Overview
Domain 4 is the conceptual core of the generative AI material. It begins with the difference between generative models, which learn the distribution of their training data and can produce new samples, and discriminative models, which only learn to separate classes. It then explains large language models: text is split into tokens, the model assigns a probability to every possible next token, and a decoding strategy picks one at a time.
You need the transformer architecture at a high level, including self-attention, encoders, decoders and why transformers process tokens in parallel. The section closes with ways to steer a model: prompt engineering techniques such as few-shot and chain-of-thought prompting, instruction tuning, and fine-tuning approaches including parameter-efficient methods. Hallucination and its mitigations also come up here.
How it's tested
Questions test whether you can explain LLM behaviour and pick the right technique for a goal. Typical items ask what a parameter does, which transformer type fits a task, or whether prompting, retrieval or fine-tuning best solves a stated problem. Cost and effort are often the deciding clue.
-
Explain a decoding setting
You predict how temperature or top p changes output, or why greedy decoding gives repetitive but stable text.
-
Pick the architecture
Embedding and classification tasks point to encoders, text generation to decoders, and translation to encoder-decoder models.
-
Prompt, retrieve or tune
A goal is stated and you choose prompt engineering, retrieval-augmented generation or fine-tuning as the lowest-effort fix.
A legal team finds that a model answers questions about its contracts with confident but invented clauses. The problem is missing knowledge, not style, so fine-tuning on writing tone would not help. Retrieval-augmented generation, which fetches relevant contract passages and adds them to the prompt, grounds the answer in real text and reduces hallucination without retraining. That is the expected choice.
Objectives 5
Official objectives from the Oracle outline, in outline order.
4.1 Explain generative AI fundamentalsHow generative models differ from predictive ones.
- What it means
- Generative AI refers to models that learn patterns in training data well enough to create new content: text, images, audio, code or video. Discriminative models, by contrast, learn boundaries between classes and output a label or score.
- What to understand
- Generative models learn an approximation of the data distribution, so sampling from them yields new examples that resemble the training data. Text generation uses transformer LLMs, while image generation often uses diffusion models that learn to remove noise step by step. Business uses include drafting, summarising, chat assistants and code generation. Risks include hallucination, bias copied from training data and intellectual property concerns.
- How it shows up
- Expect a question that separates a generative use case from a classification one.
- Common confusion
- Calling any deep learning model generative, even when its only output is a class label.
4.2 Explain LLM basicsTokens, next-token prediction, decoding and hallucination.
- What it means
- A large language model is a transformer trained on very large text corpora to predict the next token. A token is a word or word fragment. Generation repeats next-token prediction until a stop condition or the token limit is reached.
- What to understand
- The model outputs a probability distribution over its vocabulary at each step. Greedy decoding always picks the most likely token, which is stable but can repeat. Sampling picks according to the probabilities, and temperature reshapes them: low values sharpen the distribution for predictable output, high values flatten it for variety. Top k and top p restrict choices to the most likely tokens. The context window caps how many tokens the model can consider at once. Hallucination is fluent but false output.
- How it shows up
- You may be asked what raising temperature does or why an LLM can state false facts confidently.
- Common confusion
- Thinking temperature changes what the model knows; it only changes how tokens are sampled.
4.3 Explain transformers fundamentalsSelf-attention and encoder, decoder and encoder-decoder designs.
- What it means
- Transformers process all tokens of an input together and use self-attention to let each token weigh the relevance of every other token. This captures long-range context and allows parallel training, which recurrent networks could not do efficiently.
- What to understand
- Because tokens are processed together, positional encoding adds word-order information. Encoder-only models, such as BERT-style models, build contextual representations used for embeddings, search and classification. Decoder-only models generate text token by token and power most chat LLMs. Encoder-decoder models read an input sequence and produce a different output sequence, a natural fit for translation and summarisation.
- How it shows up
- Questions ask which transformer type suits embeddings, chat or translation, or what attention provides.
- Common confusion
- Assuming every LLM uses both an encoder and a decoder; most chat models are decoder-only.
4.4 Explain prompt engineering and instruction tuningShaping outputs through input design and training on instructions.
- What it means
- Prompt engineering is designing the input text so the model produces the desired output without changing its weights. Instruction tuning is a training stage in which a model learns from many instruction and response pairs so it follows requests well.
- What to understand
- Zero-shot prompting gives only the task, few-shot prompting adds worked examples in the prompt, and this in-context learning needs no training. Chain-of-thought prompting asks the model to reason step by step, which helps on multi-step problems. Clear roles, formats and constraints improve consistency. Instruction tuning, often followed by reinforcement learning from human feedback, is why chat models respond to plain requests. Prompt injection is a risk where crafted input overrides the intended instructions.
- How it shows up
- Expect to distinguish few-shot prompting from fine-tuning, or to identify chain-of-thought prompting.
- Common confusion
- Believing few-shot examples permanently teach the model; they only affect the current request.
4.5 Explain LLM fine-tuningAdapting a pretrained model with additional training.
- What it means
- Fine-tuning continues training a pretrained model on a smaller, task-specific dataset so it adapts to a domain, style or format. It changes model weights, unlike prompting or retrieval.
- What to understand
- Full fine-tuning updates all parameters and is costly in compute and storage. Parameter-efficient fine-tuning updates only a small set of added or selected weights; LoRA and T-Few are examples, and OCI Generative AI offers such methods on dedicated AI clusters. Fine-tuning suits consistent tone, specialised formats or narrow tasks. For knowledge that changes often, retrieval-augmented generation is usually cheaper and easier to keep current. Fine-tuning needs good-quality labelled examples.
- How it shows up
- A scenario weighs fine-tuning against prompting or RAG and asks for the most cost-effective approach.
- Common confusion
- Treating fine-tuning as the fix for missing or frequently changing facts.
High-yield concepts
The terms, rules and formulas this domain turns into questions most often.
-
Token
A word or sub-word unit the model reads and writes.
- Remember
- Limits and pricing are counted in tokens.
- Tested as
- Context and cost questions.
- Trap
- Equating one token with one word.
-
Next-token prediction
LLMs generate by repeatedly predicting the next token.
- Remember
- Output is probabilistic.
- Tested as
- Explaining LLM behaviour.
- Trap
- Thinking the model retrieves stored sentences.
-
Temperature
Scales token probabilities before sampling.
- Remember
- Low is focused, high is varied.
- Tested as
- Parameter questions.
- Trap
- Thinking high temperature improves accuracy.
-
Top k and top p
Limit sampling to the k most likely tokens or a cumulative probability mass.
- Remember
- Both narrow the candidate pool.
- Tested as
- Parameter questions.
- Trap
- Mixing up count-based k with probability-based p.
-
Context window
Maximum tokens the model can attend to at once.
- Remember
- Includes prompt and response.
- Tested as
- Long document scenarios.
- Trap
- Assuming it only counts the response.
-
Hallucination
Fluent output that is factually wrong.
- Remember
- Grounding with retrieval reduces it.
- Tested as
- Risk and mitigation items.
- Trap
- Believing lower temperature eliminates it.
-
Self-attention
Each token weighs every other token for context.
- Remember
- Enables parallel processing and long-range context.
- Tested as
- Transformer questions.
- Trap
- Linking it to RNN hidden state.
-
Encoder vs decoder
Encoders build representations; decoders generate text.
- Remember
- Embeddings use encoders; chat uses decoders.
- Tested as
- Architecture choice items.
- Trap
- Choosing a decoder for semantic search embeddings.
-
Few-shot prompting
Including examples in the prompt.
- Remember
- No weights change.
- Tested as
- Prompt versus tuning items.
- Trap
- Calling it fine-tuning.
-
Chain-of-thought
Prompting the model to show intermediate reasoning steps.
- Remember
- Helps multi-step problems.
- Tested as
- Technique identification.
- Trap
- Thinking it guarantees correct answers.
-
Instruction tuning
Training on instruction and response pairs.
- Remember
- Makes models follow plain requests.
- Tested as
- Why chat models behave as they do.
- Trap
- Confusing it with writing better prompts.
-
Parameter-efficient fine-tuning
Updating a small subset of weights, for example LoRA or T-Few.
- Remember
- Cheaper than full fine-tuning.
- Tested as
- Cost-focused tuning items.
- Trap
- Assuming it retrains the entire model.
Side by side
Concepts candidates mix up, compared on the point that decides the answer.
Ways to adapt an LLM
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Prompt engineering | Change the input text. | No training, instant, cheapest. | Expecting it to add large new knowledge bases. |
| Retrieval-augmented generation | Add retrieved documents to the prompt. | Supplies current facts without retraining. | Assuming it changes model weights. |
| Fine-tuning | Further train on task data. | Changes weights for style, format or narrow skills. | Using it for facts that change weekly. |
Transformer families
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Encoder-only | Produces contextual vectors. | Suited to embeddings and classification. | Using it for free-form chat. |
| Decoder-only | Generates tokens left to right. | Basis of most chat LLMs. | Expecting it to output embeddings by default. |
| Encoder-decoder | Maps one sequence to another. | Natural fit for translation and summarisation. | Thinking every LLM needs both parts. |
Decoding settings
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Greedy | Always takes the top token. | Deterministic, can repeat. | Assuming it is the most creative. |
| High temperature | Flattens probabilities. | More varied, more risk of errors. | Using it for factual extraction. |
| Low temperature | Sharpens probabilities. | More predictable output. | Believing it stops hallucination entirely. |
Exam traps
Each trap is an answer that looks right. Learn the clue that should change your mind.
-
Temperature and accuracy
- What candidates pick
- Raise temperature for more accurate answers
- Why it looks right
- Higher settings sound more powerful.
- Why it's wrong
- Temperature controls randomness in token selection; higher values increase variety and the chance of drifting off-fact.
Clue to watch for Words like creative versus consistent.
-
Fine-tune for fresh facts
- What candidates pick
- Fine-tune the model weekly
- Why it looks right
- Training sounds like teaching new information.
- Why it's wrong
- Retraining is slow and costly for changing data, while retrieval adds current documents at query time.
Clue to watch for Content that updates often.
-
Few-shot equals training
- What candidates pick
- Few-shot learning changes weights
- Why it looks right
- The word learning suggests training.
- Why it's wrong
- Examples in the prompt guide only that request; the model parameters stay exactly the same.
Clue to watch for Examples placed inside the prompt.
-
RNN memory in transformers
- What candidates pick
- Transformers pass a hidden state step by step
- Why it looks right
- Both handle sequences.
- Why it's wrong
- Transformers use self-attention over all tokens at once rather than a recurrent hidden state.
Clue to watch for Question about parallel processing.
-
Bigger context fixes hallucination
- What candidates pick
- Increase the context window
- Why it looks right
- More room seems to mean more knowledge.
- Why it's wrong
- A larger window only helps if relevant facts are actually supplied; it does not add knowledge by itself.
Clue to watch for Model invents facts it was never given.
-
Generative for classification
- What candidates pick
- Use a generative model to label emails
- Why it looks right
- Generative AI is the topical choice.
- Why it's wrong
- If the output is only a class, a discriminative classifier is the natural fit; generation is not required.
Clue to watch for Output described as a category.
Answer framework
Use this sequence when a Domain 4 question asks how to get better output from an LLM.
- Diagnose the problem as one of format or style, missing or outdated knowledge, or poor reasoning on multi-step tasks, since each has a different first remedy.
- Try prompt engineering first for format, tone and reasoning issues, using clear instructions, examples or chain-of-thought.
- Choose retrieval-augmented generation when the model lacks private or current facts, so answers are grounded in supplied documents.
- Choose fine-tuning only when consistent behaviour or a specialised style is needed and good labelled examples exist, accepting the extra cost of training and hosting a custom model.
- Adjust decoding settings last: lower temperature or top p for consistency, raise them when varied creative output is wanted, and remember these settings never add knowledge the model lacks.
Practice question
Want more like this? The 1Z0-1122-26 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
Allow seven to nine hours, the largest block among the conceptual sections. Watch the generative AI and LLM lessons in Oracle's AI Foundations course and take notes in four groups: what an LLM is, how it decodes, how the transformer is built, and how to adapt it.
Get hands-on if you can. Open any chat model, or the OCI Generative AI playground if you have tenancy access, and run the same prompt at temperature 0.1 and 0.9. Then try zero-shot versus few-shot versions of a formatting task. Seeing the difference fixes these ideas better than reading definitions.
Next, build a three-row table comparing prompt engineering, RAG and fine-tuning by cost, speed, whether weights change and what problem each solves. Most scenario questions in this section are answered from that table.
Finish with practice questions, focusing on parameter behaviour and transformer family choices. Before moving on, note which of these concepts Oracle exposes in its own service, such as temperature, embeddings and dedicated clusters, to prepare for Domain 6.
FAQ
How heavily is Intro to Generative AI & LLMs weighted?
Oracle publishes no weights for 1Z0-1122-26. Secondary estimates place this generative AI section at about six of the 40 questions, and its concepts also support Domain 6.
Do I need to know specific model names?
Not for this section. It tests concepts. Model families offered on OCI matter more in Domain 6.
Is RAG part of this section or Domain 6?
Both touch it. Here you learn why grounding reduces hallucination; Domain 6 covers how Oracle implements retrieval with Vector Search and Select AI.
What does instruction tuning add?
It trains a base model on instruction and answer pairs so it follows plain requests, which is why chat models respond helpfully.
Will the exam ask about diffusion models?
Possibly at a recognition level: diffusion models generate images by learning to remove noise. Text generation questions are more common.
What is the safest default when choosing an adaptation method?
Start with prompting, add retrieval for missing knowledge, and fine-tune only when style or task consistency still falls short.
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 4 knowledge
You have studied LLM behaviour, transformers, prompting and fine-tuning. Check that you can pick the right technique under time pressure.
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 1Z0-1122-26 AI tutor Get any concept explained your way