Free Oracle Cloud Infrastructure 2026 Generative AI Professional practice test: 30 questions with answers

Dedicated AI clusters provide reserved capacity for predictable performance in production. On-demand inference shares capacity, and trial or laptop environments are not production hosting.

IAM policies grant groups permissions on resources in compartments, which is how access is controlled. Public buckets, shared root passwords and laptop allowlists are poor controls.

Compartments group resources for access control and cost management. Subnets, boot volumes and listeners are networking or compute components.

Oracle Database 23ai with AI Vector Search stores and queries embeddings alongside relational data. Text files, printed indexes and spreadsheets do not provide vector search.

Predictable performance for production and hosting a custom fine-tuned model that needs reserved capacity are sound reasons. Removing access control, one-off experiments and laptop hosting are not.

Endpoint configuration in the generative AI service manages scaling and updates through the console or API. Laptop files, spreadsheets and printed guides do not manage endpoints.

Filtering personal data and applying access controls limit exposure through responses. Adding all records, public storage and disabled access controls each increase exposure.

An output safety filter checks responses and blocks or flags harmful content before it is shown. Unchecked output, higher temperature and removed rules each allow harm through.

Bias can enter through training data, prompts and retrieval sources and should be tested across groups. The other statements are inaccurate.

Human review of high-impact outputs and escalation of uncertain responses are sound uses. Removing review after tests, reviewing every low-risk draft and hiding reviews are not appropriate in the way described.

Monitoring quality and safety against a baseline with alerts detects drift. Assumptions, disabled logs and sampling only the first answer each miss it.

Disclosing AI use and explaining limitations supports transparency. Pretending to be human, hiding retrieval and removing sources each mislead users.

Few-shot prompting uses examples in the prompt to guide format and style without retraining. Zero-shot has no examples, and clipping and batch inference are unrelated.

Fine-tuning on many labelled examples teaches a consistent style in the model's weights. Longer prompts, higher temperature and bigger context help less consistently.

Parameter-efficient methods such as T-Few or adapter-based fine-tuning update a small set of parameters. Full retraining, quantisation and distillation work differently.

Strong training performance with weak generalisation is overfitting. Underfitting, tokenisation drift and latency are different problems.

A held-out test set that the model did not see during fine-tuning measures generalisation. Evaluating on training data, self-rating and speed-only tests each give misleading results.

Clear instructions, relevant context and self-checks against requirements improve output. Leaving the format open and mixing unrelated tasks each reduce reliability.

The retriever finds relevant chunks for the query, which the model then uses to ground its answer. Training, encryption and answer storage are different functions.

Chunk size trades relevance against context: large chunks dilute relevance, and small chunks lose surrounding meaning. The other options are incorrect.

Groundedness is shown when each claim can be traced to retrieved passages. Word counts, latency and corpus size do not measure grounding.

Hybrid search combines vector similarity with keyword matching for better recall and precision. Random sampling, batch normalisation and quantisation are unrelated to retrieval quality.

Re-indexing on change with a refresh schedule and recording source and version keep answers current and traceable. Keeping outdated versions, memory-based rebuilds and personal storage each undermine accuracy.

Query rewriting reformulates the question to improve retrieval. Returning arbitrary documents, huge chunks and removed metadata each reduce relevance.

Self-attention computes weighted relationships between tokens, which lets the model use context across the sequence. The tokenizer, output layer and learning rate schedule have different roles.

Prompt and response tokens share the context window, so long inputs must be chunked or retrieved selectively. The other options describe unrelated limits.

Embeddings are vectors that capture semantic meaning and support similarity search. Tokens, checkpoints and temperature are different concepts.

Temperature controls sampling randomness, where lower values make outputs more deterministic. Output length, GPU allocation and training data length are different settings.

Hallucinations are confident but unsupported outputs, and grounding with citations reduces them. Raising temperature increases them, and no single prompt eliminates them.

Pre-training learns general patterns from large corpora, and fine-tuning adapts the model to tasks or styles. The reversed and conflated statements are incorrect.