Training
Instruction Tuning
Also known as: instruction fine-tuning
Fine-tuning a raw model on examples of instructions and good responses, so it learns to follow requests and behave like a helpful assistant.
Instruction tuning is the step that turns a raw text-predicting AI into an assistant that actually does what you ask. Straight after pretraining, a model is good at continuing text but not at following instructions. By training it on lots of examples of "here is a request, here is a good answer", it learns to be helpful and to respond to what you want, the difference between a rambling autocomplete and a useful helper.
Instruction tuning is a form of fine-tuning where a model is trained on many (instruction, good response) pairs across diverse tasks. It teaches the model the general skill of following instructions, rather than any single task.
Why it is needed
A pretrained base model predicts likely text, so asked "Translate this to French", it might continue with more English instead of translating. Instruction tuning aligns its behaviour with user intent and greatly improves zero-shot performance, doing new tasks it was not specifically trained on, just from a clear prompt.
Where it sits
It usually comes first in alignment, often followed by preference tuning (RLHF) to further refine helpfulness, honesty and safety.
Instruction tuning is supervised fine-tuning on a broad mixture of tasks framed as instruction-response pairs, which induces generalised instruction-following and markedly improves zero-shot and held-out task performance (Wei et al. 2021, FLAN). It is typically the first alignment stage after pretraining (the "SFT" step), converting a next-token predictor into an assistant, and is often followed by preference optimisation (RLHF/DPO) for finer control of helpfulness and harmlessness (Ouyang et al. 2022).
Effectiveness depends heavily on data: task diversity, instruction phrasing variety, response quality and format coverage matter more than raw quantity, and small high-quality sets can rival large noisy ones. It shapes behaviour and interaction style rather than adding substantial new knowledge, and can introduce artefacts such as verbosity or sycophancy if the demonstration distribution is skewed. Instruction tuning is what makes plain prompting reliable, and thus underpins the everyday "just ask" usability of modern LLMs.