Microsoft AI-300 Practice Test 5
Question 40
While preparing a secure Foundry model or agent release, the following condition occurs: A document-assistant project exists in Foundry but cannot infer because its configured model deployment name is from a different project and unavailable to the target identity. The remediation must meet enterprise access and reliability requirements without a broad bypass. What measured outcome validates the solution while preserving the stated constraints?
- Representative chat queries receive valid answers within the approved latency target without waiting for a batch output artifact.
- Test queries reach the intended deployed model with attributable project usage and without cross-project credential exposure.
- A test prompt or model deployment change affects only the development environment until an approved promotion occurs.
- The reviewed deployment succeeds without reusable secrets in committed files and old credentials no longer authorize calls.
Explanation
Correct Answer
B:Test queries reach the intended deployed model with attributable project usage and without cross-project credential exposure. A project workspace and a model deployment are distinct resources; using a logical model name without the correct deployment binding is insufficient. a direct corrective action is to bind the application to the reviewed compatible model deployment in the correct Foundry project/resource and validate the connection.
Incorrect Answers
A:Representative chat queries receive valid answers within the approved latency target without waiting for a batch output artifact. That addresses a different configuration issue rather than the main requirement. Verify instead: Test queries reach the intended deployed model with attributable project usage and without cross-project credential exposure.
C:A test prompt or model deployment change affects only the development environment until an approved promotion occurs. That addresses a different configuration issue rather than the main requirement. Verify instead: Test queries reach the intended deployed model with attributable project usage and without cross-project credential exposure.
D:The reviewed deployment succeeds without reusable secrets in committed files and old credentials no longer authorize calls. That addresses a different configuration issue rather than the main requirement. Verify instead: Test queries reach the intended deployed model with attributable project usage and without cross-project credential exposure.
Question 41
As an AI operations team checks production agent evaluation evidence, the following condition occurs: A service dashboard reports 600 requests per minute and assumes model capacity is sufficient, even though occasional huge prompts cause token throughput saturation and throttling. The workflow must produce measurable quality, safety and performance evidence. Which TWO complementary choices identify the issue and correct it? Select TWO.
- Use relevant retrieval context, justified response limits and supported caching after verifying quality on evaluation examples.
- Inspect tokens per request, input/output token ratios, concurrency, throttling status and model deployment quota.
- Inspect agent instrumentation, OpenTelemetry trace propagation, tool span durations, error attributes and correlated resource logs.
- Add token-aware throughput and queue metrics and apply capacity/pacing policies matching actual peak token demand.
- Add context-groundedness and relevance evaluation on representative answers alongside coherence and fluency.
Explanation
Correct Answers
B:Inspect tokens per request, input/output token ratios, concurrency, throttling status and model deployment quota. This is the most relevant evidence for genai throughput report mixes responses and tokens. Request count and generated-token throughput are different workload dimensions; equal request rates can create very different serving loads. The matching controlled remediation is: Add token-aware throughput and queue metrics and apply capacity/pacing policies matching actual peak token demand.
D:Add token-aware throughput and queue metrics and apply capacity/pacing policies matching actual peak token demand. This targets genai throughput report mixes responses and tokens. Request count and generated-token throughput are different workload dimensions; equal request rates can create very different serving loads. Verify the failure through: Inspect tokens per request, input/output token ratios, concurrency, throttling status and model deployment quota. Validate success with: The service meets per-minute token and tail-latency targets under representative large-context requests.
Incorrect Answers
A:Use relevant retrieval context, justified response limits and supported caching after verifying quality on evaluation examples. This is associated with prompt truncation strategy increases generated-token spend, not the essential evidence-and-fix pair for genai throughput report mixes responses and tokens. First examine: Inspect tokens per request, input/output token ratios, concurrency, throttling status and model deployment quota. Then correct: Add token-aware throughput and queue metrics and apply capacity/pacing policies matching actual peak token demand.
C:Inspect agent instrumentation, OpenTelemetry trace propagation, tool span durations, error attributes and correlated resource logs. This is associated with agent traces omit downstream tool failure reason, not the essential evidence-and-fix pair for genai throughput report mixes responses and tokens. First examine: Inspect tokens per request, input/output token ratios, concurrency, throttling status and model deployment quota. Then correct: Add token-aware throughput and queue metrics and apply capacity/pacing policies matching actual peak token demand.
E:Add context-groundedness and relevance evaluation on representative answers alongside coherence and fluency. This is associated with groundedness metric conflated with response fluency, not the essential evidence-and-fix pair for genai throughput report mixes responses and tokens. First examine: Inspect tokens per request, input/output token ratios, concurrency, throttling status and model deployment quota. Then correct: Add token-aware throughput and queue metrics and apply capacity/pacing policies matching actual peak token demand.
Question 42
While optimizing a production RAG or fine-tuned model service, the following condition occurs: A RAG search index blends keyword and vector results but weights retired versions of a procedure equally with current publications, so old policy text appears first. The resulting model must meet held-out quality and security criteria. Which observable postchange result proves the intended remediation succeeded?
- Current authoritative procedures rank ahead of retired versions for evaluated test queries while archival retrieval remains available to approved users.
- Representative domain queries retrieve relevant passages with aligned embedding provenance and the old index can be rolled back safely.
- Previously unseen product queries are monitored and subsequent vetted versions meet current quality targets without erasing old task competence.
- Domain terminology retrieval improves without unacceptable loss on routine guidance queries.
Explanation
Correct Answer
A:Current authoritative procedures rank ahead of retired versions for evaluated test queries while archival retrieval remains available to approved users. Hybrid scoring does not inherently enforce document freshness or authoritative version; metadata filtering is a separate ranking control. a direct corrective action is to filter or downrank superseded records before hybrid retrieval and use governed version metadata for source selection.
Incorrect Answers
B:Representative domain queries retrieve relevant passages with aligned embedding provenance and the old index can be rolled back safely. That addresses a different configuration issue rather than the main requirement. Verify instead: Current authoritative procedures rank ahead of retired versions for evaluated test queries while archival retrieval remains available to approved users.
C:Previously unseen product queries are monitored and subsequent vetted versions meet current quality targets without erasing old task competence. That addresses a different configuration issue rather than the main requirement. Verify instead: Current authoritative procedures rank ahead of retired versions for evaluated test queries while archival retrieval remains available to approved users.
D:Domain terminology retrieval improves without unacceptable loss on routine guidance queries. That addresses a different configuration issue rather than the main requirement. Verify instead: Current authoritative procedures rank ahead of retired versions for evaluated test queries while archival retrieval remains available to approved users.
Question 43
While deploying reproducible Azure Machine Learning infrastructure, the following condition occurs: An ML team edits training notebooks directly on a shared VM and deploys model versions without recording the corresponding Git commit, preventing a reproducible incident rollback. The change must remain limited to authorized workspaces and avoid unreviewed production impact. Which production-safe remediation should the engineer apply to this AI workflow?
- Rebuild or correct the versioned MLTable asset transformation so the expected columns and types are supplied to downstream training.
- Bind every training job and model registration to a reviewed Git commit SHA and immutable environment/data version in the workflow.
- Gate production deployment behind approved GitHub environment protection, protected refs and status checks while keeping test deployments separate.
- Declare proper resource references or explicit dependencies and staged network prerequisites in the reviewed Bicep deployment.
Explanation
Correct Answer
B:Bind every training job and model registration to a reviewed Git commit SHA and immutable environment/data version in the workflow. This targets production training script commit cannot be traced. Source control alone does not provide provenance unless the pipeline records which revision produced the deployed asset. Verify the failure through: Inspect model metadata, job source snapshot, repository history, CI release manifest and available recorded commit identifiers. Validate success with: An incident analyst can reproduce the model version's code and dependencies from its recorded source commit.
Incorrect Answers
A:Rebuild or correct the versioned MLTable asset transformation so the expected columns and types are supplied to downstream training. This could address mltable data asset schema mismatches training code, but it does not solve the specified production training script commit cannot be traced failure. The matching remediation is: Bind every training job and model registration to a reviewed Git commit SHA and immutable environment/data version in the workflow.
C:Gate production deployment behind approved GitHub environment protection, protected refs and status checks while keeping test deployments separate. This could address github deployment pipeline omits secure environment approvals, but it does not solve the specified production training script commit cannot be traced failure. The matching remediation is: Bind every training job and model registration to a reviewed Git commit SHA and immutable environment/data version in the workflow.
D:Declare proper resource references or explicit dependencies and staged network prerequisites in the reviewed Bicep deployment. This could address bicep workspace declaration misses dependent private resources, but it does not solve the specified production training script commit cannot be traced failure. The matching remediation is: Bind every training job and model registration to a reviewed Git commit SHA and immutable environment/data version in the workflow.
Question 44
While optimizing a production RAG or fine-tuned model service, the following condition occurs: A customer support model fine-tuned on historical chats becomes inconsistent because examples mix strict JSON tool calls with casual prose for identical task types. The resulting model must meet held-out quality and security criteria. Which diagnostic evidence most directly distinguishes the described AI operations failure?
- Examine fine-tuning run outputs, model registry versions, promotion approvals, held-out evaluations and rollout routing settings.
- Inspect information update frequency, retrieval freshness requirement, model task errors, training cost and data-change workflow.
- Review A/B cohort assignment, groundedness samples, relevance labels, business success measures and unsupported-claim incidents.
- Inspect fine-tuning records, role/message formatting, target response schemas, label consistency and curated holdout behavior.
Explanation
Correct Answer
D:Inspect fine-tuning records, role/message formatting, target response schemas, label consistency and curated holdout behavior. This is the most relevant evidence for supervised fine-tuning dataset includes conflicting assistant behaviors. Supervised fine-tuning learns patterns in labeled examples; contradictory targets can degrade reliability even when many examples are provided. The matching controlled remediation is: Clean and stratify the supervised fine-tuning data to a coherent supported task format before retraining and evaluating.
Incorrect Answers
A:Examine fine-tuning run outputs, model registry versions, promotion approvals, held-out evaluations and rollout routing settings. This would investigate fine-tuned candidate bypasses promotion review after training rather than the documented supervised fine-tuning dataset includes conflicting assistant behaviors issue. The relevant investigation is: Inspect fine-tuning records, role/message formatting, target response schemas, label consistency and curated holdout behavior.
B:Inspect information update frequency, retrieval freshness requirement, model task errors, training cost and data-change workflow. This would investigate fine-tuning task selected when retrieval would better handle changing facts rather than the documented supervised fine-tuning dataset includes conflicting assistant behaviors issue. The relevant investigation is: Inspect fine-tuning records, role/message formatting, target response schemas, label consistency and curated holdout behavior.
C:Review A/B cohort assignment, groundedness samples, relevance labels, business success measures and unsupported-claim incidents. This would investigate rag a/b experiment judges only click-through not factuality rather than the documented supervised fine-tuning dataset includes conflicting assistant behaviors issue. The relevant investigation is: Inspect fine-tuning records, role/message formatting, target response schemas, label consistency and curated holdout behavior.
Question 45
During a regulated Azure ML workspace infrastructure review, the following condition occurs: A training job reads a similarly named blob container containing last quarter's features because its Azure ML datastore URI targets the staging location rather than the approved production store. The team must preserve approved access, model provenance and private network constraints. Which TWO complementary choices identify the issue and correct it? Select TWO.
- Bind every training job and model registration to a reviewed Git commit SHA and immutable environment/data version in the workflow.
- Update the datastore or data asset binding to the governed storage path with an explicitly pinned version and validated read access.
- Check job identity selection, storage data-plane authorization, private endpoint reachability and storage diagnostic access-denied records.
- Request approved additional quota or rescope the job to a permitted compute size and region while retaining necessary GPU capability.
- Compare the datastore URI and backing storage container, recorded data asset path, job input mount path and expected data version checksum.
Explanation
Correct Answers
B:Update the datastore or data asset binding to the governed storage path with an explicitly pinned version and validated read access. This targets datastore uri selects the incorrect storage container. A datastore connection and data asset version identify where training inputs come from; paths must be validated independently from authentication success. Verify the failure through: Compare the datastore URI and backing storage container, recorded data asset path, job input mount path and expected data version checksum. Validate success with: The job consumes the intended dataset version and staging files are not silently mixed into production runs.
E:Compare the datastore URI and backing storage container, recorded data asset path, job input mount path and expected data version checksum. This is the most relevant evidence for datastore uri selects the incorrect storage container. A datastore connection and data asset version identify where training inputs come from; paths must be validated independently from authentication success. The matching controlled remediation is: Update the datastore or data asset binding to the governed storage path with an explicitly pinned version and validated read access.
Incorrect Answers
A:Bind every training job and model registration to a reviewed Git commit SHA and immutable environment/data version in the workflow. This is associated with production training script commit cannot be traced, not the essential evidence-and-fix pair for datastore uri selects the incorrect storage container. First examine: Compare the datastore URI and backing storage container, recorded data asset path, job input mount path and expected data version checksum. Then correct: Update the datastore or data asset binding to the governed storage path with an explicitly pinned version and validated read access.
C:Check job identity selection, storage data-plane authorization, private endpoint reachability and storage diagnostic access-denied records. This is associated with job identity lacks blob reader data-plane rights, not the essential evidence-and-fix pair for datastore uri selects the incorrect storage container. First examine: Compare the datastore URI and backing storage container, recorded data asset path, job input mount path and expected data version checksum. Then correct: Update the datastore or data asset binding to the governed storage path with an explicitly pinned version and validated read access.
D:Request approved additional quota or rescope the job to a permitted compute size and region while retaining necessary GPU capability. This is associated with compute cluster cannot allocate requested gpu quota, not the essential evidence-and-fix pair for datastore uri selects the incorrect storage container. First examine: Compare the datastore URI and backing storage container, recorded data asset path, job input mount path and expected data version checksum. Then correct: Update the datastore or data asset binding to the governed storage path with an explicitly pinned version and validated read access.
Question 46
During a production Azure Machine Learning model lifecycle incident, the following condition occurs: A deployed churn model receives increasing nulls in a new CRM field but its data drift monitor tracks only average values and ignores missingness and schema validity. The operator must safeguard training reproducibility and serving continuity. What should the engineer examine first to confirm the specific AI-300 workload issue?
- Inspect prediction and ground-truth join keys, label availability delay, inference logging and published monitoring metrics.
- Inspect model input data quality signals, null fractions, expected feature schema, training baseline and monitoring sampling.
- Review monitoring window, input sample count, drift statistic, alert routing and retraining schedule history.
- Inspect distributed job settings, node count, rendezvous endpoint configuration, compute network security rules and worker log timeouts.
Explanation
Correct Answer
B:Inspect model input data quality signals, null fractions, expected feature schema, training baseline and monitoring sampling. This is the most relevant evidence for missing feature values escape input drift checks. Data quality and data distribution drift are related but distinct signals; averages alone can conceal feature loss. The matching controlled remediation is: Add suitable data quality and schema drift signals with thresholds for missingness alongside distribution drift.
Incorrect Answers
A:Inspect prediction and ground-truth join keys, label availability delay, inference logging and published monitoring metrics. This would investigate accuracy monitoring lacks delayed outcome labels rather than the documented missing feature values escape input drift checks issue. The relevant investigation is: Inspect model input data quality signals, null fractions, expected feature schema, training baseline and monitoring sampling.
C:Review monitoring window, input sample count, drift statistic, alert routing and retraining schedule history. This would investigate retraining alarm triggers on insufficient sample volume rather than the documented missing feature values escape input drift checks issue. The relevant investigation is: Inspect model input data quality signals, null fractions, expected feature schema, training baseline and monitoring sampling.
D:Inspect distributed job settings, node count, rendezvous endpoint configuration, compute network security rules and worker log timeouts. This would investigate distributed job workers cannot communicate over required network path rather than the documented missing feature values escape input drift checks issue. The relevant investigation is: Inspect model input data quality signals, null fractions, expected feature schema, training baseline and monitoring sampling.
Question 47
As an MLOps engineer reviews a model training or inference pipeline, the following condition occurs: An Azure ML sweep job ranks candidate forecasting models by validation loss although the operational requirement is minimum mean absolute error, causing a poorer business model to win. The resolution must preserve quality gates, documented artifacts and rollback options. What should the engineer examine first to confirm the specific AI-300 workload issue?
- Examine cohort labels, fairness metrics, stratified error rates, consent constraints and model evaluation dataset representativeness.
- Examine sweep primary_metric, goal direction, logged metric names, candidate ranking and the held-out evaluation calculation.
- Check pipeline YAML, component inputs and outputs, resource dependency graph and job execution chronology.
- Inspect Python RNG seeding, framework deterministic settings, dataset split creation, logged run parameters and artifact hashes.
Explanation
Correct Answer
B:Examine sweep primary_metric, goal direction, logged metric names, candidate ranking and the held-out evaluation calculation. This is the most relevant evidence for hyperparameter sweep minimizes the wrong objective. Hyperparameter sweep selection depends on the configured primary metric and goal, not an implicit interpretation of business success. The matching controlled remediation is: Configure the sweep to optimize the explicitly logged business-aligned metric with the correct minimize/maximize goal.
Incorrect Answers
A:Examine cohort labels, fairness metrics, stratified error rates, consent constraints and model evaluation dataset representativeness. This would investigate responsible ai evaluation omits demographic cohort slices rather than the documented hyperparameter sweep minimizes the wrong objective issue. The relevant investigation is: Examine sweep primary_metric, goal direction, logged metric names, candidate ranking and the held-out evaluation calculation.
C:Check pipeline YAML, component inputs and outputs, resource dependency graph and job execution chronology. This would investigate pipeline graph omits feature-build dependency rather than the documented hyperparameter sweep minimizes the wrong objective issue. The relevant investigation is: Examine sweep primary_metric, goal direction, logged metric names, candidate ranking and the held-out evaluation calculation.
D:Inspect Python RNG seeding, framework deterministic settings, dataset split creation, logged run parameters and artifact hashes. This would investigate training script ignores reproducible seed and input contract rather than the documented hyperparameter sweep minimizes the wrong objective issue. The relevant investigation is: Examine sweep primary_metric, goal direction, logged metric names, candidate ranking and the held-out evaluation calculation.
Question 48
During a production Azure Machine Learning model lifecycle incident, the following condition occurs: An interactive fraud-screening API must respond in under 200 milliseconds, but engineers deploy an Azure ML batch endpoint requiring asynchronous data job submission. The operator must safeguard training reproducibility and serving continuity. What approved correction resolves the failure without modifying unrelated healthy services?
- Deploy the approved model through a suitable managed online endpoint and provision capacity for synchronous requests.
- Fix the image or artifact mount path and ensure init() loads the registered model before promoting the deployment.
- Align the published endpoint contract and client payload mapping, then deploy the reviewed scoring code and test representative requests.
- Write the model into the declared Azure ML job output path and track the versioned artifact used in registration.
Explanation
Correct Answer
A:Deploy the approved model through a suitable managed online endpoint and provision capacity for synchronous requests. This targets online scoring requirement deployed as batch endpoint. Azure ML online endpoints support synchronous prediction while batch endpoints are designed for asynchronous high-volume jobs. Verify the failure through: Compare application latency SLA, current endpoint type, request/response behavior, compute resources and intended inference pattern. Validate success with: A controlled API load test meets latency and correctness targets without requiring clients to poll batch jobs.
Incorrect Answers
B:Fix the image or artifact mount path and ensure init() loads the registered model before promoting the deployment. This could address endpoint deployment health probe fails before model initialization, but it does not solve the specified online scoring requirement deployed as batch endpoint failure. The matching remediation is: Deploy the approved model through a suitable managed online endpoint and provision capacity for synchronous requests.
C:Align the published endpoint contract and client payload mapping, then deploy the reviewed scoring code and test representative requests. This could address endpoint scoring script expects wrong json schema, but it does not solve the specified online scoring requirement deployed as batch endpoint failure. The matching remediation is: Deploy the approved model through a suitable managed online endpoint and provision capacity for synchronous requests.
D:Write the model into the declared Azure ML job output path and track the versioned artifact used in registration. This could address training job does not persist output artifacts, but it does not solve the specified online scoring requirement deployed as batch endpoint failure. The matching remediation is: Deploy the approved model through a suitable managed online endpoint and provision capacity for synchronous requests.
Question 49
While optimizing a production RAG or fine-tuned model service, the following condition occurs: A benefits assistant returns a false eligibility statement because its document chunker splits a policy paragraph away from a nearby exception that changes the rule. The resulting model must meet held-out quality and security criteria. What approved correction resolves the failure without modifying unrelated healthy services?
- Use an appropriate hybrid keyword-plus-vector search strategy with evaluated fusion and optional semantic reranking.
- Tune chunk size and overlap with structure-aware splitting so qualifying rules and exceptions remain retrievable together.
- Retain governed immutable fine-tune artifacts and record deployment lineage with tested restore/rollback procedures.
- Evaluate both retrievers with relevance, factual support and business outcomes under stable randomized cohorts and explicit guardrails.
Explanation
Correct Answer
B:Tune chunk size and overlap with structure-aware splitting so qualifying rules and exceptions remain retrievable together. This targets rag retrieval chunks truncate critical policy exception. Chunk segmentation changes which context is available to the model; cutting a qualification out of its parent rule can produce confidently unsupported answers. Verify the failure through: Examine source document structure, chunk boundaries, overlap size, retrieved context, query examples and groundedness errors. Validate success with: Evaluation questions about eligibility include the correct exception and still return concise supported citations.
Incorrect Answers
A:Use an appropriate hybrid keyword-plus-vector search strategy with evaluated fusion and optional semantic reranking. This could address pure vector search misses exact regulatory clause codes, but it does not solve the specified rag retrieval chunks truncate critical policy exception failure. The matching remediation is: Tune chunk size and overlap with structure-aware splitting so qualifying rules and exceptions remain retrievable together.
C:Retain governed immutable fine-tune artifacts and record deployment lineage with tested restore/rollback procedures. This could address fine-tune production endpoint references deleted training artifact, but it does not solve the specified rag retrieval chunks truncate critical policy exception failure. The matching remediation is: Tune chunk size and overlap with structure-aware splitting so qualifying rules and exceptions remain retrievable together.
D:Evaluate both retrievers with relevance, factual support and business outcomes under stable randomized cohorts and explicit guardrails. This could address rag a/b experiment judges only click-through not factuality, but it does not solve the specified rag retrieval chunks truncate critical policy exception failure. The matching remediation is: Tune chunk size and overlap with structure-aware splitting so qualifying rules and exceptions remain retrievable together.
Question 50
While deploying reproducible Azure Machine Learning infrastructure, the following condition occurs: A model retraining pipeline silently picks up new records every week because its input references an Azure ML data asset's latest version rather than the approved immutable snapshot. The change must remain limited to authorized workspaces and avoid unreviewed production impact. What should the engineer examine first to confirm the specific AI-300 workload issue?
- Check the component's registry name and version, target workspace identity, registry read permissions and referenced environment dependencies.
- Inspect the data asset version resolution, pipeline input expression, snapshot timestamps, storage content hashes and previous run provenance.
- Inspect container image registry URI and tag, Azure ML compute managed identity, ACR pull role, firewall path and worker startup events.
- Inspect workflow trigger, environment configuration, protected branch policies, required reviewers and current deployment audit records.
Explanation
Correct Answer
B:Inspect the data asset version resolution, pipeline input expression, snapshot timestamps, storage content hashes and previous run provenance. This is the most relevant evidence for versionless dataset reference breaks reproducibility. A data asset reference without explicit versioning can change the effective training corpus between runs and invalidate reproducible model comparisons. The matching controlled remediation is: Pin the pipeline input to an approved versioned data asset and advance versions through controlled evaluation gates.
Incorrect Answers
A:Check the component's registry name and version, target workspace identity, registry read permissions and referenced environment dependencies. This would investigate registry model asset not available to second workspace rather than the documented versionless dataset reference breaks reproducibility issue. The relevant investigation is: Inspect the data asset version resolution, pipeline input expression, snapshot timestamps, storage content hashes and previous run provenance.
C:Inspect container image registry URI and tag, Azure ML compute managed identity, ACR pull role, firewall path and worker startup events. This would investigate docker environment image cannot pull from private registry rather than the documented versionless dataset reference breaks reproducibility issue. The relevant investigation is: Inspect the data asset version resolution, pipeline input expression, snapshot timestamps, storage content hashes and previous run provenance.
D:Inspect workflow trigger, environment configuration, protected branch policies, required reviewers and current deployment audit records. This would investigate github deployment pipeline omits secure environment approvals rather than the documented versionless dataset reference breaks reproducibility issue. The relevant investigation is: Inspect the data asset version resolution, pipeline input expression, snapshot timestamps, storage content hashes and previous run provenance.