AI Dictionary › AI Fundamentals

Instruction Tuning

Instruction tuning is a training phase in which a language model is refined on a set of examples made of instructions and correct responses, teaching it to follow commands expressed in natural language. Unlike pre-training, which teaches the model to predict the next piece of text, this phase teaches it to behave like an assistant that carries out specific requests. It is one of the steps that turns a base model into something usable in a conversational product.

Definition

How it works

The model is exposed to many instruction-response pairs covering different tasks: answering questions, summarizing text, writing code, translating. Through this supervised training, the model generalizes the behavior of following instructions even to tasks not seen during tuning, as long as they are structurally similar to the ones observed.

Applications

It is a common step in building conversational assistants based on LLMs, often combined with later techniques such as RLHF to further refine response style and safety. It is applied both by those who build base models and by those who customize them for a specific domain.

History & etymology

It is associated with the shift of language models from simple text-completion systems to assistants capable of executing instructions, a key change in the recent evolution of generative models.

Related terms

More in AI Fundamentals

Put it into practice

From our network

AGORÀ Intelligence: Enterprise AI Governance Platform

Govern AI at scale: policies, adoption and measurable results on your data. Built for boards and C-suite.

Visit agora-intelligence.com →

From the Agora Intelligence blog

More on agora-intelligence.com →

📱 Download the Android app (beta) iOS coming soon

Say what you mean. Get what you need.

Grace Certified, the AI coach that trains and certifies your prompt engineering, by Agora Intelligence.