100% Latest & Updated Microsoft AI-300 Practice Test Questions, Exam Dumps & Verified Answers!
30 Days Free Updates, Instant Download!
Question 90
An AI operations engineer must resolve 3 INDEPENDENT workstreams. Workstream A: A multinational data-science platform has shared training components but project workspaces cannot discover a governed enterprise-wide component version. Workstream B: An AML training sweep exhausts expensive GPU quota because no trial budget, early stopping or primary validation metric is configured. Workstream C: An HR assistant repeatedly ignores a requirement to cite policy sections because its prompt gives vague instructions and no output schema. Select exactly 3 actions, ONE directly solving EACH listed independent workstream.
- Design task-specific prompts with clear context, instructions, input boundaries, structured output criteria and failure handling, and test on realistic cases.
- Configure an AML hyperparameter sweep with defined search space, primary metric, resource budget and appropriate early-termination policy.
- Run appropriate Foundry evaluation metrics for groundedness, relevance, coherence and fluency using aligned inputs, responses and source context.
- Publish reusable models, components or environments to an Azure Machine Learning registry and reference the intended registered version across eligible workspaces.
- Compare fine-tuned checkpoints against baseline and held-out data for task quality, robustness, safety and runtime cost, then tune or roll back as indicated.
- Measure Foundry latency, percentile response time, throughput, error and token metrics by deployed model and realistic workload cohorts.
Explanation
Correct Answers
A:This operation directly solves a LISTED independent workstream: An HR assistant repeatedly ignores a requirement to cite policy sections because its prompt gives vague instructions and no output schema. Recommended action: Design task-specific prompts with clear context, instructions, input boundaries, structured output criteria and failure handling, and test on realistic cases. Technical reason: Prompt engineering shapes model behavior but cannot guarantee correctness or override retrieval quality, safety constraints and model capabilities. Acceptance evidence: A representative evaluation shows the prompt produces the specified answer structure, grounded context and safe fallback behavior across edge cases. Material limitation: Prompts should treat retrieved content as untrusted and avoid embedding credentials; natural-language instructions alone cannot enforce every access boundary.
B:This operation directly solves a LISTED independent workstream: An AML training sweep exhausts expensive GPU quota because no trial budget, early stopping or primary validation metric is configured. Recommended action: Configure an AML hyperparameter sweep with defined search space, primary metric, resource budget and appropriate early-termination policy. Technical reason: AML sweeps coordinate multiple training trials and compare explicitly logged metrics; search policy and budgets influence cost and selection. Acceptance evidence: Recorded trials show candidate hyperparameters, the nominated optimization metric and a reproducible best-run selection within quota limits. Material limitation: Poor validation data and incorrect metric direction can select a systematically weak model despite completing every sweep trial.
D:This operation directly solves a LISTED independent workstream: A multinational data-science platform has shared training components but project workspaces cannot discover a governed enterprise-wide component version. Recommended action: Publish reusable models, components or environments to an Azure Machine Learning registry and reference the intended registered version across eligible workspaces. Technical reason: AML registries support discoverable shared assets across workspaces with explicit versioning and access; the registry is not an inference endpoint. Acceptance evidence: An authorized workspace can resolve and reuse the published asset version with traceable lineage without re-authoring the source artifact. Material limitation: Registry access and regional replication may have distinct identity, network and artifact availability requirements across workspaces.
Incorrect Answers
C:This would be correct for an UNLISTED and separate AI operations situation: Two Foundry agent candidates have similar latency but one invents citations; the team must choose meaningful quality evaluation criteria. It would require: Run appropriate Foundry evaluation metrics for groundedness, relevance, coherence and fluency using aligned inputs, responses and source context. The appropriate success evidence would be: An evaluation report separates the named metrics and identifies unsupported citations under a reviewed rubric and dataset. The given workstreams do not ask for this operation, so it cannot replace an action needed for a listed requirement.
E:This would be correct for an UNLISTED and separate AI operations situation: A legal extraction model improves on its fine-tuning training set but performs worse on new contract templates and rare jurisdiction clauses. It would require: Compare fine-tuned checkpoints against baseline and held-out data for task quality, robustness, safety and runtime cost, then tune or roll back as indicated. The appropriate success evidence would be: A documented checkpoint comparison identifies a model meeting held-out quality and safety criteria and tracks its exact trained version. The given workstreams do not ask for this operation, so it cannot replace an action needed for a listed requirement.
F:This would be correct for an UNLISTED and separate AI operations situation: A Foundry rollout changes capacity, but the release owner has no p50/p95/p99 latency and failure-rate comparison between old and new deployments. It would require: Measure Foundry latency, percentile response time, throughput, error and token metrics by deployed model and realistic workload cohorts. The appropriate success evidence would be: A representative load test and production telemetry show p95 response time and throughput meet service objectives under expected concurrency. The given workstreams do not ask for this operation, so it cannot replace an action needed for a listed requirement.
Question 91
An Azure machine learning operations team must resolve a specific production requirement: A GitHub workflow must deploy an Azure ML training pipeline without storing a long-lived service-principal client secret in repository variables. Which engineering action MOST directly addresses this requirement?
- Measure Foundry latency, percentile response time, throughput, error and token metrics by deployed model and realistic workload cohorts.
- Store parameterized prompt templates in a protected Git repository, tag releases and link deployed prompt commit and evaluation evidence.
- Assemble representative versioned evaluation datasets with appropriate input, context, reference and response-field mapping for the target Foundry evaluator.
- Configure GitHub-to-Azure workload identity federation or another supported secure login with scoped roles and a validated AML workflow.
Explanation
Correct Answer
D:The appropriate operation for this situation is: Configure GitHub-to-Azure workload identity federation or another supported secure login with scoped roles and a validated AML workflow. Technical explanation: OIDC federation allows trusted workflow identities to request short-lived Azure tokens; repository trust and Azure RBAC both determine permitted actions. Expected validation: A permitted GitHub Actions run authenticates without a persistent checked-in secret and accesses only the approved AML workspace resources. Critical boundary: A federated credential linked to the wrong repository branch or audience can fail even if the Azure identity has Contributor permission. A misleading shortcut to avoid: Commit a subscription-wide service principal password into the Git repo because GitHub Actions cannot authenticate to Azure securely.
Incorrect Answers
A:This is a plausible engineering operation for a DIFFERENT Azure AI case: A team reports model throughput based only on request counts even though output token volumes vary drastically across workloads. The separate case would use: Measure Foundry latency, percentile response time, throughput, error and token metrics by deployed model and realistic workload cohorts. Its technical principle is: Tail latency and output token length can dominate perceived AI service quality even when average CPU or average response time appears healthy. However, the question actually describes: A GitHub workflow must deploy an Azure ML training pipeline without storing a long-lived service-principal client secret in repository variables. For that requirement, the relevant decision is: Configure GitHub-to-Azure workload identity federation or another supported secure login with scoped roles and a validated AML workflow. The offered answer does not directly resolve or verify the specified case, although it is useful in the other context.
B:This is a plausible engineering operation for a DIFFERENT Azure AI case: A team uses the same prompt file across three projects and must track which commit generated each evaluation run and production deployment. The separate case would use: Store parameterized prompt templates in a protected Git repository, tag releases and link deployed prompt commit and evaluation evidence. Its technical principle is: Git versioning provides a reviewable prompt history when the deployment pipeline actually binds the intended source commit to running configuration. However, the question actually describes: A GitHub workflow must deploy an Azure ML training pipeline without storing a long-lived service-principal client secret in repository variables. For that requirement, the relevant decision is: Configure GitHub-to-Azure workload identity federation or another supported secure login with scoped roles and a validated AML workflow. The offered answer does not directly resolve or verify the specified case, although it is useful in the other context.
C:This is a plausible engineering operation for a DIFFERENT Azure AI case: An agent evaluator receives JSONL records with mismatched input field names, so the prompt, response and context mappings are empty. The separate case would use: Assemble representative versioned evaluation datasets with appropriate input, context, reference and response-field mapping for the target Foundry evaluator. Its technical principle is: Generative evaluation relies on aligned test records and correct field mapping; incomplete or unrepresentative datasets undermine measured quality. However, the question actually describes: A GitHub workflow must deploy an Azure ML training pipeline without storing a long-lived service-principal client secret in repository variables. For that requirement, the relevant decision is: Configure GitHub-to-Azure workload identity federation or another supported secure login with scoped roles and a validated AML workflow. The offered answer does not directly resolve or verify the specified case, although it is useful in the other context.
Question 92
The engineering team proposes to: Configure GitHub-to-Azure workload identity federation or another supported secure login with scoped roles and a validated AML workflow. The original incident was: A GitHub workflow must deploy an Azure ML training pipeline without storing a long-lived service-principal client secret in repository variables. Which observable result BEST verifies success?
- A monitoring dashboard associates recent production samples with active deployment versions and surfaces defined quality or safety threshold breaches.
- A selected MLflow run shows expected parameters, dataset or code provenance, metrics and versioned artifacts accessible to authorized reviewers.
- A permitted GitHub Actions run authenticates without a persistent checked-in secret and accesses only the approved AML workspace resources.
- A usage report reconciles billed units with per-workflow token consumption and documents measured savings while quality tests still pass.
Explanation
Correct Answer
C:This result gives direct evidence: A permitted GitHub Actions run authenticates without a persistent checked-in secret and accesses only the approved AML workspace resources. It validates the selected operation: Configure GitHub-to-Azure workload identity federation or another supported secure login with scoped roles and a validated AML workflow. The technical basis is: OIDC federation allows trusted workflow identities to request short-lived Azure tokens; repository trust and Azure RBAC both determine permitted actions. Interpret the result with this limitation: A federated credential linked to the wrong repository branch or audience can fail even if the Azure identity has Contributor permission.
Incorrect Answers
A:This is a plausible engineering operation for a DIFFERENT Azure AI case: A production agent is deployed to multiple regions and the operations team needs continuous view of response quality and safety drift across versions. The separate case would use: A monitoring dashboard associates recent production samples with active deployment versions and surfaces defined quality or safety threshold breaches. Its technical principle is: Offline predeployment checks and continuous postdeployment monitoring address different risks as user distribution, dependencies and prompts evolve. However, the question actually describes: A GitHub workflow must deploy an Azure ML training pipeline without storing a long-lived service-principal client secret in repository variables. For that requirement, the relevant acceptance is: A permitted GitHub Actions run authenticates without a persistent checked-in secret and accesses only the approved AML workspace resources. The offered answer does not directly resolve or verify the specified case, although it is useful in the other context.
B:This is a plausible engineering operation for a DIFFERENT Azure AI case: A manufacturing team reports improved model accuracy but cannot identify the MLflow run that recorded the original training parameters and dataset version. The separate case would use: A selected MLflow run shows expected parameters, dataset or code provenance, metrics and versioned artifacts accessible to authorized reviewers. Its technical principle is: MLflow tracking stores per-run lineage and performance evidence; it must be integrated into actual training code and governed experiment storage. However, the question actually describes: A GitHub workflow must deploy an Azure ML training pipeline without storing a long-lived service-principal client secret in repository variables. For that requirement, the relevant acceptance is: A permitted GitHub Actions run authenticates without a persistent checked-in secret and accesses only the approved AML workspace resources. The offered answer does not directly resolve or verify the specified case, although it is useful in the other context.
D:This is a plausible engineering operation for a DIFFERENT Azure AI case: An agent loops through multiple expensive tool calls for each question and engineers need to attribute token and retrieval costs by workflow. The separate case would use: A usage report reconciles billed units with per-workflow token consumption and documents measured savings while quality tests still pass. Its technical principle is: Generative AI cost depends on token usage, model and capacity, retrieval and agent tool workânot just the number of API calls. However, the question actually describes: A GitHub workflow must deploy an Azure ML training pipeline without storing a long-lived service-principal client secret in repository variables. For that requirement, the relevant acceptance is: A permitted GitHub Actions run authenticates without a persistent checked-in secret and accesses only the approved AML workspace resources. The offered answer does not directly resolve or verify the specified case, although it is useful in the other context.
Question 93
A skeptical operations reviewer is documenting what an apparently sound engineering plan cannot guarantee. Proposed change: Configure GitHub-to-Azure workload identity federation or another supported secure login with scoped roles and a validated AML workflow. The initiating incident is: A GitHub workflow must deploy an Azure ML training pipeline without storing a long-lived service-principal client secret in repository variables. Which specific hidden constraint or failure condition must remain in the risk register?
- Uncalibrated vector-score thresholds can discard useful rare documents or admit irrelevant context, and arbitrary chunking can split facts.
- A federated credential linked to the wrong repository branch or audience can fail even if the Azure identity has Contributor permission.
- Hybrid ranking cannot cure missing indexed content, invalid embedding fields or violated security filters; tuning must compare relevant cohorts.
- Automatic metrics and LLM-as-judge scores can be noisy, model-dependent and should be checked against representative human review.
Explanation
Correct Answer
B:The key limitation is: A federated credential linked to the wrong repository branch or audience can fail even if the Azure identity has Contributor permission. This matters for: A GitHub workflow must deploy an Azure ML training pipeline without storing a long-lived service-principal client secret in repository variables. The team plans: Configure GitHub-to-Azure workload identity federation or another supported secure login with scoped roles and a validated AML workflow. Its measurable success criterion is: A permitted GitHub Actions run authenticates without a persistent checked-in secret and accesses only the approved AML workspace resources. Underlying principle: OIDC federation allows trusted workflow identities to request short-lived Azure tokens; repository trust and Azure RBAC both determine permitted actions.
Incorrect Answers
A:This is a plausible engineering operation for a DIFFERENT Azure AI case: A RAG internal handbook bot retrieves entire 200-page policy manuals and loses relevant paragraphs because no text chunk boundaries were planned. The separate case would use: Uncalibrated vector-score thresholds can discard useful rare documents or admit irrelevant context, and arbitrary chunking can split facts. Its technical principle is: Retrieval quality depends on chunk segmentation, embedding representation and search parameters; changing one threshold shifts recall and precision. However, the question actually describes: A GitHub workflow must deploy an Azure ML training pipeline without storing a long-lived service-principal client secret in repository variables. For that requirement, the relevant risk is: A federated credential linked to the wrong repository branch or audience can fail even if the Azure identity has Contributor permission. The offered answer does not directly resolve or verify the specified case, although it is useful in the other context.
C:This is a plausible engineering operation for a DIFFERENT Azure AI case: An IT-helpdesk RAG bot answers conceptual questions but fails to retrieve exact error codes such as AADSTS50076 from a policy knowledge base. The separate case would use: Hybrid ranking cannot cure missing indexed content, invalid embedding fields or violated security filters; tuning must compare relevant cohorts. Its technical principle is: Hybrid search merges complementary keyword and semantic candidates using rank fusion and may improve recall across exact identifiers and paraphrases. However, the question actually describes: A GitHub workflow must deploy an Azure ML training pipeline without storing a long-lived service-principal client secret in repository variables. For that requirement, the relevant risk is: A federated credential linked to the wrong repository branch or audience can fail even if the Azure identity has Contributor permission. The offered answer does not directly resolve or verify the specified case, although it is useful in the other context.
D:This is a plausible engineering operation for a DIFFERENT Azure AI case: A legal assistant produces fluent summaries that sometimes cite passages unrelated to the user's contract, and developers need distinct quality metrics. The separate case would use: Automatic metrics and LLM-as-judge scores can be noisy, model-dependent and should be checked against representative human review. Its technical principle is: Quality dimensions capture different failure modes; a fluent, coherent answer can still be ungrounded or irrelevant to the user request. However, the question actually describes: A GitHub workflow must deploy an Azure ML training pipeline without storing a long-lived service-principal client secret in repository variables. For that requirement, the relevant risk is: A federated credential linked to the wrong repository branch or audience can fail even if the Azure identity has Contributor permission. The offered answer does not directly resolve or verify the specified case, although it is useful in the other context.
Question 94
A model-operations architect is explaining why a solution fits this incident: A GitHub workflow must deploy an Azure ML training pipeline without storing a long-lived service-principal client secret in repository variables. Which underlying technical principle best supports an appropriate remedy?
- A pipeline coordinates a dependency graph of reproducible steps; data lineage and component cache semantics depend on declared inputs.
- Fine-tuning adjusts model behavior and has distinct training and compatibility trade-offs compared with RAG or prompt-only configuration.
- OIDC federation allows trusted workflow identities to request short-lived Azure tokens; repository trust and Azure RBAC both determine permitted actions.
- Safety evaluations probe defined risk categories and attack surfaces; detection performance depends on datasets, language, policy and model behavior.
Explanation
Correct Answer
C:The correct technical principle is: OIDC federation allows trusted workflow identities to request short-lived Azure tokens; repository trust and Azure RBAC both determine permitted actions. For the described situation: A GitHub workflow must deploy an Azure ML training pipeline without storing a long-lived service-principal client secret in repository variables. The relevant operation is: Configure GitHub-to-Azure workload identity federation or another supported secure login with scoped roles and a validated AML workflow. A useful acceptance test is: A permitted GitHub Actions run authenticates without a persistent checked-in secret and accesses only the approved AML workspace resources. Important exception or risk: A federated credential linked to the wrong repository branch or audience can fail even if the Azure identity has Contributor permission. Avoid the mistaken assumption: Commit a subscription-wide service principal password into the Git repo because GitHub Actions cannot authenticate to Azure securely.
Incorrect Answers
A:This is a plausible engineering operation for a DIFFERENT Azure AI case: An AML pipeline must reuse cached preprocessing outputs only when the source data and component versions genuinely match its inputs. The separate case would use: A pipeline coordinates a dependency graph of reproducible steps; data lineage and component cache semantics depend on declared inputs. Its technical principle is: A pipeline coordinates a dependency graph of reproducible steps; data lineage and component cache semantics depend on declared inputs. However, the question actually describes: A GitHub workflow must deploy an Azure ML training pipeline without storing a long-lived service-principal client secret in repository variables. For that requirement, the relevant principle is: OIDC federation allows trusted workflow identities to request short-lived Azure tokens; repository trust and Azure RBAC both determine permitted actions. The offered answer does not directly resolve or verify the specified case, although it is useful in the other context.
B:This is a plausible engineering operation for a DIFFERENT Azure AI case: An experimental fine-tuning run has unstable loss and catastrophic forgetting after a data format change, so the training method needs reevaluation. The separate case would use: Fine-tuning adjusts model behavior and has distinct training and compatibility trade-offs compared with RAG or prompt-only configuration. Its technical principle is: Fine-tuning adjusts model behavior and has distinct training and compatibility trade-offs compared with RAG or prompt-only configuration. However, the question actually describes: A GitHub workflow must deploy an Azure ML training pipeline without storing a long-lived service-principal client secret in repository variables. For that requirement, the relevant principle is: OIDC federation allows trusted workflow identities to request short-lived Azure tokens; repository trust and Azure RBAC both determine permitted actions. The offered answer does not directly resolve or verify the specified case, although it is useful in the other context.
D:This is a plausible engineering operation for a DIFFERENT Azure AI case: A multilingual public agent needs a risk evaluation dataset covering prompt injection, protected information and abusive output across supported languages. The separate case would use: Safety evaluations probe defined risk categories and attack surfaces; detection performance depends on datasets, language, policy and model behavior. Its technical principle is: Safety evaluations probe defined risk categories and attack surfaces; detection performance depends on datasets, language, policy and model behavior. However, the question actually describes: A GitHub workflow must deploy an Azure ML training pipeline without storing a long-lived service-principal client secret in repository variables. For that requirement, the relevant principle is: OIDC federation allows trusted workflow identities to request short-lived Azure tokens; repository trust and Azure RBAC both determine permitted actions. The offered answer does not directly resolve or verify the specified case, although it is useful in the other context.
Question 95
An AI operations engineer must resolve 2 INDEPENDENT workstreams. Workstream A: A GitHub workflow must deploy an Azure ML training pipeline without storing a long-lived service-principal client secret in repository variables. Workstream B: A data scientist is iterating feature transformations interactively and must persist experimental results from an AML compute instance into a governed workspace. Select exactly 2 actions, ONE directly solving EACH listed independent workstream.
- Deploy supported Foundry resources, managed identities and dependencies through versioned Bicep templates and scoped Azure CLI automation with change review.
- Configure GitHub-to-Azure workload identity federation or another supported secure login with scoped roles and a validated AML workflow.
- Generate and curate synthetic fine-tuning data with controlled prompts, provenance, privacy checks, deduplication, label validation and balanced domain coverage.
- Implement Azure AI Search hybrid queries combining full-text BM25, vector retrieval and optional semantic reranking, then tune fusion and filters.
- Use an approved AML-connected notebook environment with reproducible dependencies, versioned code and identity-backed data access for experiments.
Explanation
Correct Answers
B:This operation directly solves a LISTED independent workstream: A GitHub workflow must deploy an Azure ML training pipeline without storing a long-lived service-principal client secret in repository variables. Recommended action: Configure GitHub-to-Azure workload identity federation or another supported secure login with scoped roles and a validated AML workflow. Technical reason: OIDC federation allows trusted workflow identities to request short-lived Azure tokens; repository trust and Azure RBAC both determine permitted actions. Acceptance evidence: A permitted GitHub Actions run authenticates without a persistent checked-in secret and accesses only the approved AML workspace resources. Material limitation: A federated credential linked to the wrong repository branch or audience can fail even if the Azure identity has Contributor permission.
E:This operation directly solves a LISTED independent workstream: A data scientist is iterating feature transformations interactively and must persist experimental results from an AML compute instance into a governed workspace. Recommended action: Use an approved AML-connected notebook environment with reproducible dependencies, versioned code and identity-backed data access for experiments. Technical reason: Notebooks support exploratory execution but cell-order state, local files and secret handling can make results irreproducible unless governed. Acceptance evidence: A fresh notebook kernel with approved environment and data versions reproduces the intended experiment without exposing credentials in source. Material limitation: Executing cells out of order can leave stale state; notebook success does not establish production-grade scheduled pipelines.
Incorrect Answers
A:This would be correct for an UNLISTED and separate AI operations situation: A platform team manually builds Foundry projects with inconsistent network restrictions across three regions and needs a repeatable infrastructure definition. It would require: Deploy supported Foundry resources, managed identities and dependencies through versioned Bicep templates and scoped Azure CLI automation with change review. The appropriate success evidence would be: A reviewed infrastructure deployment creates only the intended Foundry resources with correct identities, network properties and documented change history. The given workstreams do not ask for this operation, so it cannot replace an action needed for a listed requirement.
C:This would be correct for an UNLISTED and separate AI operations situation: An AI training dataset includes generated financial transactions but the team cannot distinguish synthetic labels from audited real outcome records. It would require: Generate and curate synthetic fine-tuning data with controlled prompts, provenance, privacy checks, deduplication, label validation and balanced domain coverage. The appropriate success evidence would be: A curated evaluation set confirms useful edge-case coverage, accurate labels and no protected information copied verbatim into synthetic examples. The given workstreams do not ask for this operation, so it cannot replace an action needed for a listed requirement.
D:This would be correct for an UNLISTED and separate AI operations situation: A procurement assistant needs precise SKU matching and semantic similarity ranked together, with tenant filter access restrictions preserved. It would require: Implement Azure AI Search hybrid queries combining full-text BM25, vector retrieval and optional semantic reranking, then tune fusion and filters. The appropriate success evidence would be: A labeled benchmark retrieves both exact codes and semantic matches while preserving required tenant and document-level filters. The given workstreams do not ask for this operation, so it cannot replace an action needed for a listed requirement.
Question 96
An AI-300 engineer receives the following deployment report: A repository clone cannot read a private AML resource after a team rotates an expired GitHub deployment credential, despite unchanged training code. Select the BEST next configuration or operation for this case.
- Configure GitHub-to-Azure workload identity federation or another supported secure login with scoped roles and a validated AML workflow.
- Automate Foundry evaluations with supported built-in and carefully validated custom evaluators, versioned datasets, thresholds and release gates.
- Apply documented Azure ML model versioning, tags and archiving or stage governance, with explicit retention, ownership and rollback policy.
- Configure guarded Azure ML model or endpoint monitoring alerts and approved pipeline retraining triggers with threshold windows and ownership.
Explanation
Correct Answer
A:The appropriate operation for this situation is: Configure GitHub-to-Azure workload identity federation or another supported secure login with scoped roles and a validated AML workflow. Technical explanation: OIDC federation allows trusted workflow identities to request short-lived Azure tokens; repository trust and Azure RBAC both determine permitted actions. Expected validation: A permitted GitHub Actions run authenticates without a persistent checked-in secret and accesses only the approved AML workspace resources. Critical boundary: A federated credential linked to the wrong repository branch or audience can fail even if the Azure identity has Contributor permission. A misleading shortcut to avoid: Commit a subscription-wide service principal password into the Git repo because GitHub Actions cannot authenticate to Azure securely.
Incorrect Answers
B:This is a plausible engineering operation for a DIFFERENT Azure AI case: A GenAI team has manual quality checks before deployment but one prompt regression was released because no evaluation workflow ran in CI. The separate case would use: Automate Foundry evaluations with supported built-in and carefully validated custom evaluators, versioned datasets, thresholds and release gates. Its technical principle is: Repeatable evaluation workflows reduce manual omissions while custom scoring logic must be reviewed, calibrated and versioned alongside models. However, the question actually describes: A repository clone cannot read a private AML resource after a team rotates an expired GitHub deployment credential, despite unchanged training code. For that requirement, the relevant decision is: Configure GitHub-to-Azure workload identity federation or another supported secure login with scoped roles and a validated AML workflow. The offered answer does not directly resolve or verify the specified case, although it is useful in the other context.
C:This is a plausible engineering operation for a DIFFERENT Azure AI case: An audit asks why an older approved model version was replaced and what rollback artifact is still retained after archiving. The separate case would use: Apply documented Azure ML model versioning, tags and archiving or stage governance, with explicit retention, ownership and rollback policy. Its technical principle is: Lifecycle governance distinguishes model registration, approval, archive state and active endpoint traffic; each must be managed independently. However, the question actually describes: A repository clone cannot read a private AML resource after a team rotates an expired GitHub deployment credential, despite unchanged training code. For that requirement, the relevant decision is: Configure GitHub-to-Azure workload identity federation or another supported secure login with scoped roles and a validated AML workflow. The offered answer does not directly resolve or verify the specified case, although it is useful in the other context.
D:This is a plausible engineering operation for a DIFFERENT Azure AI case: A classifier's p95 prediction latency exceeds its service objective during a traffic surge yet no alert or remediation owner receives the signal. The separate case would use: Configure guarded Azure ML model or endpoint monitoring alerts and approved pipeline retraining triggers with threshold windows and ownership. Its technical principle is: Automated alerts and retraining must be tied to validated signals and governed deployment gates rather than unreviewed model replacement. However, the question actually describes: A repository clone cannot read a private AML resource after a team rotates an expired GitHub deployment credential, despite unchanged training code. For that requirement, the relevant decision is: Configure GitHub-to-Azure workload identity federation or another supported secure login with scoped roles and a validated AML workflow. The offered answer does not directly resolve or verify the specified case, although it is useful in the other context.
Question 97
A post-change review considers this requirement: A repository clone cannot read a private AML resource after a team rotates an expired GitHub deployment credential, despite unchanged training code. Engineers chose: Configure GitHub-to-Azure workload identity federation or another supported secure login with scoped roles and a validated AML workflow. Which acceptance evidence is most convincing?
- A release ledger identifies the exact served model revision and a test rollback restores an approved baseline when a quality gate fails.
- A representative client authenticates to the selected model endpoint and meets documented latency, cost and request-capacity expectations.
- A permitted GitHub Actions run authenticates without a persistent checked-in secret and accesses only the approved AML workspace resources.
- Recorded trials show candidate hyperparameters, the nominated optimization metric and a reproducible best-run selection within quota limits.
Explanation
Correct Answer
C:This result gives direct evidence: A permitted GitHub Actions run authenticates without a persistent checked-in secret and accesses only the approved AML workspace resources. It validates the selected operation: Configure GitHub-to-Azure workload identity federation or another supported secure login with scoped roles and a validated AML workflow. The technical basis is: OIDC federation allows trusted workflow identities to request short-lived Azure tokens; repository trust and Azure RBAC both determine permitted actions. Interpret the result with this limitation: A federated credential linked to the wrong repository branch or audience can fail even if the Azure identity has Contributor permission.
Incorrect Answers
A:This is a plausible engineering operation for a DIFFERENT Azure AI case: A Foundry model provider retires an old revision and teams must identify affected endpoints before the scheduled model rollover. The separate case would use: A release ledger identifies the exact served model revision and a test rollback restores an approved baseline when a quality gate fails. Its technical principle is: Model release management links versioned model, endpoint deployment, prompts and evaluation evidence; updating one component can alter output behavior. However, the question actually describes: A repository clone cannot read a private AML resource after a team rotates an expired GitHub deployment credential, despite unchanged training code. For that requirement, the relevant acceptance is: A permitted GitHub Actions run authenticates without a persistent checked-in secret and accesses only the approved AML workspace resources. The offered answer does not directly resolve or verify the specified case, although it is useful in the other context.
B:This is a plausible engineering operation for a DIFFERENT Azure AI case: A research firm must deploy a custom open foundation model that needs its own managed GPU inference environment rather than a serverless API. The separate case would use: A representative client authenticates to the selected model endpoint and meets documented latency, cost and request-capacity expectations. Its technical principle is: Foundation model serving modes have different capacity, billing, region, version and runtime requirements; they are not interchangeable. However, the question actually describes: A repository clone cannot read a private AML resource after a team rotates an expired GitHub deployment credential, despite unchanged training code. For that requirement, the relevant acceptance is: A permitted GitHub Actions run authenticates without a persistent checked-in secret and accesses only the approved AML workspace resources. The offered answer does not directly resolve or verify the specified case, although it is useful in the other context.
D:This is a plausible engineering operation for a DIFFERENT Azure AI case: An AML training sweep exhausts expensive GPU quota because no trial budget, early stopping or primary validation metric is configured. The separate case would use: Recorded trials show candidate hyperparameters, the nominated optimization metric and a reproducible best-run selection within quota limits. Its technical principle is: AML sweeps coordinate multiple training trials and compare explicitly logged metrics; search policy and budgets influence cost and selection. However, the question actually describes: A repository clone cannot read a private AML resource after a team rotates an expired GitHub deployment credential, despite unchanged training code. For that requirement, the relevant acceptance is: A permitted GitHub Actions run authenticates without a persistent checked-in secret and accesses only the approved AML workspace resources. The offered answer does not directly resolve or verify the specified case, although it is useful in the other context.
Question 98
The owner is considering deployment prerequisites and possible unintended consequences, not simply whether a configuration exists. Planned response: Configure GitHub-to-Azure workload identity federation or another supported secure login with scoped roles and a validated AML workflow. Original business problem: A repository clone cannot read a private AML resource after a team rotates an expired GitHub deployment credential, despite unchanged training code. What qualification is most important before release?
- Explanations and group metrics depend on data quality and context; dashboard availability is not proof that a model is legally or ethically safe.
- Synthetic datasets require privacy, edge-case coverage and bias review; a clean schema alone does not demonstrate representative production behavior.
- A federated credential linked to the wrong repository branch or audience can fail even if the Azure identity has Contributor permission.
- Tracking run metadata is not proof that a metric was calculated fairly or that the underlying dataset is immutable.
Explanation
Correct Answer
C:The key limitation is: A federated credential linked to the wrong repository branch or audience can fail even if the Azure identity has Contributor permission. This matters for: A repository clone cannot read a private AML resource after a team rotates an expired GitHub deployment credential, despite unchanged training code. The team plans: Configure GitHub-to-Azure workload identity federation or another supported secure login with scoped roles and a validated AML workflow. Its measurable success criterion is: A permitted GitHub Actions run authenticates without a persistent checked-in secret and accesses only the approved AML workspace resources. Underlying principle: OIDC federation allows trusted workflow identities to request short-lived Azure tokens; repository trust and Azure RBAC both determine permitted actions.
Incorrect Answers
A:This is a plausible engineering operation for a DIFFERENT Azure AI case: A healthcare risk model is retrained using historical decisions and the audit team wants to inspect feature explanations and error disparities. The separate case would use: Explanations and group metrics depend on data quality and context; dashboard availability is not proof that a model is legally or ethically safe. Its technical principle is: Responsible AI review considers reliability, fairness, transparency and relevant safety requirements beyond a single aggregate accuracy metric. However, the question actually describes: A repository clone cannot read a private AML resource after a team rotates an expired GitHub deployment credential, despite unchanged training code. For that requirement, the relevant risk is: A federated credential linked to the wrong repository branch or audience can fail even if the Azure identity has Contributor permission. The offered answer does not directly resolve or verify the specified case, although it is useful in the other context.
B:This is a plausible engineering operation for a DIFFERENT Azure AI case: An agent evaluator receives JSONL records with mismatched input field names, so the prompt, response and context mappings are empty. The separate case would use: Synthetic datasets require privacy, edge-case coverage and bias review; a clean schema alone does not demonstrate representative production behavior. Its technical principle is: Generative evaluation relies on aligned test records and correct field mapping; incomplete or unrepresentative datasets undermine measured quality. However, the question actually describes: A repository clone cannot read a private AML resource after a team rotates an expired GitHub deployment credential, despite unchanged training code. For that requirement, the relevant risk is: A federated credential linked to the wrong repository branch or audience can fail even if the Azure identity has Contributor permission. The offered answer does not directly resolve or verify the specified case, although it is useful in the other context.
D:This is a plausible engineering operation for a DIFFERENT Azure AI case: An MLOps auditor asks for the exact registered model and run identifier that produced a promoted binary after a series of notebook experiments. The separate case would use: Tracking run metadata is not proof that a metric was calculated fairly or that the underlying dataset is immutable. Its technical principle is: MLflow tracking stores per-run lineage and performance evidence; it must be integrated into actual training code and governed experiment storage. However, the question actually describes: A repository clone cannot read a private AML resource after a team rotates an expired GitHub deployment credential, despite unchanged training code. For that requirement, the relevant risk is: A federated credential linked to the wrong repository branch or audience can fail even if the Azure identity has Contributor permission. The offered answer does not directly resolve or verify the specified case, although it is useful in the other context.
Question 99
An Azure AI design review considers the following failure: A repository clone cannot read a private AML resource after a team rotates an expired GitHub deployment credential, despite unchanged training code. Which platform principle justifies a correct resolution?
- RAG evaluation must connect document relevance to generated-answer grounding; rank scores alone are not proof of factual response quality.
- OIDC federation allows trusted workflow identities to request short-lived Azure tokens; repository trust and Azure RBAC both determine permitted actions.
- AML sweeps coordinate multiple training trials and compare explicitly logged metrics; search policy and budgets influence cost and selection.
- Foundry application identity, control-plane roles and downstream search or storage data access are separate authorization steps.
Explanation
Correct Answer
B:The correct technical principle is: OIDC federation allows trusted workflow identities to request short-lived Azure tokens; repository trust and Azure RBAC both determine permitted actions. For the described situation: A repository clone cannot read a private AML resource after a team rotates an expired GitHub deployment credential, despite unchanged training code. The relevant operation is: Configure GitHub-to-Azure workload identity federation or another supported secure login with scoped roles and a validated AML workflow. A useful acceptance test is: A permitted GitHub Actions run authenticates without a persistent checked-in secret and accesses only the approved AML workspace resources. Important exception or risk: A federated credential linked to the wrong repository branch or audience can fail even if the Azure identity has Contributor permission. Avoid the mistaken assumption: Commit a subscription-wide service principal password into the Git repo because GitHub Actions cannot authenticate to Azure securely.
Incorrect Answers
A:This is a plausible engineering operation for a DIFFERENT Azure AI case: A support knowledge search update changes chunk overlap and semantic reranker settings without regression tests on low-frequency product questions. The separate case would use: RAG evaluation must connect document relevance to generated-answer grounding; rank scores alone are not proof of factual response quality. Its technical principle is: RAG evaluation must connect document relevance to generated-answer grounding; rank scores alone are not proof of factual response quality. However, the question actually describes: A repository clone cannot read a private AML resource after a team rotates an expired GitHub deployment credential, despite unchanged training code. For that requirement, the relevant principle is: OIDC federation allows trusted workflow identities to request short-lived Azure tokens; repository trust and Azure RBAC both determine permitted actions. The offered answer does not directly resolve or verify the specified case, although it is useful in the other context.
C:This is a plausible engineering operation for a DIFFERENT Azure AI case: An AML training sweep exhausts expensive GPU quota because no trial budget, early stopping or primary validation metric is configured. The separate case would use: AML sweeps coordinate multiple training trials and compare explicitly logged metrics; search policy and budgets influence cost and selection. Its technical principle is: AML sweeps coordinate multiple training trials and compare explicitly logged metrics; search policy and budgets influence cost and selection. However, the question actually describes: A repository clone cannot read a private AML resource after a team rotates an expired GitHub deployment credential, despite unchanged training code. For that requirement, the relevant principle is: OIDC federation allows trusted workflow identities to request short-lived Azure tokens; repository trust and Azure RBAC both determine permitted actions. The offered answer does not directly resolve or verify the specified case, although it is useful in the other context.
D:This is a plausible engineering operation for a DIFFERENT Azure AI case: An engineer grants subscription Contributor to every prompt author to avoid repeated Foundry connection failures, violating least-privilege policy. The separate case would use: Foundry application identity, control-plane roles and downstream search or storage data access are separate authorization steps. Its technical principle is: Foundry application identity, control-plane roles and downstream search or storage data access are separate authorization steps. However, the question actually describes: A repository clone cannot read a private AML resource after a team rotates an expired GitHub deployment credential, despite unchanged training code. For that requirement, the relevant principle is: OIDC federation allows trusted workflow identities to request short-lived Azure tokens; repository trust and Azure RBAC both determine permitted actions. The offered answer does not directly resolve or verify the specified case, although it is useful in the other context.
Question 100
An AI operations engineer must resolve 3 INDEPENDENT workstreams. Workstream A: A repository clone cannot read a private AML resource after a team rotates an expired GitHub deployment credential, despite unchanged training code. Workstream B: An image classification team wants a randomized sweep instead of repeating hand-selected hyperparameters and must preserve each trial's outcome. Workstream C: A regulated AI agent needs to roll back a prompt and base-model combination after safety metrics worsen during a phased rollout. Select exactly 3 actions, ONE directly solving EACH listed independent workstream.
- Implement Azure AI Search hybrid queries combining full-text BM25, vector retrieval and optional semantic reranking, then tune fusion and filters.
- Select or adapt an eligible embedding model for the domain, consistently apply it to documents and queries, and reevaluate search index similarity.
- Configure GitHub-to-Azure workload identity federation or another supported secure login with scoped roles and a validated AML workflow.
- Configure an AML hyperparameter sweep with defined search space, primary metric, resource budget and appropriate early-termination policy.
- Configure Foundry risk and safety evaluation using representative adversarial cases, harmful-content criteria and approved safety gates.
- Track model versions and deployment identifiers, compare candidate revisions on a fixed evaluation suite, and govern staged promotion or rollback.
Explanation
Correct Answers
C:This operation directly solves a LISTED independent workstream: A repository clone cannot read a private AML resource after a team rotates an expired GitHub deployment credential, despite unchanged training code. Recommended action: Configure GitHub-to-Azure workload identity federation or another supported secure login with scoped roles and a validated AML workflow. Technical reason: OIDC federation allows trusted workflow identities to request short-lived Azure tokens; repository trust and Azure RBAC both determine permitted actions. Acceptance evidence: A permitted GitHub Actions run authenticates without a persistent checked-in secret and accesses only the approved AML workspace resources. Material limitation: A federated credential linked to the wrong repository branch or audience can fail even if the Azure identity has Contributor permission.
D:This operation directly solves a LISTED independent workstream: An image classification team wants a randomized sweep instead of repeating hand-selected hyperparameters and must preserve each trial's outcome. Recommended action: Configure an AML hyperparameter sweep with defined search space, primary metric, resource budget and appropriate early-termination policy. Technical reason: AML sweeps coordinate multiple training trials and compare explicitly logged metrics; search policy and budgets influence cost and selection. Acceptance evidence: Recorded trials show candidate hyperparameters, the nominated optimization metric and a reproducible best-run selection within quota limits. Material limitation: Poor validation data and incorrect metric direction can select a systematically weak model despite completing every sweep trial.
F:This operation directly solves a LISTED independent workstream: A regulated AI agent needs to roll back a prompt and base-model combination after safety metrics worsen during a phased rollout. Recommended action: Track model versions and deployment identifiers, compare candidate revisions on a fixed evaluation suite, and govern staged promotion or rollback. Technical reason: Model release management links versioned model, endpoint deployment, prompts and evaluation evidence; updating one component can alter output behavior. Acceptance evidence: A release ledger identifies the exact served model revision and a test rollback restores an approved baseline when a quality gate fails. Material limitation: Provider-managed version updates, availability and retirement timelines differ; not every deployment type supports identical rollback semantics.
Incorrect Answers
A:This would be correct for an UNLISTED and separate AI operations situation: An IT-helpdesk RAG bot answers conceptual questions but fails to retrieve exact error codes such as AADSTS50076 from a policy knowledge base. It would require: Implement Azure AI Search hybrid queries combining full-text BM25, vector retrieval and optional semantic reranking, then tune fusion and filters. The appropriate success evidence would be: A labeled benchmark retrieves both exact codes and semantic matches while preserving required tenant and document-level filters. The given workstreams do not ask for this operation, so it cannot replace an action needed for a listed requirement.
B:This would be correct for an UNLISTED and separate AI operations situation: An enterprise glossary has rare technical abbreviations and a generic embedding model repeatedly retrieves unrelated consumer meanings for its domain-specific queries. It would require: Select or adapt an eligible embedding model for the domain, consistently apply it to documents and queries, and reevaluate search index similarity. The appropriate success evidence would be: A controlled multilingual or terminology benchmark shows improved retrieval relevance with consistent index/query vector configuration. The given workstreams do not ask for this operation, so it cannot replace an action needed for a listed requirement.
E:This would be correct for an UNLISTED and separate AI operations situation: A multilingual public agent needs a risk evaluation dataset covering prompt injection, protected information and abusive output across supported languages. It would require: Configure Foundry risk and safety evaluation using representative adversarial cases, harmful-content criteria and approved safety gates. The appropriate success evidence would be: Controlled adversarial prompts produce the expected safety signals, refusal tests and documented gaps before deployment approval. The given workstreams do not ask for this operation, so it cannot replace an action needed for a listed requirement.
Operationalizing Machine Learning and Generative AI Solutions (AI-300) Practice Tests