> ## Documentation Index
> Fetch the complete documentation index at: https://arize-ax.mintlify.site/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Curate a Dataset from Traces

> Save production spans as dataset examples for experiments, fine-tuning, and few-shot prompts

You need a test dataset, but making one from scratch is painful. Your best examples are already in your production traces — the edge cases your users actually hit, the queries that tripped up your model, the responses that were perfect. Arize AX lets you turn those into a dataset directly, no pipeline needed.

## How to do it

**Individual:** Open a span → click **Add to** → **Dataset** → choose or create a dataset.

**Bulk:** Select multiple spans in the traces table → **Add to** → **Dataset**.

**Using Alyx:** *"Create a dataset from the filtered spans"* or *"Add these error traces to my test dataset"*

\[screenshot: add to dataset dialog]

## Common workflows

| Workflow                     | How                                                                                  |
| ---------------------------- | ------------------------------------------------------------------------------------ |
| **Test set from production** | Filter to edge cases → Add to Dataset → Run experiments                              |
| **Few-shot examples**        | Find high-quality responses → Add to Dataset → Reference in prompts                  |
| **Fine-tuning data**         | Filter for correct responses → Add to Dataset → Export                               |
| **Human-in-the-loop**        | [Labeling Queue](/docs/ax/observe/take-action/labeling-queue) → Annotate → Create Dataset |

## What gets saved

Each span's **input**, **output**, and **metadata** are added as a dataset example — everything you need to replay the scenario in an experiment.