Free AI-300 practice test: 30 questions with answers

MLflow tracking records parameters, metrics, and artifacts for each run, which allows comparison across jobs. Manual spreadsheets are error-prone, and wallpapers and deletion do not record experiment results.

A registered, versioned model can be promoted unchanged and traced back to the run that produced it. Registration does not replace evaluation or guarantee accuracy, and deleting old versions removes the ability to roll back.

Automated machine learning searches algorithms and settings and ranks candidate models against a metric. Hand coding, unevaluated copies, and oversized compute do not systematically find a good model.

A sweep over a defined search space, optimizing a primary metric, systematically finds good hyperparameters. Random unrecorded changes, defaults only, and deleting the metric remove the basis for choosing.

Drift detection compares production data with the training distribution, which signals when the model may become unreliable. Prediction counts, names, and anecdotes do not measure distribution change.

Groundedness measures whether the answer's claims are supported by the provided context. Disk rates, CPU, and deployment counts are infrastructure measures and do not assess support.

Token consumption per feature and model over time tracks the main driver of model spend and shows which feature costs the most. Laptop counts, logo views, and file counts do not reflect running costs.

Risk and safety evaluations check outputs for harmful content categories and reveal gaps before release. Visual, database load, and comment reviews do not assess content safety.

Distributed tracing shows each call and its timing within one request, which locates slow steps. Monthly totals, screenshots, and deployment counts do not show the path of an individual request.

Relevance measures whether the response addresses the question asked. Disk latency, storage tiers, and prompt length are not measures of answer quality.

Chunk size directly affects retrieval precision, and measured relevance shows whether the change helps. Longer outputs, removing retrieval, and more users do not address mixed-topic chunks.

Synthetic examples can expand scarce data when they are validated by experts to avoid errors in the training set. Unlabeled random text, unreviewed copies of test data, and skipped preparation harm the model.

Hybrid search covers exact codes through keywords and descriptive questions through semantic matching. Using only one approach misses the other kind of query, and random sampling returns no reliable matches.

Raising the similarity threshold keeps only strong matches, which reduces dilution. Lowering it adds more loose matches, removing retrieval loses grounding, and longer outputs do not improve relevance.

An A/B test with predefined metrics and split traffic measures real impact under comparable conditions. One user, a single handpicked question, and version numbers do not show which strategy performs better.

A managed identity with a scoped role assignment gives the compute access without storing secrets. Pasted keys, long-lived SAS tokens, and public containers all expose data or credentials.

Registries let assets such as environments and components be shared across workspaces and regions. Screenshots, manual recreation, and renaming do not share the definition.

Bicep templates applied through the CLI or a pipeline make environments repeatable and reviewable. Manual clicks, screenshots, and ad hoc creation cause drift between environments.

GitHub Actions triggers the pipeline on repository events, which automates the workflow. Reminders, manual copies, and printouts depend on people acting consistently.

Private endpoints and restricted public access limit who can reach the workspace over the network. Open ports, shared public addresses, and disabled authentication expand access.

Version control gives each prompt a history and lets results be tied to a specific version. Unrecorded production edits, chat threads, and desktop files lose that link.

Provisioned throughput reserves capacity so that high-volume workloads meet latency and throughput needs. Pay-per-call endpoints do not reserve capacity, and shared keys or user limits do not address throughput.

Running both variants on the same dataset and comparing metrics gives an evidence-based choice. Length, seniority, and recency are not measures of quality.

Selecting by measured quality on the actual task balances cost and latency against the required accuracy. The largest model by default, name-based choice, and skipping evaluation do not justify the selection.

A gradual traffic split and a retained previous deployment allow the team to detect problems and roll back quickly. One-step replacement, disabled monitoring, and unchecked global rollouts remove those safeguards.

Batch endpoints process large files efficiently as jobs, which suits nightly volume. Per-record real-time calls and manual notebooks are slower and do not scale to that volume.

Shifting traffic back and redeploying a recorded version from the registry restore a known good state quickly. Rebuilding from memory, hiding the regression, and disabling monitoring each delay recovery or conceal the problem.

Deployment logs and startup output show errors in the scoring code, dependencies, or environment. Dashboard colors, user counts, and unrelated equipment do not explain server errors.

A drift monitor compares feature statistics with a reference and alerts when a threshold is crossed. Manual checks are slow, printed names give no alert, and features do change over time.

A sustained quality decline and persistent input drift both point to a model that no longer fits current data. A single corrected prediction, naming style, and planned restarts are not signs of degradation.