Home › Key concepts › Pre-training, fine-tuning and alignment

Key concept · Walk (intermediate) · the same on every platform

Pre-training, fine-tuning and alignment

Three stages turn raw text into a helpful assistant.

What it is

Pre-training teaches language and knowledge from trillions of tokens. Instruction tuning (supervised fine-tuning) teaches it to follow requests from examples. Preference tuning (such as RLHF or DPO) teaches it which answers people prefer.

Why it matters

It tells you which stage to touch for which problem, and why fine-tuning rarely adds new knowledge.

An example

A base model continues text; after instruction tuning it answers questions; after preference tuning it answers them politely and safely.

The common mistake

Fine-tuning to teach facts that change monthly, when retrieval would do it better.

Where you meet it on each platform

The idea stays the same; only the product name changes.

JobAWSAzureGoogle CloudDatabricksOpen source
Fine-tuningteach a model your style or task with examplesBedrock custom models / SageMakerFoundry fine-tuningVertex AI tuningMosaic AI fine-tuningHugging Face TRL or Unsloth

Related ideas

Where it shows up

Part of the Walk level on these paths: LLM Engineering (adapting models).