# Testing with a Public Dataset

If you want to try Starter Edition but do not have a dataset to hand, or would rather not use your own, you can use a public dataset instead. The places below are a few ideas for where to look, not the only ones.

## Before you choose one

Licences vary from dataset to dataset, so read the licence on the dataset's own page and follow everything it asks for, such as attribution. Some licences allow only personal or non-commercial use, which can rule out testing at work. A copy uploaded by someone else can carry a different licence from the original, so where you can, check the original source too.

Where a site has its own terms of use, check those as well as the dataset's licence.

The rule on personal, confidential and sensitive data applies to a public dataset as well, and that includes health data, financial data and anything similar. None of the data values, table or column names, descriptions or the context you give your agent while authoring a contract may contain it (see Legal and Data Use).

Choose a dataset that holds no data about individual people, including pseudonymous IDs such as user or customer numbers. If a dataset has any, choose a different one.

> [!WARNING]
> **Treat a downloaded dataset as untrusted.**
>
> Where you can, download from the original publisher over HTTPS, and on sites where anyone can upload, such as Kaggle, check who uploaded the data. Use only data files such as CSV, and never run or load anything else that comes with a dataset, or enable macros in a spreadsheet.
>
> Text inside a dataset can contain instructions aimed at your AI agent, and no setting is known to prevent this fully. While your agent works with downloaded data, keep it asking for your approval before it runs commands or edits files, and read each request before you approve it.
>
> These steps lower the risk but do not remove it.

## Where to find public data

Some example sites with public data, among many others:

- **[data.gov](https://catalog.data.gov/):** the US government's open data catalogue, run by the General Services Administration. Not every entry states a licence, so check with the publisher where it does not.
- **[data.gov.uk](https://www.data.gov.uk/):** UK public data, from the National Data Library. Check the licence shown for each dataset, or on the publisher's site it links to.
- **[UCI Machine Learning Repository](https://archive.ics.uci.edu/):** a collection of datasets for machine learning, maintained at the University of California, Irvine. Check the licence on each dataset's page.
- **[Kaggle Datasets](https://www.kaggle.com/datasets):** a large collection uploaded by Kaggle users, which you can filter by licence type. Kaggle's terms of use allow use of the site only "for your own internal, personal, non-commercial use", so check they suit your purpose. Some datasets are re-uploads of data published elsewhere, so check where the data came from.
- **[Google Dataset Search](https://datasetsearch.research.google.com/):** a search engine for datasets held in thousands of repositories across the web. Check the licence on the page where each dataset is held.