Instruction tuning
The step that turns a model which continues text into a model that does what it is asked.
Definition
Instruction tuning is supervised fine-tuning on instruction-and-response pairs, done so that a model learns to follow directions in general rather than to perform one task. The training set is deliberately heterogeneous — summarize this, classify that, rewrite this in plain language, answer this question, extract these fields — on the premise that a model exposed to enough varied instructions generalizes to instructions it has never seen. It is the step that converts a model which continues text into a model that does what it is asked.
The technique came out of work showing that multi-task training on natural-language task descriptions produced large gains on unseen tasks, and it now anchors the first stage of post-training almost everywhere. Its practical properties are reasonably well understood: diversity of instruction type matters more than sheer count; response quality sets the ceiling on style and on factual habits; and the format of the training pairs — how a system prompt, a user turn and an assistant turn are marked — carries over into the deployed model's behaviour.
It is often confused with alignment, and it is not the same thing. It teaches a model the shape of compliance with instructions; it does not by itself teach a model which instructions to refuse, nor make it accurate on material the demonstrations never covered. Refusal behaviour, appropriate expressions of uncertainty and fine-grained stylistic preference are normally shaped afterwards by preference-based methods. And where the instructions concern specialist work, the same constraint as any supervised stage applies: an instruction-and-response pair written by someone who does not know the domain teaches the model to answer confidently and incorrectly.
How DeepenSkill approaches it
Writing instruction-and-response pairs that hold up in a specialist domain needs someone who works in it, and that is the talent DeepenSkill places. Everyone clears the same bar before starting: credentials verified, a standardized domain test passed, then a practical task scored against gold reference data. On live work the team measuring the experts is separate from the team running them — the platform rejects a validation report authored by the party that delivered the work — agreement statistics are reported with the method named, the item and rater counts stated and a 95% interval attached, and every judgment traces to a verified contributor, a task spec and a timestamp. See expert SFT work, the preference stage that follows it, and a sample verification bundle you can recompute yourself.
FAQ
What is instruction tuning?
Supervised fine-tuning on a deliberately varied mix of instruction-and-response pairs, so a model learns to follow directions across tasks rather than to perform one task well.
Is instruction tuning the same as SFT?
Instruction tuning is a form of supervised fine-tuning. SFT is the general technique of training on prompt-and-response pairs; instruction tuning is SFT on a varied instruction mix aimed at general direction-following.
Does instruction tuning make a model aligned?
No. It teaches a model to follow instructions, not which instructions to refuse or how to express uncertainty. Those behaviours are normally shaped afterwards by preference-based methods such as RLHF.