OCI AI: Generative AI Professional — Lesson 3: Model selection, customization choices, fine-tuning data, evaluation sets, endpoint deployment, SDK usage, and application integration

Learning Objectives

• Identify the configuration and data choices that shape Model selection, customization choices, fine-tuning data, evaluation sets, endpoint deployment, SDK usage, and application integration.
• Run normal and boundary tests and record the outcome in model comparison scorecard and integrated prototype trace.
• Separate root cause from symptom when investigating a customized model improving style while reducing factual task accuracy.

Detailed Topics

• Model selection
• Customization choices
• Fine-tuning data
• Evaluation sets
• Endpoint deployment
• SDK usage
• And application integration
• Decision ownership, prerequisites, and acceptance evidence for Model selection, customization choices, fine-tuning data, evaluation sets, endpoint deployment, SDK usage, and application integration
• Relevant access, data, integration, performance, recovery, and support considerations

Instructor Demonstration

The demonstration traces Model selection, customization choices, fine-tuning data, evaluation sets, endpoint deployment, SDK usage, and application integration from requirement through execution and review. Learners see where evidence should be collected, how model comparison scorecard and integrated prototype trace supports verification, and which warning signs demand investigation.

Hands-On Exercise

Build or execute Model selection, customization choices, fine-tuning data, evaluation sets, endpoint deployment, SDK usage, and application integration for a defined business case. Introduce the condition “a customized model improving style while reducing factual task accuracy,” inspect the resulting evidence, correct the cause without discarding the audit trail, and document the final state in model comparison scorecard and integrated prototype trace.

1. Confirm scope, access, sample records, and acceptance criteria.
2. Perform the configuration or transaction using a controlled sequence.
3. Inspect outputs and downstream effects rather than relying on screen messages alone.
4. Close with a support note, preventive action, and reproducible validation step.

Scenario Analysis

• Clarify the business trigger, affected roles, required records, and acceptable result for Model selection, customization choices, fine-tuning data, evaluation sets, endpoint deployment, SDK usage, and application integration.
• Compare at least two implementation or correction choices and identify which evidence would distinguish them.
• Decide how model comparison scorecard and integrated prototype trace should be reviewed, approved, retained, and handed to the next owner.

Troubleshooting Guidance

Diagnose Model selection, customization choices, fine-tuning data, evaluation sets, endpoint deployment, SDK usage, and application integration by checking effective settings, ownership, source records, dependencies, and recent releases. Use model comparison scorecard and integrated prototype trace to locate the earliest divergence. Treat “a customized model improving style while reducing factual task accuracy” as a hypothesis until the same test passes after correction and still passes for a boundary case.

Validation Checklist

• The stated requirement for Model selection, customization choices, fine-tuning data, evaluation sets, endpoint deployment, SDK usage, and application integration is linked to an owner and an observable acceptance condition.
• Model comparison scorecard and integrated prototype trace contains enough detail for another practitioner to repeat the check.
• Normal processing, one boundary condition, and the identified failure path have all been tested.
• Access, data integrity, downstream impact, rollback, and support ownership are recorded.

Lesson Outcome

Learners can apply Model selection, customization choices, fine-tuning data, evaluation sets, endpoint deployment, SDK usage, and application integration, explain the decisions made, and defend the result with model comparison scorecard and integrated prototype trace.