A thousand examples, carefully chosen

figure 1 · the hypothesis

The knowledge is already in there

A base language model has read trillions of tokens. Somewhere in its weights it already knows how to plan a trip, explain a proof, write a cover letter. What it does not know is that it is supposed to behave like an assistant — answer the question, in a helpful register, and stop. LIMA’s claim, stated as the Superficial Alignment Hypothesis: a model’s knowledge and capabilities are learnt almost entirely during pretraining, while alignment merely teaches it which subdistribution of formats to use when talking to a user.

click any figure to zoom · scroll to magnify · drag to pan

If that is true, alignment is a small lesson, not a second education — and a small, perfect dataset should be enough to deliver it. The paper is one long test of that corollary.

figure 2 · the dataset

750 borrowed, 250 written by hand

The entire training set is 1,000 prompt–response pairs — about 750,000 tokens. Three quarters are mined from community sites where good answers already exist: Stack Exchange, wikiHow, and Reddit’s WritingPrompts. The last quarter the authors wrote themselves, deliberately, to stamp one uniform assistant voice across the set. A further 300 prompts are held out as the test set — the model never sees them.

The hand-written 250 are the quiet trick. Community answers bring knowledge and variety; the authored examples bring consistency of persona. The mix is the point: diversity in the prompts, uniformity in the voice.

figure 3 · the filter

Millions of threads in, a thousand answers out

“Curated” is doing heavy lifting, so here is what it meant. From millions of Stack Exchange threads: keep only highly-voted answers to self-contained questions; drop answers shorter than 1,200 or longer than 4,096 characters; strip links, HTML, and first-person asides; sample across topics rather than deep within one. What survives is rewritten where needed into one calm, helpful register.

The funnel is the work. Nothing in it is clever machine learning — it is editorial judgment applied a thousand times, which is exactly why so few teams do it.

figure 4 · worked by hand

Five parts in ten million

It is worth feeling the ratio. LLaMA 65B pretrained on roughly 1.4 trillion tokens. LIMA’s alignment set is 750 thousand. Divide them and the entire alignment lesson is about 0.00005% of what the model has read — five parts in ten million. For contrast, Alpaca fine-tuned on 52,000 examples (52× more), and InstructGPT’s pipeline consumed tens of thousands of demonstrations and rankings plus a reinforcement-learning stage.

results · human preference

Holding its own against the RLHF heavyweights

The test: show human judges a prompt with two responses — LIMA’s and a rival’s — and ask which is better, or whether they tie. LIMA, with its thousand examples and no preference tuning of any kind, beats Alpaca 65B (trained on 52× more data) and OpenAI’s RLHF-trained DaVinci003. Against the era’s product models it loses the match but wins a startling share of rounds.

Read the last row carefully: on 43% of prompts, judges found the 1,000-example model at least as good as GPT-4. Absolute grading agreed — 88% of LIMA’s responses met the prompt’s requirements, half rated excellent. That is the hypothesis surviving contact with the evidence.

figure 6 · the ablations

Doubling the data did nothing

The ablations separate the three ingredients. Quality: training on filtered versus unfiltered Stack Exchange answers is worth half a point on a six-point grading scale. Diversity: heterogeneous Stack Exchange prompts beat wikiHow’s uniform how-to formula at the same size and quality. Quantity: doubling the training set from 2,000 examples — and doubling it again, and again, up to 16,000 — produced no improvement at all.

the shape of the finding quantity is the only knob that costs nothing to turn, and it is the one that does nothing. The two that work — quality and diversity — are exactly the ones that demand human judgment per example. “More data” is not a data strategy.

figure 7 · the limits

Where the seams show

The authors are candid about the edges. Authoring uniform, excellent examples is laborious and does not scale like scraping does. LIMA is not robust: an unlucky sampling seed or an adversarial prompt can produce a weak answer where a product model would recover. And on a small safety probe of 30 sensitive prompts, it responded safely only 80% of the time — style transfer alone does not install guardrails.

This is precisely the gap the rest of the arc fills: entry 15 shows what the full human-feedback pipeline buys, and entry 16 shows how cheaply its core can now be had.

the equations · worked

The only loss in the paper, and how the verdict is counted

applied · our reading

The data bar for a domain fine-tune

LIMA is the first paper we hand anyone who says “we need ten thousand training examples.” Three things we carry into our own fine-tuning work:

• Budget for curation, not collection. A few hundred expert-written, style-uniform examples are a real alignment dataset. The cost centre is editorial time from people who know the domain — not scraping, not labeling farms.
• Diversify the prompts, unify the voice. LIMA’s split — varied community questions plus a hand-authored core that fixes the persona — is a recipe. For a domain assistant: real user questions for coverage, expert-authored answers for tone.
• Let the plateau save you money. Before commissioning example 10,001, re-read the scaling ablation: past a small, well-chosen core, quantity bought nothing. If your fine-tune is weak, the fix is almost never “more of the same.”

the caveat we would flag LIMA aligned a 65B base model — a big vessel with plenty of latent capability to surface. A small model with thin pretraining has less to reveal, and a thousand examples cannot conjure knowledge that was never there. That failure mode — fine-tuning to inject facts — is the most expensive mistake in applied fine-tuning, and no amount of curation rescues it.