RLHF vs SFT: what is the difference?

RLHF and SFT are two post-training methods: RLHF uses human rankings of model outputs to teach preference, while SFT uses human-written ideal answers to teach imitation. Labs usually run SFT first and RLHF after. The two methods need different data, and that data is priced differently.

What is SFT?

Supervised fine-tuning trains a pretrained model on prompt and response pairs. Humans write or edit the responses, so the model learns to follow instructions in the right format and tone. The training data is a set of SFT examples, each showing one good answer.

What is RLHF?

Reinforcement learning from human feedback trains a reward model from human comparisons, then uses that reward to guide the language model. The input is preference data: a prompt, several model responses and a human ranking. The model learns which answers people prefer, not just what a good answer looks like.

How do they differ?

Attribute	SFT	RLHF
Human input	Write or edit an answer	Rank or compare answers
Data unit	Prompt and ideal response	Prompt and ranked responses
What the model learns	Imitate	Prefer
Main risk	Narrow coverage, copied style	Reward model bias, reward hacking
Typical vendors	Expert data labs, talent marketplaces	Labs such as Surge AI and Scale AI

How do data costs compare?

SFT responses are written by hand, so each example costs expert time. Preference comparisons are quicker to produce, but reward models need many of them and consistent reviewers. In both cases, quality depends on rubrics and reviewer calibration. Read how AI training data is priced for the pricing models.

When should you use which?

Use SFT to teach new formats, domains or behaviors when you can show good examples. Use RLHF or related preference methods to fine-tune quality, tone and safety when you can compare outputs but not easily write the ideal one. Most teams use both inside post-training.