OCI 1Z0-1122-25 FAQ for exam format, topics, prep strategy, practice, and common candidate traps.
This exam is about AI fundamentals and judgment, not deep model-building math. Strong answers usually classify the issue as data, metric, model behavior, GenAI grounding, or governance before they talk about tools.
| Question | Short answer |
|---|---|
| Do I need to be a data scientist? | No. You need clear fundamentals around data quality, metrics, GenAI basics, and responsible AI. |
| What is the highest-yield area? | Metric choice, leakage and overfitting, and grounding or retrieval basics. |
| What does the exam punish most? | Hype-driven answers that skip evaluation, monitoring, privacy, or fairness. |
| What hands-on work matters most? | One small end-to-end flow: define the task, inspect data quality, choose a metric, and reason about monitoring or grounding. |
| What should I trust if summaries disagree? | The current Oracle exam page and OCI documentation. |
No. You need practical AI-system instincts more than advanced research depth.
Questions usually collapse into one of these lanes:
| Lane | What it is really testing |
|---|---|
| data | data quality, labels, imbalance, leakage, and preprocessing risk |
| evaluation | metric fit, generalization quality, and whether the score means anything |
| GenAI | prompting, grounding, retrieval, and hallucination control |
| governance | privacy, fairness, safety, and responsible-use boundaries |
| lifecycle | deployment, monitoring, drift, and operational feedback |
The highest-yield area is usually metric and evaluation judgment.
| If the question is mostly about… | Start with… | Strongest first move |
|---|---|---|
| a suspiciously strong score | leakage or broken evaluation design | distrust the score before praising the model |
| bad production behavior | data drift, poor retrieval, or task-metric mismatch | find the weak layer first |
| fluent but unreliable GenAI output | grounding and retrieval quality | model size alone is rarely the answer |
| fairness or privacy concern | responsible AI controls | governance belongs in the first answer, not the appendix |
It punishes shallow AI thinking.
Common traps:
| Trap | Better reading |
|---|---|
| “The metric is high, so the model is good.” | first ask whether the metric matches the task and whether leakage is inflating it |
| “A bigger model will fix the problem.” | weak data, wrong metric, or bad grounding often matter more |
| “Generative AI is mostly prompting.” | retrieval, grounding, safety, and evaluation still matter |
| “Governance is separate from model quality.” | privacy, fairness, and safety are part of system quality |
You do not need a giant ML project. You need one believable AI workflow.
Route the miss by failure layer.
| If your misses sound like… | Weak lane | Fix next |
|---|---|---|
| “I chose accuracy and it looked fine.” | evaluation | review imbalance, precision, recall, F1, and task-fit metrics |
| “The score looked amazing, but I missed the trick.” | data and evaluation | review leakage, split design, and held-out validation logic |
| “I blamed the model instead of retrieval.” | GenAI | review grounding, context quality, and retrieval fit |
| “I ignored fairness, privacy, or safety.” | governance | review responsible AI controls and operational ownership |
Use this order:
1Z0-1122-25If a summary sounds cleaner than the Oracle source, downgrade it.
Do less broad reading and more classification practice.
| Keep doing | Stop doing |
|---|---|
| rereading the cheat sheet and glossary | opening unrelated new ML frameworks |
| drilling confusion pairs like leakage vs overfitting and grounding vs fine-tuning | treating every AI question like a product question |
| checking Oracle docs for disputed boundaries | building a large late-stage AI project |
| practicing data -> metric -> model -> governance order | trusting unsupported community summaries over Oracle docs |