← AI Terminology

Instruction Tuning

Instruction tuning (also called supervised fine-tuning for instruction following) is the process of fine-tuning a pre-trained language model on a dataset of (instruction, response) pairs — teaching it to follow natural language commands rather than just complete text.

It is the first step that converts a raw pre-trained LLM into a useful assistant.
Why It Matters in AI
A raw pre-trained LLM continues text; it does not follow instructions. GPT-3 would complete "Write a poem about the sea" with more instructions, not a poem. Instruction tuning flips this: the model learns that an instruction is a command, not text to continue. FLAN, Alpaca, InstructGPT, and all modern chat models are instruction-tuned. It is the lowest-cost, highest-impact fine-tuning step — often achievable with thousands of examples.
Key Points
Aspect Description
SFT Supervised Fine-Tuning is the technical name — cross-entropy loss on response tokens only
FLAN Fine-Tuned Language Net: instruction tuned across 1,000+ tasks — proved massive generalisation
Scale 1,000–100,000 high-quality examples is typically enough; quality > quantity
After SFT Often followed by RLHF/DPO alignment to improve safety and preference alignment
Data format (instruction, [input,] response) triples — e.g. {"instruction": "Summarise:", "input": article, "output": summary}
Key datasets FLAN (Google, 1800 tasks), Alpaca (52K GPT-4 generated), Dolly, ShareGPT, Open-Hermes
Simple Analogy
A brilliant but socially oblivious genius (pre-trained LLM) who knows everything but answers any question by rambling about related topics. Instruction tuning is etiquette school: they learn that when someone asks a question, they should answer it — and when asked to write a poem, they should write a poem, not discuss the history of poetry.
Common Usage Examples
  • trl.SFTTrainer(model, train_dataset=alpaca_dataset, formatting_func=format_instruct_sample)
  • Alpaca format: {"instruction": "...", "input": "...", "output": "..."} — standard instruction tuning schema
  • --data_path data/alpaca_data.json in Stanford Alpaca fine-tuning script
  • Open-Hermes 2.5: high-quality instruction dataset used to fine-tune Mistral-7B to near-GPT-4 quality
  • chat_template in HuggingFace tokenizer: formats conversation history into instruction-tuning prompt format
Summary
In short: Instruction tuning teaches a pre-trained model to follow commands rather than just complete text — the transformation that turns a language model into a useful assistant.