aligning

Definition

Bringing a model's behaviour into line with human intent, instructions, and safety policy. A purely pre-trained model only continues text; aligning is the later stage — instruction tuning, preference optimisation — that makes it answer the question actually asked and refuse what it should refuse. Keep it separate from capability: alignment work mostly changes how existing ability is expressed, not how much of it there is.

Explainers using this term

• Building a Dataset in Practice — Collect, Clean, BlendSelf-Instruct: Aligning Language Models with Self-Generated Instructions