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

# Datasets overview

> Curated collections of examples — the fuel for simulation, tests, and improvement.

A **Dataset** is a collection of examples you can run agents against. Each example is an input (a question, a task, a payload) paired with an expected output (the correct answer, or "any of these are acceptable", or just a rubric).

Datasets are how you measure agent quality objectively. Without a dataset, "did the change help" is a vibe. With one, it's a number.

## When to use datasets

* **Regression testing** — every deploy runs the dataset, catches regressions before users do.
* **Simulation** — proposed improvements are tested on the dataset, not on production.
* **Benchmarking** — compare Flow versions, model choices, prompt variants.
* **Coverage** — check that your Flow handles the range of inputs it's supposed to.

## Types of datasets

<CardGroup cols={2}>
  <Card title="Golden set" icon="star">
    Hand-curated correct examples. Small, high-quality.
  </Card>

  <Card title="Holdout" icon="lock">
    Set-aside test data — not used for training/improvement, only for evaluation.
  </Card>

  <Card title="Regression suite" icon="shield">
    Examples of previously-fixed failures. Guards against regressions.
  </Card>

  <Card title="Production sample" icon="chart-column">
    Real traces from production — check quality on real traffic.
  </Card>
</CardGroup>

Each is just a Dataset with a different purpose. Nora doesn't force separate types — but tagging your datasets by role makes them easy to use appropriately.

## How to build one

<CardGroup cols={2}>
  <Card title="Create a dataset" icon="plus" href="/reliable/datasets/create">
    From production traces, hand-authored, imports, or clusters.
  </Card>

  <Card title="Annotate examples" icon="pen" href="/reliable/datasets/annotate">
    Add expected outputs, rubrics, or verdict labels.
  </Card>
</CardGroup>

## Coverage and versioning

<CardGroup cols={2}>
  <Card title="Coverage" icon="chart-pie" href="/reliable/datasets/coverage">
    Are your examples representative of production?
  </Card>

  <Card title="Versioning" icon="clock-rotate-left" href="/reliable/datasets/versioning">
    Datasets change over time — track versions.
  </Card>
</CardGroup>

## The dataset lifecycle

1. **Seed** — small manual set of 10-30 examples.
2. **Grow** — add every fixed cluster's failing traces as new examples.
3. **Prune** — retire examples that are no longer relevant (dead features, outdated cases).
4. **Split** — as it grows, keep a holdout that never trains anything, only evaluates.

A healthy Flow has 100-1000 examples per major intent. Growth is organic — the Improvement flow adds examples every time it fixes a cluster.

## Where datasets live

Under **Reliable → Datasets** in the workspace. Each Dataset shows:

* Example count.
* Last update.
* Which Flows / Improvements have used it.
* Coverage stats (see [Coverage](/reliable/datasets/coverage)).