Quality Beats Quantity
Lesson 3 of 4 in Supervised Fine-Tuning.
How many demonstrations do you need? The instinct from pre-training — more data, better model — turns out to mislead here, and one paper became the emblem of why. LIMA (Zhou et al. 2023) fine-tuned a 65-billion-parameter base model on just 1,000 carefully curated demonstrations — no preference training at all — and in human preference comparisons its responses held up remarkably well against assistants built with far larger post-training pipelines. The authors read this as evidence for what they call the superficial alignment hypothesis: nearly all of a model’s knowledge and capability comes from pre-training, and Alignment-stage tuning mostly teaches which format and style to use when engaging users.
Treat that as a finding under its stated conditions, not a law. One base model at one scale, one curation effort, one evaluation protocol built on human preference judgments — those caveats are real, and later work has complicated the picture in both directions. But the direction of the lesson has held up broadly across the field: past a modest scale, the marginal curated example beats the marginal scraped one by a wide margin, because the model is not learning facts from SFT data — it is learning a behavioral target, and a noisy target is a worse target.
Bar chart with three bars showing an illustrative human-preference score for three fine-tuning datasets: one thousand excellent curated demonstrations scores 78, twenty thousand mixed-quality demonstrations scores 71, and two hundred thousand noisy scraped demonstrations scores 60. The bars illustrate that curation can beat scale for SFT data.
“Quality” sounds like a platitude until you make it operational. In an SFT dataset it decomposes into three properties you can actually inspect and enforce — and a fourth that goes without saying (the responses must be correct, since every error in the set is a demonstrated behavior).
Coverage — the prompts span what the assistant will face
Diversity of prompts matters more than volume of responses. A thousand near-duplicate “summarize this article” pairs teach one behavior a thousand times; the same budget spread across tasks, formats, difficulty levels, and edge cases (ambiguous requests, impossible requests, requests that deserve a Refusal) teaches the distribution of situations. Gaps in coverage become gaps in behavior — a model shown no examples of declining gracefully will improvise when it has to.
Consistency — the responses agree on what “good” looks like
The model fits all the demonstrations. If half your examples answer in tight bullet points and half in leisurely essays — or worse, if annotators disagreed about when to hedge and when to commit — the model learns a blur of the styles, not the best of them. This is why serious annotation efforts run on style guides and inter-annotator review: consistency in the data is what becomes reliability in the model.
Tone — every demonstration is a vote for a persona
The assistant’s voice — how formal, how warm, how it opens and closes answers, how it says no — is not configured anywhere. It is averaged out of the demonstration set. If you have ever noticed a model’s signature verbal habits, you have been reading its SFT data at one remove. Curating tone means reading responses the way an editor reads copy, and it is a large share of what the “quality” in quality-beats-quantity actually paid for.
Interactive checkpoint quiz (1 questions) — open this page in a browser to take it.