Microsoft AI-300 Practice Test 35
Question 340
As an MLOps engineer reviews a model training or inference pipeline, the following condition occurs: A withdrawn credit-risk model is marked deprecated in a wiki but remains exposed as an active registry version and continues to be selected by automated deployment tools. The resolution must preserve quality gates, documented artifacts and rollback options. Which investigation best verifies the cause before changing production configuration?
- Inspect run metric definitions, test cohort composition, split timestamps, class prevalence and the comparison dashboard filter.
- Inspect registry model lifecycle status, alias assignments, endpoint references, deployment logs and retirement approvals.
- Examine model registry history, artifact retention rules, endpoint references, promotion record and rollback runbook.
- Inspect registered model metadata, online feature lookup mapping, feature timestamps, key names and schema contract from training.
Explanation
Correct Answer
B:Inspect registry model lifecycle status, alias assignments, endpoint references, deployment logs and retirement approvals. This is the most relevant evidence for retired model kept active as discoverable deployment candidate. Model lifecycle management must reflect state in the model registry and deployment references, not just informal documentation. The matching controlled remediation is: Archive or retire the version through a governed registry lifecycle and remove it from production deployment aliases after dependents migrate. This matters here because as an MLOps engineer reviews a model training or inference pipeline, the following condition occurs: A withdrawn….
Incorrect Answers
A:Inspect run metric definitions, test cohort composition, split timestamps, class prevalence and the comparison dashboard filter. This would investigate run comparison uses incompatible validation populations rather than the documented retired model kept active as discoverable deployment candidate issue. The relevant investigation is: Inspect registry model lifecycle status, alias assignments, endpoint references, deployment logs and retirement approvals. This matters here because as an MLOps engineer reviews a model training or inference pipeline, the following condition occurs: A withdrawn….
C:Examine model registry history, artifact retention rules, endpoint references, promotion record and rollback runbook. This would investigate incident rollback artifact was deleted instead of archived rather than the documented retired model kept active as discoverable deployment candidate issue. The relevant investigation is: Inspect registry model lifecycle status, alias assignments, endpoint references, deployment logs and retirement approvals. This matters here because as an MLOps engineer reviews a model training or inference pipeline, the following condition occurs: A withdrawn….
D:Inspect registered model metadata, online feature lookup mapping, feature timestamps, key names and schema contract from training. This would investigate feature retrieval specification missing from registered model rather than the documented retired model kept active as discoverable deployment candidate issue. The relevant investigation is: Inspect registry model lifecycle status, alias assignments, endpoint references, deployment logs and retirement approvals.
Question 341
During a regulated Azure ML workspace infrastructure review, the following condition occurs: A tabular classification pipeline expects named columns but a newly registered MLTable asset points to CSV files with a different header and type inference rules. The team must preserve approved access, model provenance and private network constraints. Which TWO complementary choices identify the issue and correct it? Select TWO.
- Give the approved compute or workspace identity the necessary ACR pull access and use a validated immutable image reference.
- Rebuild or correct the versioned MLTable asset transformation so the expected columns and types are supplied to downstream training.
- Compare environment image digest, Conda or pip dependencies, Docker build logs, locally resolved versions and job import traceback.
- Require protected pull requests and automated component tests before changes to production training code are merged.
- Inspect MLTable path definitions, transformation instructions, sampled file headers, declared feature schema and job parsing errors.
Explanation
Correct Answers
B:Rebuild or correct the versioned MLTable asset transformation so the expected columns and types are supplied to downstream training. This targets mltable data asset schema mismatches training code. An Azure ML data asset describes dataset structure and access; registration success does not prove its schema matches a consuming model. Verify the failure through: Inspect MLTable path definitions, transformation instructions, sampled file headers, declared feature schema and job parsing errors. Validate success with: A validation job loads the intended features with the declared schema and rejects incompatible input files.
E:Inspect MLTable path definitions, transformation instructions, sampled file headers, declared feature schema and job parsing errors. This is the most relevant evidence for mltable data asset schema mismatches training code. An Azure ML data asset describes dataset structure and access; registration success does not prove its schema matches a consuming model. The matching controlled remediation is: Rebuild or correct the versioned MLTable asset transformation so the expected columns and types are supplied to downstream training.
Incorrect Answers
A:Give the approved compute or workspace identity the necessary ACR pull access and use a validated immutable image reference. This is associated with docker environment image cannot pull from private registry, not the essential evidence-and-fix pair for mltable data asset schema mismatches training code. First examine: Inspect MLTable path definitions, transformation instructions, sampled file headers, declared feature schema and job parsing errors. Then correct: Rebuild or correct the versioned MLTable asset transformation so the expected columns and types are supplied to downstream training.
C:Compare environment image digest, Conda or pip dependencies, Docker build logs, locally resolved versions and job import traceback. This is associated with training environment lacks pinned package versions, not the essential evidence-and-fix pair for mltable data asset schema mismatches training code. First examine: Inspect MLTable path definitions, transformation instructions, sampled file headers, declared feature schema and job parsing errors. Then correct: Rebuild or correct the versioned MLTable asset transformation so the expected columns and types are supplied to downstream training.
D:Require protected pull requests and automated component tests before changes to production training code are merged. This is associated with notebook merge bypasses protected evaluation tests, not the essential evidence-and-fix pair for mltable data asset schema mismatches training code. First examine: Inspect MLTable path definitions, transformation instructions, sampled file headers, declared feature schema and job parsing errors. Then correct: Rebuild or correct the versioned MLTable asset transformation so the expected columns and types are supplied to downstream training.
Question 342
During a Microsoft Foundry generative AI deployment review, the following condition occurs: A legal-document summarizer routinely receives 110,000-token inputs, but the selected model deployment has an insufficient supported context window and rejects or truncates requests. The deployment must retain governed identities, endpoint contracts and auditable release history. Which diagnostic evidence most directly distinguishes the described AI operations failure?
- Inspect Foundry project and resource regions, model deployment regional availability, enterprise residency constraints and API error codes.
- Compare real tokenized input sizes, selected model version limits, deployment configuration and representative failure logs.
- Inspect deployed model modalities, model card capabilities, payload content types, supported API version and failed inference responses.
- Check application response SLA, workload concurrency, chosen Foundry model endpoint type and current request queuing behavior.
Explanation
Correct Answer
B:Compare real tokenized input sizes, selected model version limits, deployment configuration and representative failure logs. This is the most relevant evidence for foundation model selected without required context window. Model suitability depends on context limits and task capability, not just a broad 'generative model' label. The matching controlled remediation is: Choose an appropriate available model and supported prompt chunking strategy that meets the workload's context and cost requirements.
Incorrect Answers
A:Inspect Foundry project and resource regions, model deployment regional availability, enterprise residency constraints and API error codes. This would investigate foundry project created in mismatched resource location rather than the documented foundation model selected without required context window issue. The relevant investigation is: Compare real tokenized input sizes, selected model version limits, deployment configuration and representative failure logs.
C:Inspect deployed model modalities, model card capabilities, payload content types, supported API version and failed inference responses. This would investigate vision input sent to text-only foundation model rather than the documented foundation model selected without required context window issue. The relevant investigation is: Compare real tokenized input sizes, selected model version limits, deployment configuration and representative failure logs.
D:Check application response SLA, workload concurrency, chosen Foundry model endpoint type and current request queuing behavior. This would investigate interactive inference deployed on unsuitable batch-style compute rather than the documented foundation model selected without required context window issue. The relevant investigation is: Compare real tokenized input sizes, selected model version limits, deployment configuration and representative failure logs.
Question 343
During a Foundry generative AI quality and observability incident, the following condition occurs: A team upgrades a custom evaluator prompt and compares this month's agent scores to last month's, even though the grading rubric and baseline model changed. The fix must preserve evaluation validity and avoid unnecessary user-data collection. Which investigation best verifies the cause before changing production configuration?
- Review evaluator code and prompt versions, rubric labels, baseline sample scores and published metric time series.
- Examine end-to-end request traces, retrieval timing, model time to first token, generation duration, request concurrency and percentiles.
- Inspect evaluator definitions, retrieved source contexts, response citations, groundedness scoring and sampled human reference judgments.
- Inspect evaluation workflow outputs, release conditions, metric thresholds, CI job failure behavior and promotion records.
Explanation
Correct Answer
A:Review evaluator code and prompt versions, rubric labels, baseline sample scores and published metric time series. This is the most relevant evidence for custom evaluator silently changes its scoring rubric. Evaluation metrics are outputs of versioned assessment logic; changing a judge or rubric can change scores without any underlying agent change. The matching controlled remediation is: Version the custom evaluator and recalibrate on a fixed reference set before using its scores for cross-version comparison.
Incorrect Answers
B:Examine end-to-end request traces, retrieval timing, model time to first token, generation duration, request concurrency and percentiles. This would investigate generation latency spike blamed on retrieval without breakdown rather than the documented custom evaluator silently changes its scoring rubric issue. The relevant investigation is: Review evaluator code and prompt versions, rubric labels, baseline sample scores and published metric time series.
C:Inspect evaluator definitions, retrieved source contexts, response citations, groundedness scoring and sampled human reference judgments. This would investigate groundedness metric conflated with response fluency rather than the documented custom evaluator silently changes its scoring rubric issue. The relevant investigation is: Review evaluator code and prompt versions, rubric labels, baseline sample scores and published metric time series.
D:Inspect evaluation workflow outputs, release conditions, metric thresholds, CI job failure behavior and promotion records. This would investigate evaluation workflow fails to block degraded agent deployment rather than the documented custom evaluator silently changes its scoring rubric issue. The relevant investigation is: Review evaluator code and prompt versions, rubric labels, baseline sample scores and published metric time series.
Question 344
During a retrieval or model-customization performance review, the following condition occurs: An operations chatbot answers 'I cannot find it' to precise internal procedure questions because its vector search similarity threshold was raised above scores returned by the relevant documents. The change must keep citations, privacy and previously approved behavior intact. Which diagnostic evidence most directly distinguishes the described AI operations failure?
- Inspect registered fine-tuned model version, retained checkpoints, base model lineage, tokenizer files, serving configuration and retention policy.
- Review A/B cohort assignment, groundedness samples, relevance labels, business success measures and unsupported-claim incidents.
- Examine source document structure, chunk boundaries, overlap size, retrieved context, query examples and groundedness errors.
- Inspect top-k search scores, threshold, query normalization, relevant passages and test question recall.
Explanation
Correct Answer
D:Inspect top-k search scores, threshold, query normalization, relevant passages and test question recall. This is the most relevant evidence for rag similarity threshold filters all relevant evidence. Vector similarity thresholds control recall and precision; a threshold chosen without validation can suppress even the best available evidence. The matching controlled remediation is: Calibrate retrieval threshold and ranking settings using labeled queries to admit relevant context without overwhelming the model with noise.
Incorrect Answers
A:Inspect registered fine-tuned model version, retained checkpoints, base model lineage, tokenizer files, serving configuration and retention policy. This would investigate fine-tune production endpoint references deleted training artifact rather than the documented rag similarity threshold filters all relevant evidence issue. The relevant investigation is: Inspect top-k search scores, threshold, query normalization, relevant passages and test question recall.
B: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 rag similarity threshold filters all relevant evidence issue. The relevant investigation is: Inspect top-k search scores, threshold, query normalization, relevant passages and test question recall.
C:Examine source document structure, chunk boundaries, overlap size, retrieved context, query examples and groundedness errors. This would investigate rag retrieval chunks truncate critical policy exception rather than the documented rag similarity threshold filters all relevant evidence issue. The relevant investigation is: Inspect top-k search scores, threshold, query normalization, relevant passages and test question recall.
Question 345
During a production Azure Machine Learning model lifecycle incident, the following condition occurs: A notebook computes features from a mutable SQL export every time it runs, but the submitted model job cannot reproduce yesterday's feature distribution. The operator must safeguard training reproducibility and serving continuity. Which observable postchange result proves the intended remediation succeeded?
- Authorized auditors can assess model behavior while unauthorized team members cannot recover individual identifiers.
- A reviewer compares both runs and identifies the exact code, data, hyperparameters and metrics used without reconstructing them by hand.
- A historical notebook rerun uses the recorded snapshot and yields traceable feature statistics.
- Holdout performance is credible against production-like inputs and candidate selection no longer uses post-outcome fields.
Explanation
Correct Answer
C:A historical notebook rerun uses the recorded snapshot and yields traceable feature statistics. Notebook experimentation is flexible but reliable model comparison requires stable, recorded data inputs rather than live mutable queries. a direct corrective action is to capture a governed immutable dataset or versioned data asset and reference that exact version in notebook and scheduled jobs.
Incorrect Answers
A:Authorized auditors can assess model behavior while unauthorized team members cannot recover individual identifiers. That addresses a different configuration issue rather than the main requirement. Verify instead: A historical notebook rerun uses the recorded snapshot and yields traceable feature statistics.
B:A reviewer compares both runs and identifies the exact code, data, hyperparameters and metrics used without reconstructing them by hand. That addresses a different configuration issue rather than the main requirement. Verify instead: A historical notebook rerun uses the recorded snapshot and yields traceable feature statistics.
D:Holdout performance is credible against production-like inputs and candidate selection no longer uses post-outcome fields. That addresses a different configuration issue rather than the main requirement. Verify instead: A historical notebook rerun uses the recorded snapshot and yields traceable feature statistics.
Question 346
During a regulated Azure ML workspace infrastructure review, the following condition occurs: A central model team registers a reusable component in a shared Azure ML registry, but application workspace users cannot resolve the component because registry access was never granted. The team must preserve approved access, model provenance and private network constraints. What measured outcome validates the solution while preserving the stated constraints?
- The same versioned component runs against authorized data in two workspaces without editing its source implementation.
- A validation job loads the intended features with the declared schema and rejects incompatible input files.
- Authorized workspaces run the shared component while unauthorized subscriptions remain blocked from the registry.
- A reviewed release can deploy production resources, but an unreviewed branch workflow cannot obtain production deployment rights.
Explanation
Correct Answer
C:Authorized workspaces run the shared component while unauthorized subscriptions remain blocked from the registry. Cross-workspace registries support reusing vetted assets but do not automatically authorize every workspace or resolve unknown versions. a direct corrective action is to grant controlled registry consumption rights and use the fully qualified versioned registry asset reference in the target workspace.
Incorrect Answers
A:The same versioned component runs against authorized data in two workspaces without editing its source implementation. That addresses a different configuration issue rather than the main requirement. Verify instead: Authorized workspaces run the shared component while unauthorized subscriptions remain blocked from the registry.
B:A validation job loads the intended features with the declared schema and rejects incompatible input files. That addresses a different configuration issue rather than the main requirement. Verify instead: Authorized workspaces run the shared component while unauthorized subscriptions remain blocked from the registry.
D:A reviewed release can deploy production resources, but an unreviewed branch workflow cannot obtain production deployment rights. That addresses a different configuration issue rather than the main requirement. Verify instead: Authorized workspaces run the shared component while unauthorized subscriptions remain blocked from the registry.
Question 347
While preparing a secure Foundry model or agent release, the following condition occurs: A project deployment script succeeds in creating resources but fails when attempting model deployment because the selected subscription's capacity quota is insufficient. The remediation must meet enterprise access and reliability requirements without a broad bypass. Which diagnostic evidence most directly distinguishes the described AI operations failure?
- Examine model deployment quota for selected region/type, Azure CLI errors, requested capacity units and existing model allocations.
- Compare project connection configuration, model deployment resource ID, endpoint URL, deployment name and invocation authorization.
- Inspect input/output token volume, peak concurrent requests, provisioned throughput allocation, throttling codes and deployment latency percentiles.
- Inspect IaC parameter history, deployment logs, credential usage, Key Vault references and access audit events.
Explanation
Correct Answer
A:Examine model deployment quota for selected region/type, Azure CLI errors, requested capacity units and existing model allocations. This is the most relevant evidence for azure cli foundry setup assumes existing model quota. Resource provisioning and model capacity allocation are separate operations; successful project creation does not reserve throughput for every selected model. The matching controlled remediation is: Validate quota in a preflight deployment stage and reduce or obtain approved capacity before creating the intended deployment. This matters here because while preparing a secure Foundry model or agent release, the following condition occurs: A project deployment script….
Incorrect Answers
B:Compare project connection configuration, model deployment resource ID, endpoint URL, deployment name and invocation authorization. This would investigate foundry project links to wrong deployed model resource rather than the documented azure cli foundry setup assumes existing model quota issue. The relevant investigation is: Examine model deployment quota for selected region/type, Azure CLI errors, requested capacity units and existing model allocations.
C:Inspect input/output token volume, peak concurrent requests, provisioned throughput allocation, throttling codes and deployment latency percentiles. This would investigate provisioned throughput capacity insufficient at peak traffic rather than the documented azure cli foundry setup assumes existing model quota issue. The relevant investigation is: Examine model deployment quota for selected region/type, Azure CLI errors, requested capacity units and existing model allocations.
D:Inspect IaC parameter history, deployment logs, credential usage, Key Vault references and access audit events. This would investigate foundry infrastructure parameters embed a secret in source repository rather than the documented azure cli foundry setup assumes existing model quota issue. The relevant investigation is: Examine model deployment quota for selected region/type, Azure CLI errors, requested capacity units and existing model allocations.
Question 348
While deploying reproducible Azure Machine Learning infrastructure, the following condition occurs: An Azure ML training job authenticates with a managed identity but fails to download input data because it has workspace rights without Storage Blob Data Reader on the data container. The change must remain limited to authorized workspaces and avoid unreviewed production impact. What approved correction resolves the failure without modifying unrelated healthy services?
- Assign the minimum permitted storage data access role to the job identity at the relevant container scope without granting secret-key access.
- Recreate or provision the workspace in the approved subscription and region through the reviewed deployment pipeline, preserving authorized datasets separately.
- Link the authorized Key Vault or repair supported access to the intended secret store through the correct workspace deployment configuration.
- Bind every training job and model registration to a reviewed Git commit SHA and immutable environment/data version in the workflow.
Explanation
Correct Answer
A:Assign the minimum permitted storage data access role to the job identity at the relevant container scope without granting secret-key access. This targets job identity lacks blob reader data-plane rights. Workspace RBAC and Azure Storage data-plane roles are distinct; successful workspace authentication does not authorize every blob read. Verify the failure through: Check job identity selection, storage data-plane authorization, private endpoint reachability and storage diagnostic access-denied records. Validate success with: The job reads the required training blobs and cannot write or list unrelated protected containers. This matters here because while deploying reproducible Azure Machine Learning infrastructure, the following condition occurs: An Azure ML….
Incorrect Answers
B:Recreate or provision the workspace in the approved subscription and region through the reviewed deployment pipeline, preserving authorized datasets separately. This could address azure ml workspace deployed in the wrong subscription, but it does not solve the specified job identity lacks blob reader data-plane rights failure. The matching remediation is: Assign the minimum permitted storage data access role to the job identity at the relevant container scope without granting secret-key access.
C:Link the authorized Key Vault or repair supported access to the intended secret store through the correct workspace deployment configuration. This could address workspace dependencies point to unrelated secrets vault, but it does not solve the specified job identity lacks blob reader data-plane rights failure. The matching remediation is: Assign the minimum permitted storage data access role to the job identity at the relevant container scope without granting secret-key access.
D:Bind every training job and model registration to a reviewed Git commit SHA and immutable environment/data version in the workflow. This could address production training script commit cannot be traced, but it does not solve the specified job identity lacks blob reader data-plane rights failure. The matching remediation is: Assign the minimum permitted storage data access role to the job identity at the relevant container scope without granting secret-key access.
Question 349
During a Microsoft Foundry generative AI deployment review, the following condition occurs: A chat assistant must return low-latency answers but the team selects an offline batch inference architecture that produces responses only after a scheduled job completes. The deployment must retain governed identities, endpoint contracts and auditable release history. What measured outcome validates the solution while preserving the stated constraints?
- A production-sized document completes evaluation without unintended context loss or unsupported token-limit errors.
- Prompt tests retain representative behavior without exposing customer identities in accessible Git history or future deployments.
- Representative chat queries receive valid answers within the approved latency target without waiting for a batch output artifact.
- The last stable response and tool-call contracts are recovered across representative transactions without rebuilding application data.
Explanation
Correct Answer
C:Representative chat queries receive valid answers within the approved latency target without waiting for a batch output artifact. Generative chat workloads ordinarily require online serving semantics; scheduled batch compute cannot satisfy an interactive response deadline. a direct corrective action is to deploy an available supported online/serverless or managed model endpoint sized for synchronous conversational requests. This matters here because a chat assistant must return low-latency answers but the team selects an offline batch inference architecture that….
Incorrect Answers
A:A production-sized document completes evaluation without unintended context loss or unsupported token-limit errors. That addresses a different configuration issue rather than the main requirement. Verify instead: Representative chat queries receive valid answers within the approved latency target without waiting for a batch output artifact. This matters here because a chat assistant must return low-latency answers but the team selects an offline batch inference architecture that….
B:Prompt tests retain representative behavior without exposing customer identities in accessible Git history or future deployments. That addresses a different configuration issue rather than the main requirement. Verify instead: Representative chat queries receive valid answers within the approved latency target without waiting for a batch output artifact.
D:The last stable response and tool-call contracts are recovered across representative transactions without rebuilding application data. That addresses a different configuration issue rather than the main requirement. Verify instead: Representative chat queries receive valid answers within the approved latency target without waiting for a batch output artifact. This matters here because a chat assistant must return low-latency answers but the team selects an offline batch inference architecture that….
Question 350
As an MLOps engineer reviews a model training or inference pipeline, the following condition occurs: A retail demand model triggers constant drift alerts because its production Christmas data is compared only against a summer reference dataset without accounting for known seasonal regimes. The resolution must preserve quality gates, documented artifacts and rollback options. Which production-safe remediation should the engineer apply to this AI workflow?
- Monitor available proxy signals immediately and compute labeled performance only after verified outcomes arrive through a governed join.
- Configure a suitable seasonal baseline or context-aware drift monitor while retaining meaningful anomaly thresholds.
- Allow the narrowly scoped supported compute-to-compute communication paths and correct the distributed training rendezvous configuration.
- Configure tail-latency and error-rate dashboards with capacity remedies appropriate to the measured bottleneck.
Explanation
Correct Answer
B:Configure a suitable seasonal baseline or context-aware drift monitor while retaining meaningful anomaly thresholds. This targets data drift monitor uses wrong reference season. Data drift measurement depends on a chosen reference distribution; a statistically different season is not automatically a harmful model change. Verify the failure through: Inspect reference data time window, drift feature metrics, alert threshold, seasonal segmentation and actual model outcomes. Validate success with: Normal seasonal variation stops generating false incidents and genuine distribution changes still cross documented alert criteria. This matters here because as an MLOps engineer reviews a model training or inference pipeline, the following condition occurs: A retail demand….
Incorrect Answers
A:Monitor available proxy signals immediately and compute labeled performance only after verified outcomes arrive through a governed join. This could address accuracy monitoring lacks delayed outcome labels, but it does not solve the specified data drift monitor uses wrong reference season failure. The matching remediation is: Configure a suitable seasonal baseline or context-aware drift monitor while retaining meaningful anomaly thresholds.
C:Allow the narrowly scoped supported compute-to-compute communication paths and correct the distributed training rendezvous configuration. This could address distributed job workers cannot communicate over required network path, but it does not solve the specified data drift monitor uses wrong reference season failure. The matching remediation is: Configure a suitable seasonal baseline or context-aware drift monitor while retaining meaningful anomaly thresholds.
D:Configure tail-latency and error-rate dashboards with capacity remedies appropriate to the measured bottleneck. This could address endpoint latency average conceals tail degradation, but it does not solve the specified data drift monitor uses wrong reference season failure. The matching remediation is: Configure a suitable seasonal baseline or context-aware drift monitor while retaining meaningful anomaly thresholds. This matters here because as an MLOps engineer reviews a model training or inference pipeline, the following condition occurs: A retail demand….