Supervised Fine-Tuning (SFT)

What is SFT?

Supervised Fine-Tuning is behavior cloning for language models: take a pretrained next-token predictor, train it further on (instruction → ideal response) pairs, and it learns to be an assistant rather than merely continue text. Every chat model you’ve used went through SFT before any preference tuning; it’s the stage that installs the format, the helpfulness reflex, and the “respond, don’t continue” behavior.

What SFT can and cannot teach

SFT excels at anything demonstrable: output formats, tone, tool-call syntax, domain conventions, refusal phrasing. It struggles with anything comparative — SFT only ever sees “here is a good answer,” never “this answer is better than that one,” so it can’t calibrate between two plausible responses. That’s precisely the gap DPO and RLHF-family methods fill afterward. SFT also inherits a subtle flaw: it trains on gold contexts (teacher forcing), so the model never practices recovering from its own mistakes — one reason long generations drift.

Data beats everything

The consistent empirical finding: a few thousand excellent, diverse examples outperform hundreds of thousands of mediocre scraped ones (the LIMA result made this famous with ~1,000 examples). Practical corollary for teams: your SFT budget belongs in curation and deduplication, not volume. One contradictory or sloppy example teaches sloppiness as effectively as a good one teaches quality.

What people get wrong

• Using SFT to inject facts. It shapes behavior; knowledge injection remains unreliable — that’s retrieval’s job.
• Skipping the held-out eval. Without one, you can’t distinguish “learned the skill” from “memorized my 800 examples.”
• Over-training: excessive epochs on narrow SFT data measurably erodes general capability (catastrophic forgetting); mixing in general instruction data is the standard antidote.

Primary source: LIMA: Less Is More for Alignment (Zhou et al., 2023)