Reasoning datasets

Reasoning is a Language type for question-and-answer columns used in GRPO-style advanced experiments (reward functions, prompting, and related wizard steps).

Create

1. New dataset → Category Language → Type Reasoning

2. Files — Upload a CSV or import from Hugging Face

3. Finish the modal; Arena opens the dataset on Data

Map columns on Data

Open the dataset and stay on Data. The heading is Select Question and Answer Columns.

Click the column badges to assign:

• Question

• Answer

The preview table paginates so you can check rows before you save. Click a selected badge again to clear that column.

Map both columns before you train. Next on the wizard stays disabled until both columns are saved.

Experiments

Use reasoning datasets only in Advanced Training projects. After the Environment step attaches this dataset, later steps cover prompting, reward functions, and LLM training settings. Any org member on a plan that supports Advanced Training can create and use reasoning datasets. Reasoning datasets do not require an Enterprise plan.

See also

• LLM algorithms

• Datasets overview