Concept guide · LLMs and generative AI
What is the difference between pretraining, instruction tuning, and alignment?
A slower explanation for when the short interview answer is not enough. Start with the intuition, then connect it to the technical idea and the way an interviewer expects you to reason about it.
ML area
LLMs and generative AI → Core interview concept
Interview format
Core
Question
ML-093
Start with the idea
In simple terms
This question is asking you to explain What is the difference between pretraining, instruction tuning, and alignment in a way that connects the idea to a real machine-learning decision. Start with the role this concept plays, then explain what changes when the data, target, or deployment situation changes.
Build the answer
How to reason about it
- 1Define the concept before naming algorithms or implementation details.
- 2Connect it to the data, target, decision, or failure mode in the question.
- 3State the assumption or evaluation check that would make the explanation trustworthy.
Go one level deeper
The technical explanation
Pretraining learns broad language or multimodal patterns from large datasets, usually through a self-supervised objective. Instruction tuning adapts the model to follow examples of tasks and responses. Alignment adds preferences or constraints through methods such as preference optimisation, human feedback, or rule-based evaluation. None of these guarantees factual accuracy for a new domain. The safest way to understand this topic is to connect the definition to assumptions, data available at prediction time, the decision it supports, and how you would evaluate whether the result is useful.
Say this in the interview
Pretraining learns broad language or multimodal patterns from large datasets, usually through a self-supervised objective. Instruction tuning adapts the model to follow examples of tasks and responses. Alignment adds preferences or constraints through methods such as preference optimisation, human feedback, or rule-based evaluation. None of these guarantees factual accuracy for a new domain.
Keep preparing