> ## Documentation Index
> Fetch the complete documentation index at: https://docs.captar.aurat.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Datasets

> Build append-only datasets from trace exports or file imports, then use them for manual review.

# Datasets

Datasets in Captar are project-scoped collections of rows that come from either
trace exports or imported files.

They are meant to be an organized review surface, not a generic data warehouse.
Keep each row readable enough that a reviewer can understand the example without
opening half a dozen unrelated tabs.

## Supported dataset flows

* Export a trace into a dataset row
* Create a new dataset inside a project
* Import rows from `json`, `jsonl`, or `csv`
* Export dataset rows back out for offline work

Each row can keep input, output, metadata, and source information such as trace
ID, external trace ID, span ID, and payload retention mode.

## When to export

* Export traces that show the behavior you want to preserve or compare.
* Keep representative failures, not only successful runs.
* Avoid filling datasets with noisy duplicates.
* Use metadata to explain why the row matters.