SFT datasets

SFT (supervised fine-tuning) is a Language type for fixed prompt-and-target pairs on advanced LLM experiments.

Create

New dataset → Language → SFT → add data on Files (CSV or Hugging Face).

Map columns on Data

On Data, use Select Prompt and Target Columns:

• Prompt

• Target (the completion the model should imitate)

Save both mappings before you leave the dataset or continue the experiment wizard. The Finish control on the wizard’s dataset step expects saved columns.

Experiments

For Advanced Training projects only. The Agent step offers SFT when this dataset type is attached. Any plan that includes Advanced Training can use SFT datasets. Tabular and non-tabular types additionally require an Enterprise plan.

Preference vs SFT

If you have chosen and rejected pairs, use Preference. If you only have prompt and target text, use SFT. The Language type switcher notes that preference rows can sometimes feed SFT-style work, but Arena keeps the types separate.

See also

• Supervised training (algorithm SFT on the agent step)

• Datasets overview