OCI AI Portfolio practice questions

Worked examples for OCI AI Portfolio

1. Question 1

   Which statement about compute selection in OCI Data Science is correct?

   1. A. Data Science runs exclusively on CPU shapes; GPU work must be moved to a separate self-managed Compute instance

      The 'Data Science cannot use GPUs' misconception. GPU shapes are selectable for Data Science resources such as notebook sessions, jobs, and deployments.

   2. B. The compute shape is fixed by the model catalog once a model is saved and cannot differ between training and serving

      Confuses artifact storage with resource sizing. The catalog holds artifacts and metadata; shape is chosen per resource, so training and deployment can use different shapes.

   3. C. You select a compute shape — CPU or GPU — for Data Science resources such as notebook sessions, jobs, and model deployments, sizing it to the workloadCorrect answer

      Data Science resources are backed by OCI Compute shapes chosen at creation, including GPU shapes for accelerated training or inference, letting each resource be sized independently.

   4. D. GPU shapes are available only for inference in Data Science, never for training

      The 'GPUs are inference-only' misconception. Accelerated compute is commonly used for training in notebook sessions and jobs as well as for serving.

   Explanation

   OCI Data Science resources run on OCI Compute shapes that you choose when creating each resource, and GPU shapes are available for the accelerated training and inference that deep-learning workloads need. Because the choice is per resource, a job can train on one shape while a deployment serves on another; the model catalog stores artifacts and metadata and does not pin compute. Claiming Data Science is CPU-only, or that GPUs there serve inference alone, both contradict the platform's use of the underlying accelerated infrastructure.

2. Question 2

   A team has decided the pretrained model in an OCI AI Service does not recognize the domain-specific categories their business uses, so they plan to train a custom model in that service. Which TWO statements about custom models in OCI AI Services are accurate? (Select all that apply.)

   1. A. Training a custom model permanently replaces the pretrained model for that tenancy, so the built-in capability is no longer callable.

      Assumes customization is a swap rather than an addition. The pretrained model stays available; a custom model is an additional model your requests can target.

   2. B. A custom model is trained on your own labeled data so the service can recognize entities, classes, or document types the pretrained model does not cover.Correct answer

      Correct. Custom training exists precisely to adapt a service to business-specific vocabulary, image classes, or document layouts using examples you label.

   3. C. The trained custom model must be exported and hosted on your own compute instances before it can serve predictions.

      Assumes you must operate the serving infrastructure. The service manages training and hosting; you invoke the model through the same managed endpoint.

   4. D. Custom training is mandatory before any request to these services will succeed, because the pretrained models are only samples.

      The 'training is always required' misconception. Pretrained models are fully usable in production immediately; custom training is an optional path for domain gaps.

   5. E. Once trained, the custom model is invoked through the same service API used for the pretrained model, by referencing the custom model.Correct answer

      Correct. Customization does not change the integration pattern — the application keeps calling the same service and simply points the request at the custom model.

   Explanation

   Custom models in OCI AI Services are an optional extension of the prebuilt capability: you supply labeled examples covering your own entities, classes, or document types, the service trains and hosts the model, and your application targets it through the same API it already uses. Customization neither disables the pretrained model, nor requires you to export and self-host anything, nor becomes a precondition for using the service, since the pretrained models are production-ready on their own.

3. Question 3

   Which TWO practices reflect correct use of the OCI Data Science model catalog in a team's ML workflow? (Choose two.)

   1. A. Save each trained model with metadata describing its provenance and input/output schema so results can be reproduced and reviewed laterCorrect answer

      Capturing provenance and schema metadata alongside the artifact is the catalog's core purpose, supporting reproducibility, review, and correct invocation after deployment.

   2. B. Use the catalog as the shared hand-off point so a model built in one notebook session can be deployed or reused by other team membersCorrect answer

      The catalog is tenancy-scoped and access-controlled, making it the intended place to share models across sessions and teammates rather than passing files informally.

   3. C. Rely on the catalog to automatically retrain each stored model whenever new data lands in Object Storage

      Attributes training orchestration to a repository. Retraining is triggered by a job or pipeline; the catalog only stores artifacts and metadata.

   4. D. Keep models only on the notebook session's block storage, since a session's disk is the permanent system of record

      Misconception that a notebook session is durable shared storage. Session storage is tied to that session and is not the shared, governed repository the catalog provides.

   5. E. Store the catalog entry inside an OCI AI Service project so the prebuilt API serves the custom model

      The prebuilt-versus-custom conflation. AI Services serve Oracle's pretrained models for fixed tasks; a catalog model is served through Data Science model deployment.

   Explanation

   The model catalog is the managed, access-controlled repository where Data Science models are saved with metadata — provenance, taxonomy, and input/output schemas — which makes runs reproducible and lets teammates find, review, and deploy a model built elsewhere. It is a repository, not an orchestrator, so it does not trigger retraining on new data; leaving models only on a notebook session's disk forfeits durability and sharing; and a custom model is served by Data Science model deployment rather than by a prebuilt AI Service's fixed-task API.

4. Question 4

   What is the primary purpose of the model catalog in OCI Data Science?

   1. A. It is a marketplace where Oracle publishes pretrained models that replace the need to train your own

      Confuses the catalog with a public model marketplace. The catalog is a private, tenancy-scoped repository for models your team produces, not an Oracle-published storefront.

   2. B. It provisions and autoscales the GPU compute used to train models

      Confuses the artifact repository with compute provisioning. Compute shapes are chosen for notebook sessions, jobs, and deployments; the catalog only stores artifacts and metadata.

   3. C. It is a data catalog that profiles and governs the training datasets used by a project

      Confuses model catalog with data cataloging/governance. The model catalog tracks models, not dataset discovery or lineage over source data.

   4. D. It stores model artifacts together with their metadata so models can be versioned, shared across a team, and later deployedCorrect answer

      The model catalog is the centralized, managed repository of model artifacts plus metadata (provenance, taxonomy, schemas), enabling reproducibility, sharing, and deployment from a saved model.

   Explanation

   In the OCI Data Science lifecycle, a trained model is saved to the model catalog, which persists the serialized artifact along with metadata that captures provenance and input/output schemas. That makes models reproducible, discoverable by teammates, and directly deployable as endpoints later. It is not a public store of Oracle-pretrained models, it does not provision or scale compute (shapes are chosen per notebook session, job, or deployment), and it is not a dataset governance catalog.

5. Question 5

   A data scientist finishes training a model in an OCI Data Science notebook session and saves it to the model catalog. A colleague asks what the catalog entry actually gives the team. What is the primary purpose of the OCI Data Science model catalog?

   1. A. It is a live serving tier: saving a model to the catalog immediately publishes an inference endpoint that applications can call.

      Conflates cataloguing with deployment. Storing an artifact and serving it are distinct steps — a cataloged model must be explicitly deployed before an HTTP endpoint exists.

   2. B. It is a repository of Oracle-supplied pretrained models that customers browse and call, equivalent to the prebuilt AI Services catalog.

      Confuses a customer-owned artifact store with a vendor model marketplace. The catalog holds the models your own team creates and saves, not a shelf of Oracle pretrained models.

   3. C. It is a centralized, managed repository for model artifacts and their associated metadata, so models can be versioned, shared across a team, audited for provenance, and later deployed.Correct answer

      Correct. The model catalog stores model artifacts together with metadata such as provenance and taxonomy, giving teams a shared, governed place to preserve, discover, and reuse models before deployment.

   4. D. It is a feature store that persists the engineered input features so downstream training runs can reuse consistent feature definitions.

      Mistakes the model catalog for feature management. The catalog's unit of storage is the trained model artifact and its metadata, not engineered feature data.

   Explanation

   The model catalog in OCI Data Science is a managed repository where completed model artifacts are preserved along with metadata — provenance, taxonomy, and other descriptive attributes — so a team can version, discover, share, and audit models rather than leaving them scattered in individual notebook sessions. Cataloguing is deliberately separate from serving: a stored model becomes callable only when it is deployed as a model deployment endpoint. It is not a shelf of Oracle-supplied pretrained models, which is what the prebuilt AI Services provide through their own APIs, and it stores models rather than the engineered features used to train them.

6. Question 6

   Which TWO statements correctly describe how applications consume OCI AI Services? (Choose two.)

   1. A. They are invoked through REST APIs, SDKs and the CLI, so the capability can be embedded directly into an application.Correct answer

      OCI AI Services are described as API-driven services accessible through REST APIs, SDKs, the CLI and the Console.

   2. B. The underlying model serving infrastructure is managed by Oracle, so the caller does not provision or scale compute for inference.Correct answer

      AI Services are fully managed: Oracle operates the models and serving infrastructure, and the consumer only makes service calls.

   3. C. Each call requires the customer to first stand up a model deployment endpoint in OCI Data Science.

      The 'Data Science is a prerequisite' misconception. AI Services have their own managed endpoints and do not depend on a Data Science model deployment.

   4. D. They can only be used interactively from the OCI Console and offer no programmatic interface.

      The 'console-only' misconception. Programmatic access through APIs and SDKs is the primary consumption model for AI Services.

   Explanation

   OCI AI Services are prebuilt, fully managed services that developers consume programmatically through REST APIs, SDKs and the CLI, with Oracle operating the models and inference infrastructure so no compute provisioning is needed. They neither require a Data Science model deployment as a prerequisite nor restrict usage to interactive Console work. The API-driven, managed nature is precisely what lets teams without ML expertise adopt them.

7. Question 7

   A team must run the same model-training script every night on a schedule, unattended, with a compute shape sized for the job — no interactive editing required. Which OCI Data Science capability best matches this requirement?

   1. A. Keep a notebook session running permanently and rely on someone executing the cells each night

      Misuses the interactive environment for batch automation; notebook sessions are for development and this approach is neither unattended nor repeatable.

   2. B. Store the script in the model catalog, which executes any script it holds on a schedule

      Confuses the artifact repository with an execution engine. The catalog stores model artifacts and metadata and runs nothing.

   3. C. Data Science jobs, which run a defined task such as a training script on managed infrastructure as a repeatable, unattended workloadCorrect answer

      Jobs are the Data Science feature for repeatable, unattended execution of a task (for example a training script) on a chosen compute shape, independent of any interactive session.

   4. D. A model deployment, since deployments re-train the model on each scheduled invocation

      Confuses serving with training. A deployment answers inference requests against an already-trained model; it does not perform training runs.

   Explanation

   OCI Data Science jobs let a team define a task — commonly a training or data-processing script — and run it on managed infrastructure repeatably and unattended, which is exactly what a nightly training run needs. An always-on notebook session driven by hand is interactive development rather than automation, the model catalog stores artifacts and metadata but executes nothing, and a model deployment serves predictions from an existing model rather than training a new one.

8. Question 8

   A finance team wants to detect unusual spikes and dips in daily transaction volumes across several correlated business metrics, without writing statistical models themselves. Which characterisation of OCI Anomaly Detection is accurate?

   1. A. It only works on a single univariate series at a time and cannot consider relationships between signals.

      Understates the service. Anomaly Detection is designed for multivariate data and accounts for correlations across signals, not just one series in isolation.

   2. B. It classifies free-text descriptions of transactions to decide which ones are suspicious.

      Confuses Anomaly Detection with Language's text classification. Anomaly Detection consumes numeric signals, not free-text narratives.

   3. C. It is a prebuilt service that identifies outliers in time-series and multivariate business or sensor data, and it can be trained on the customer's own historical data to learn what normal looks like.Correct answer

      This matches the documented role of Anomaly Detection: a prebuilt AI Service for outlier detection in time-series/multivariate data, with models trained on the customer's own data to define normal behaviour.

   4. D. It requires no data from the customer at all, because a single global pretrained model already knows every business's normal behaviour.

      The 'universal pretrained model' misconception. Normal behaviour is business-specific, so the service learns from the customer's own historical data.

   Explanation

   OCI Anomaly Detection is a prebuilt AI Service for spotting outliers in time-series and multivariate data such as business metrics and sensor telemetry, and it learns a customer-specific notion of normal from that customer's historical data. It is not limited to a single univariate series, it does not classify free text — that is Language's domain — and no global model could define normal for every business, which is why training on the customer's own data matters. Matching numeric, correlated, time-ordered input to the service is the key skill.