Overview
Domain 3 is about getting the right model ready to serve requests in OCI Enterprise AI, which runs on the OCI Generative AI service. You need to know the access modes: on-demand pretrained models called by model ID and billed by usage, dedicated AI clusters that reserve GPU capacity for your workloads, and imported models brought from Hugging Face or Object Storage and hosted on a dedicated cluster.
The section covers the OpenAI-compatible endpoint, where the model value can be an on-demand model ID, a dedicated endpoint OCID or a routing profile OCID. It then walks through fine-tuning: preparing a JSONL dataset of prompt and completion pairs, creating a fine-tuning cluster, training a custom model, reading accuracy and loss, and serving the result from a hosting cluster endpoint. Region availability, model retirement dates and IAM rules that limit which models a group can use also appear here.
How it's tested
Model questions present a workload with requirements on isolation, throughput, customisation or cost and ask which access mode or configuration fits. Others give a fine-tuning sequence or an API call and ask what is missing or in the wrong order.
-
Access mode fit
Steady high-volume traffic, data isolation or an open-weight model points to a dedicated AI cluster; light or bursty use points to on-demand.
-
Fine-tuning sequence
You order dataset upload, fine-tuning cluster, custom model, hosting cluster and endpoint, or spot the missing step.
-
Model value in a call
You choose between a model ID, an endpoint OCID and a routing profile OCID for the model field.
A team imports an open-weight model from Hugging Face and tries to call it by name through the on-demand endpoint, which fails. Imported models are not served on demand. They need a hosting dedicated AI cluster and an endpoint, and the call then passes the endpoint OCID as the model value. Requesting a fine-tuning cluster would not help, because OCI does not fine-tune imported models.
Topics 5
Official topics from the Oracle outline, in outline order.
3.1 Select OCI Enterprise AI models and access modesOn-demand, dedicated and imported model options.
- What it means
- OCI Enterprise AI offers pretrained chat and embedding models from several providers, including Cohere, Meta, xAI and OpenAI open-weight gpt-oss models, in selected regions. They can be used on demand, served on a dedicated AI cluster, or replaced by your own imported model.
- What to understand
- On-demand access needs no provisioning and is billed by usage, which suits experiments and variable traffic. Dedicated AI clusters reserve capacity for predictable latency, isolation and custom models. Availability differs by region and model, and Oracle publishes retirement dates, so production designs track them. Oracle added a Model Discovery API to list models, capabilities and retirement dates, and IAM policies can allow or exclude specific model IDs for groups.
- How it shows up
- Scenarios describe traffic, isolation and customisation needs and ask which access mode fits.
- Common confusion
- Assuming every model is available in every region and in every access mode.
3.2 Call models through OpenAI-compatible endpointsBase URL, project OCID and the model value.
- What it means
- OCI exposes OpenAI-compatible APIs at an inference endpoint of the form inference.generativeai.<region>.oci.oraclecloud.com/openai/v1. Existing OpenAI SDK code works by changing the base URL, supplying OCI credentials and passing a Generative AI project OCID.
- What to understand
- The model field takes a model ID for on-demand models, the endpoint OCID for a model on a dedicated cluster, or a routing profile OCID in regions without local processing. API keys suit testing; IAM-based authentication is recommended for production. Requests run in OCI regions, which matters for data residency. Oracle's Smart Model Router can spread on-demand inference across regions a customer selects.
- How it shows up
- A question shows an SDK call and asks which value to put in the model field for a dedicated endpoint.
- Common confusion
- Pointing the SDK at the public OpenAI URL and expecting OCI models, credentials or residency.
3.3 Provision dedicated AI clustersHosting and fine-tuning clusters, units and endpoints.
- What it means
- A dedicated AI cluster is a set of GPU units reserved for one tenancy. A fine-tuning cluster trains custom models; a hosting cluster serves base, custom or imported models through endpoints. Unit types are tied to model families, so the model you choose fixes the unit shape.
- What to understand
- Clusters are billed per unit-hour, not per call. Oracle's pricing documentation sets a minimum of 744 unit-hours for a hosting cluster and one unit-hour per fine-tuning job, and some models carry a hosting price multiplier. One hosting cluster can serve several fine-tuned models that share a base model, with Oracle citing up to 50. Endpoint guardrails for content moderation, prompt injection and PII apply on dedicated endpoints.
- How it shows up
- Expect a cost or capacity scenario that asks whether a dedicated cluster or on-demand access is justified.
- Common confusion
- Using a fine-tuning cluster to serve production traffic, or assuming a hosting cluster can be paid by the call.
3.4 Fine-tune base models into custom modelsDatasets, training methods, metrics and serving.
- What it means
- Fine-tuning creates a custom model from a supported base model using your labelled data. The dataset is a JSONL file of prompt and completion pairs stored in Object Storage; Oracle requires at least 32 pairs and splits the single dataset 80/20 into training and validation automatically.
- What to understand
- The sequence is: prepare and upload data, create a fine-tuning cluster, create the custom model or a new version, create a hosting cluster, create an endpoint, then test. Parameter-efficient methods are used: T-Few for supported Cohere models and LoRA for supported Meta Llama models, both updating a small fraction of weights. Accuracy and loss on validation data show progress; falling loss is good, and near-perfect accuracy can signal overfitting.
- How it shows up
- Questions give the workflow out of order, or ask what accuracy and loss tell you after training.
- Common confusion
- Believing a custom model is callable as soon as training finishes, without a hosting endpoint.
3.5 Import and host external modelsBringing open-weight models into OCI Enterprise AI.
- What it means
- Model import lets you bring supported open-weight models, for example from Hugging Face or an Object Storage bucket, into OCI Generative AI. Imported models run on a hosting dedicated AI cluster and are reached through an endpoint like any custom model.
- What to understand
- Import suits teams that need a specific open model, a model already fine-tuned elsewhere, or a size not offered on demand. OCI does not fine-tune imported models, so any adaptation happens before import. Oracle keeps adding supported families, and the Responses API can call select imported models on a dedicated cluster. Check the supported list and required unit shape before planning capacity.
- How it shows up
- A scenario mentions a model trained outside OCI and asks how to serve it.
- Common confusion
- Expecting to fine-tune an imported model inside OCI or call it through on-demand inference.
High-yield concepts
The terms, rules and formulas this domain turns into questions most often.
-
On-demand model
Pretrained model called by ID with usage-based billing.
- Remember
- No provisioning; shared capacity.
- Tested as
- Cost and experimentation scenarios.
- Trap
- Expecting reserved throughput.
-
Dedicated AI cluster
GPU units reserved for one tenancy.
- Remember
- Two kinds: fine-tuning and hosting.
- Tested as
- Isolation and throughput questions.
- Trap
- Mixing up the two cluster kinds.
-
Unit-hour
Billing measure for dedicated clusters.
- Remember
- Hosting has a 744 unit-hour minimum commitment.
- Tested as
- Cost comparison items.
- Trap
- Treating cluster use as pay per request.
-
Endpoint
Callable address for a model on a hosting cluster.
- Remember
- Its OCID goes in the model field of API calls.
- Tested as
- API configuration questions.
- Trap
- Calling a custom model by base model ID.
-
Custom model
Result of fine-tuning a base model on your data.
- Remember
- Needs a hosting cluster and endpoint before use.
- Tested as
- Fine-tuning workflow questions.
- Trap
- Assuming it is live after training.
-
Training dataset format
JSONL lines with prompt and completion fields.
- Remember
- At least 32 pairs; auto split 80/20.
- Tested as
- Data preparation questions.
- Trap
- Supplying separate validation files.
-
T-Few
Parameter-efficient method for supported Cohere models.
- Remember
- Updates a small subset of weights.
- Tested as
- Method matching questions.
- Trap
- Pairing it with Llama models.
-
LoRA
Low-rank adapter method for supported Llama models.
- Remember
- Base weights stay frozen.
- Tested as
- Method matching questions.
- Trap
- Thinking it retrains the full model.
-
Accuracy and loss
Validation metrics reported after fine-tuning.
- Remember
- Falling loss is good; perfect accuracy may mean overfitting.
- Tested as
- Training evaluation questions.
- Trap
- Reading high accuracy as always good.
-
Model import
Bringing supported open-weight models into OCI.
- Remember
- Hosted on a dedicated cluster; not fine-tuned by OCI.
- Tested as
- External model scenarios.
- Trap
- Expecting on-demand access.
-
Routing profile
OCID used as the model value where local processing is unavailable.
- Remember
- Sends requests to a region that serves the model.
- Tested as
- Regional availability questions.
- Trap
- Using a plain model ID in that region.
-
Project OCID
Generative AI project identifier passed with OpenAI-compatible calls.
- Remember
- Set through the SDK project argument.
- Tested as
- SDK setup questions.
- Trap
- Omitting it and blaming the model.
Side by side
Concepts candidates mix up, compared on the point that decides the answer.
Model access modes
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| On-demand | Shared capacity, called by model ID | Pay for usage, no commitment | Assuming isolation and fixed latency |
| Dedicated hosting cluster | Reserved units serving endpoints | Predictable performance, monthly minimum | Using it for occasional tests |
| Imported model on dedicated cluster | Your chosen open-weight model on reserved units | Bring your own weights | Trying to fine-tune it in OCI |
Dedicated cluster types
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Fine-tuning cluster | Runs training jobs for custom models | Billed per job, minimum one unit-hour | Serving inference from it |
| Hosting cluster | Serves models through endpoints | 744 unit-hour minimum; many custom models per cluster | Creating one per custom model by default |
Fine-tuning methods
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| T-Few | Inserts and trains a small set of extra weights | Used with supported Cohere models | Applying it to Llama |
| LoRA | Adds trainable low-rank adapter matrices | Used with supported Llama models | Expecting full weight updates |
Exam traps
Each trap is an answer that looks right. Learn the clue that should change your mind.
-
Custom model is live after training
- What candidates pick
- Call the custom model right after the job completes.
- Why it looks right
- Training feels like the last step.
- Why it's wrong
- A custom model must be placed on a hosting cluster and given an endpoint before any request can reach it.
Clue to watch for The scenario reports errors calling a newly trained model.
-
Dedicated means pay per call
- What candidates pick
- Pick a hosting cluster for a pilot with a few hundred requests.
- Why it looks right
- Dedicated sounds faster.
- Why it's wrong
- Hosting clusters bill unit-hours with a 744 unit-hour minimum, so low volume is cheaper on demand.
Clue to watch for Low or unpredictable traffic and a tight budget.
-
Fine-tune imported models in OCI
- What candidates pick
- Create a fine-tuning cluster for a Hugging Face model.
- Why it looks right
- OCI fine-tunes its own base models.
- Why it's wrong
- OCI does not fine-tune imported models; adapt them before import and host them on a dedicated cluster.
Clue to watch for The model came from outside OCI.
-
Two datasets for validation
- What candidates pick
- Upload separate training and validation files.
- Why it looks right
- Common practice elsewhere.
- Why it's wrong
- OCI takes one dataset and splits it 80/20 automatically, so a separate validation file is not part of the workflow.
Clue to watch for Questions about preparing fine-tuning data.
-
Perfect accuracy is the goal
- What candidates pick
- Select the run with 100% training accuracy.
- Why it looks right
- Higher looks better.
- Why it's wrong
- Near-perfect accuracy with poor generalisation suggests overfitting; check validation loss trends instead.
Clue to watch for Very small datasets or many epochs.
-
Use the OpenAI base URL
- What candidates pick
- Keep api.openai.com and change only the model name.
- Why it looks right
- The SDK is the same.
- Why it's wrong
- OCI models are served at the OCI inference endpoint with OCI credentials and a project OCID; the OpenAI URL never reaches them.
Clue to watch for Data residency or OCI credentials are mentioned.
Answer framework
Apply these steps when choosing and configuring a model for an OCI Enterprise AI workload.
- Check which supported models are available in the target region, including retirement dates and whether a routing profile is needed.
- Choose on-demand access for experiments and variable load, or a dedicated hosting cluster for steady volume, isolation or custom models.
- If prompting and retrieval fall short on style or task behaviour, prepare at least 32 JSONL prompt and completion pairs for fine-tuning.
- Run fine-tuning on a fine-tuning cluster, review accuracy and loss, then host the custom model on a hosting cluster endpoint.
- Configure callers with the OCI base URL, project OCID and the correct model value: model ID, endpoint OCID or routing profile OCID.
Practice question
Want more like this? The 1Z0-1158-26 practice test has questions for every domain.
Checklist
0 of 12
Tick each one you can do without notes. Your ticks stay in this browser.
Study strategy
Start in the OCI Generative AI documentation section on models by region and pick two chat models and one embedding model you will use throughout your preparation. Call each on demand with the OpenAI SDK, setting the OCI base URL, an API key and a project OCID, so the configuration becomes routine.
You may not have budget for a dedicated cluster, so learn that part from the docs and pricing pages. Write out the fine-tuning workflow as a numbered list from dataset to endpoint, then add the cluster type used at each step. Create a small JSONL file with forty prompt and completion pairs to confirm you understand the format. Work two cost examples by hand: one month of a hosting cluster against on-demand usage at two traffic levels.
Read the model import page and note which steps differ from fine-tuning, especially that OCI does not tune imported models. Finish by reviewing the latest Oracle release notes for new models, retirements and features such as the Model Discovery API, since this domain changes more often than the conceptual ones and outdated assumptions cost marks.
FAQ
How much of 1Z0-1158-26 is Domain 3, Configuring OCI Enterprise AI Models?
Oracle does not publish weights for this exam. Practice banks give model configuration around a fifth of items, similar to the agentic concepts section.
Is OCI Enterprise AI a different service from OCI Generative AI?
Oracle markets OCI Enterprise AI as the platform for building and governing AI workloads, and its models, clusters and APIs are documented under OCI Generative AI. Use the Generative AI docs for configuration details.
What is the hosting cluster minimum commitment?
Oracle's pricing documentation sets 744 unit-hours per hosting cluster. Usage below that is billed at the minimum.
How many training examples do I need?
At least 32 prompt and completion pairs in one JSONL file. OCI splits it into training and validation data automatically.
Can I fine-tune any model?
Only supported base models. Imported models cannot be fine-tuned inside OCI, so adapt them before you import them.
What goes in the model field for a dedicated endpoint?
The endpoint OCID. On-demand calls use the model ID, and regions without local processing can use a routing profile OCID.
Other Artificial Intelligence & Machine Learning certifications
Where the same subject shows up in other exams. Cards marked "Same subject" open the matching domain.
Test your Domain 3 knowledge
You have covered model access modes, OpenAI-compatible calls, dedicated clusters, unit-hours, fine-tuning and model import. Check how you handle configuration and cost scenarios.
Keep going
- Training Training Built on your Weak areas
- Practice test Questions by domain, with explanations
- Mock exam Full length, timed like the real exam
- Readiness test Find your weakest domain in minutes
- Cheat sheet Key terms and formulas on one page
- Ask the 1Z0-1158-26 AI tutor Get any concept explained your way