Concept
Supervised fine-tuning
Version 1 — First definition, written for the AfterQuery Terminal-Bench analysis (frontier-blogs v1).
Supervised fine-tuning is continued training of a pretrained model on curated demonstrations — input paired with the output you want — so that the demonstrated behaviour becomes the model’s default. It is imitation, not trial and error: the model never acts and never gets a reward; it learns what doing the task well looks like from examples of it done well, which makes the curation of those examples the whole game.
From the sources
In the standard post-training recipe SFT comes first, and the ordering is load-bearing rather than conventional: reinforcement learning needs occasional successes to amplify, and a base model too weak to ever succeed gives it nothing. AfterQuery’s Terminal-Bench write-up states the dependency exactly — the RL stage “only works because the SFT checkpoint is already capable enough to sometimes solve the training tasks,” or as they compress it, “SFT makes meaningful RL possible” — and attributes the larger share of their gain (3.1% → 13.5%, before RL’s 13.5% → 17.0%) to this stage, “which is entirely a function of data quality.” That attribution is also the claim a data-curation company most benefits from; the mechanism is standard, the proportion is theirs.
Used by
A data vendor's account of NVIDIA using its tasks to lift Nemotron 3 Ultra on GDPval. What the claims are, and what they rest on.
A lab's own account of taking gpt-oss-20b from 3.1% to 17.0% on Terminal-Bench 2.0. What the claims are, and what they rest on.