Dataset Overview

What is a Dataset?

A dataset is a table of examples. Prompts and evals run against it and write their results back as new columns. Experiments and optimization run on the same rows in their own tabs. You reach it from Dataset in the left nav.

Columns, rows, and cells

Each dataset is a grid: columns define what you’re capturing (a query, an expected answer, a score), rows are the individual examples, and a cell holds the value where a row meets a column. A column’s values come from you directly, or get filled in automatically by running something against the dataset. See Static & Dynamic Columns for the difference.

Where the data comes from

• File upload: bring in a CSV, JSON, or JSONL file
• The SDK: push rows from your own code
• Synthetic generation: describe a schema and get realistic rows back
• Hugging Face: import an existing dataset by name
• An existing dataset or experiment: branch off data you already have in Future AGI
• Observe traces: turn real production traffic into rows
• Manual entry: add rows and columns by hand

Dive deeper