How it works
Part of a base model (pre-trained) that predicted text. Training uses instruction-response pairs. The model learns to answer questions, execute tasks and follow commands.
The instruction tuning step is what differentiates a base model from an instructed model (instruct). Without it, the model only completes text. With it, the model answers questions coherently.
The process generally involves supervised fine-tuning (SFT) with curated instruction data. Companies like Meta, Mistral and Google use instruction tuning to create the "Instruct" versions of their models.
Practical example
Llama 3 base predicts the next token. After instruction tuning, it becomes Llama 3 Instruct, able to answer "explain what an API is" coherently. Without this step, the base model would only complete text without answering the question.
A base model may have the knowledge, but it does not know how to present it. Instruction tuning teaches the response format, tone and ability to follow complex instructions with multiple steps.
What it's for
Understanding instruction tuning helps choose between a base model and an instructed model for an application. Base models are useful for custom fine-tuning. Instructed models are useful for direct use as assistants.
For developers, this means knowing when to use each type of model. For researchers, it means understanding how to improve models with high quality instruction data.
Difference between base model and instructed model
Base model predicts the next token based on training text. It is good for completing text and for custom fine-tuning. Instructed model follows commands and answers questions. It is good for direct use as an assistant or agent.
An instructed model loses a little of the base model's flexibility (it can be harder to fine-tune). But it gains in usability and ability to follow instructions.