Instruction Tuning
Training a text predictor to follow instructions — the step that turned an autocomplete into an assistant, and it may take only a thousand examples.
When not to use it
- On a model that's already instruction-tuned. Which is all of them. You'll degrade what's there.
- To add knowledge. It teaches format and style. Use RAG or task-specific fine-tuning.
- With a huge mediocre dataset. LIMA: 1,000 good examples beat volume. Quality and diversity dominate.
- With verbose or tic-laden examples. The model learns the style very literally and permanently.
Reach for something else instead
- Prompting — if you want a different format, ask.
- Task-specific fine-tuning — if you want different behaviour on your task, train on your task.
- RAG — if the gap is knowledge.
- Preference training (DPO) — if you want better responses, not just responsive ones.
Sources & further reading
- Wei et al. (2021), Finetuned Language Models Are Zero-Shot Learners — FLAN; instruction-following generalises to unseen tasks.
- Ouyang et al. (2022), Training language models to follow instructions with human feedback — InstructGPT; instruction tuning plus preference training, the recipe everything uses.
- Zhou et al. (2023), LIMA: Less Is More for Alignment — 1,000 examples, and the Superficial Alignment Hypothesis.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Reaching for it when prompting would do. It's already been done to your model.
- Assuming more data is better. A thousand diverse, high-quality examples is the finding.
- Not noticing that response style is inherited wholesale, including the tics.
- Expecting it to add capability. The evidence says it mostly selects a mode that pretraining already built.
At a glance
FieldLanguage & LLMs
What it doesturns a text predictor into something that answers
The finding~1,000 curated examples is competitive (LIMA)
The hypothesisalignment is superficial; capability is all pretraining
What it teachesformat and style, very literally
DifficultyIntermediate
Flashcards for this concept
Question
—
Answer
—
1 / 4
Often compared with
Instruction tuning vs. pretraining — one gave the model everything it knows; the other taught it to answer when asked. LIMA's claim is that the second is a thousand examples of style.