it can process. Account, repository, and user counts do not determine this limit.

Question 42Unverified - official content domains not published on the cited page

A team wants to reduce hallucinations in a watsonx.ai question-answering assistant grounded in internal documents. Which combination is most effective?

  1. Ask the model to answer from memory without sources
  2. Retrieve relevant passages and instruct the model to answer only from the provided context Correct answer.
  3. Raise temperature and remove the context passages
  4. Shorten the instructions to a single word

Explanation

Grounding the answer in retrieved passages and instructing the model to rely only on that context reduces unsupported claims. Removing context or raising temperature increases the likelihood of fabricated answers.

Question 43Unverified - official content domains not published on the cited page

Which watsonx.ai deployment configuration change most directly reduces per-request latency for a latency-sensitive endpoint?

  1. Increase the max output tokens for every request
  2. Enable higher temperature sampling
  3. Add more few-shot examples to every prompt
  4. Select a smaller model variant with fewer parameters Correct answer.

Explanation

Smaller model variants generally require less computation per token and can lower response latency. Longer outputs, larger prompts, and higher temperature add work rather than reducing it.

Question 44Unverified - official content domains not published on the cited page

What distinguishes a watsonx.ai deployment space from a project?

  1. A deployment space hosts served assets and online endpoints, while a project is a collaboration workspace Correct answer.
  2. A deployment space stores only notebooks and no models
  3. A project serves production inference while a space is only for authoring
  4. A deployment space is a billing account for API keys

Explanation

Deployment spaces are where assets are promoted and served as online deployments, whereas projects are the collaborative authoring environment. The two roles are complementary and distinct.

Question 45Unverified - official content domains not published on the cited page

A team needs to evaluate whether a watsonx.ai model's answers are faithful to retrieved sources. Which metric is most relevant?

  1. Token count of the prompt
  2. Size of the training dataset file
  3. Number of deployment spaces used
  4. Answer faithfulness to the retrieved context Correct answer.

Explanation

Faithfulness measures whether generated claims are supported by the retrieved context, which is the core quality concern for grounded answering. Prompt length, space count, and file size do not measure grounding.

Question 46Unverified - official content domains not published on the cited page

A regulated team must document the intended use and known limitations of a foundation model they deploy. Which artifact is designed for this?

  1. A model card Correct answer.
  2. A deployment log
  3. A prompt template
  4. A tokenizer file

Explanation

Model cards summarize intended use, training context, and limitations to support transparency and governance. Deployment logs, prompt templates, and tokenizer files do not serve this documentation purpose.

Question 47Unverified - official content domains not published on the cited page

A team observes that a watsonx.ai assistant performs well on simple queries but fails when a question requires combining facts from three separate documents. Which architectural change is most appropriate?

  1. Use multi-step retrieval that gathers evidence across documents before generating the answer Correct answer.
  2. Increase temperature so the model explores more combinations
  3. Reduce retrieval to a single document to simplify the prompt
  4. Remove the instructions that define the task

Explanation

Questions requiring facts from multiple documents benefit from multi-step retrieval that collects and combines evidence before generation. Limiting retrieval or removing instructions reduces the information and guidance the model needs.

Question 48Unverified - official content domains not published on the cited page

A team wants to compare a tuned model against its base model before promotion. Which evaluation design gives the most defensible conclusion?

  1. Score each model with different temperature values
  2. Score only the tuned model and skip the baseline
  3. Score both models on the same held-out test set with identical prompts and decoding settings Correct answer.
  4. Score the tuned model on its training data and the base model on new data

Explanation

A defensible comparison holds inputs, prompts, and decoding settings constant and evaluates on held-out data that neither model was tuned on. Mismatched data or settings confound the comparison.

Question 49Unverified - official content domains not published on the cited page

An application must answer over a corpus that changes daily. Which design keeps answers current without retraining the foundation model?

  1. Raise the temperature whenever documents change
  2. Refresh the retrieval index as documents change and ground answers in retrieved passages Correct answer.
  3. Retune the model on the full corpus each day
  4. Increase max tokens to cover new documents

Explanation

Retrieval-augmented generation keeps knowledge current by updating the index, avoiding repeated expensive retraining. Retuning daily is costly and temperature or token limits do not add new knowledge.

Question 50Unverified - official content domains not published on the cited page

A team must decide whether a poor watsonx.ai result stems from retrieval or from generation. Which diagnostic step is most informative?

  1. Add more few-shot examples without reviewing retrieval
  2. Switch deployment spaces and retest
  3. Increase the model temperature and rerun the same prompt
  4. Inspect the retrieved passages to see whether they contain the needed evidence Correct answer.

Explanation

Checking the retrieved passages isolates whether the needed evidence reached the model, separating retrieval failures from generation failures. Changing temperature, adding examples, or moving spaces does not reveal where the pipeline broke.