Term
Instruction Tuning
Instruction tuning is a variant of supervised fine-tuning in which a language model is trained on a broad mix of instruction-response pairs — so it reliably follows instructions across arbitrary tasks.
Instruction tuning — explained in more detail
Classic SFT often trains a model for a narrow use case — customer support, coding assistant, medical QA. Instruction tuning pursues the opposite: pack as many task types as possible into a single training dataset so the model learns that the input is always an instruction to be carried out — whether translation, classification, summarisation, code generation or brainstorming. The approach became popular through Google’s FLAN and OpenAI’s InstructGPT paper. Datasets like FLAN, Self-Instruct, Alpaca or Dolly contain tens of thousands to millions of heterogeneous instruction-response pairs. The result: the trained model follows tasks that never appeared verbatim in the training data — the famous zero-shot generalisation.
Example / Practical context
Open-source model families make this very visible: Llama-Instruct, Mistral-Instruct, Qwen-Instruct are the instruction-tuned variants of their respective base models. Anyone running a local RAG or agent stack practically always picks an instruct variant — base models without further post-training are not capable of following instructions cleanly.
Distinction from related concepts
SFT is the umbrella term — instruction tuning is a specific data recipe of it, characterised by high task diversity. RLHF/DPO/ORPO run after instruction tuning and adjust the model further based on human preferences. Domain fine-tuning is the opposite of instruction tuning: a model is deliberately narrowed to a single field (law, medicine), often at the cost of general-purpose abilities.
Entdecke mehr
AI Workflows by Keyword: How We Make Recurring Routines Enforceable
A typed keyword triggers a fixed AI routine — and every single step must be committed before the next one appears. Why that's the actual trick.
GlossarDPO (Direct Preference Optimization)
DPO is a fine-tuning method that aligns language models directly to human preference pairs — without a separate reward model and without reinforcement learning.
LexikonFine-Tuning Explained — When, How, With What
When fine-tuning pays off, which methods exist (SFT, DPO, LoRA, QLoRA), and what AMD vs. NVIDIA hardware actually means in practice.