follow

Definition

Shorthand for instruction following: producing output that respects both the content and the constraints of a request. Base models predict plausible continuations and are poor at this; the ability is installed afterwards through instruction tuning and preference optimisation. Worth separating from knowledge — a model can know the answer and still ignore the requested format, length, or refusal boundary.

Explainers using this term

• Alignment, Explained — From RLHF to Constitutional AITraining language models to follow instructions with human feedback
• KL Divergence From Scratch — Measuring the Gap Between Two DistributionsTraining language models to follow instructions with human feedback
• Instruction Tuning and RLHF from Scratch — How a Model Learns to Follow OrdersTraining language models to follow instructions with human feedback