Pradra StudioDocs

Datasets

A capability is only as good as the data behind it. Pradra Studio treats data as a first-class, versioned asset — never a loose file on someone's laptop.

What a dataset is

A dataset is a validated, versioned collection of the data you train and evaluate on. When you upload data (or a connector syncs it in), the platform:

  1. Validates it — checks structure and column types so problems surface early, before they cost a training run.
  2. Versions it — each upload becomes an immutable version with a fixed fingerprint.
  3. Keeps it — datasets are deprecated when retired, never silently deleted.

Because every version is immutable, a result can always be traced back to the exact data that produced it. Two people training against "version 3" are guaranteed to be using the same rows — and a model in production can always answer "what was I trained on?".

A dataset moves through: Uploaded → Validated → Versioned → Deprecated.

Working with datasets in the Studio

  1. Go to Datasets in the sidebar. The list shows every dataset in your workspace with its status and latest version.
  2. Click Create and upload a CSV. Validation runs automatically; a dataset that fails validation tells you what is wrong (missing headers, inconsistent types) instead of quietly accepting bad data.
  3. Open a dataset to see its detail page: version history, column statistics and distributions (so you can eyeball the data before trusting it), and the feature sets defined on top of it.
  4. Upload again to create the next version — the previous one stays exactly as it was, still referenced by any model trained on it.

Three ways data gets in

PathWhen to use it
Manual uploadOne-off files, experiments, getting started.
ConnectorThe data lives in Postgres, MySQL, or S3 and should sync on a schedule.
Prediction feedbackDown the line, production predictions themselves become labeled data for retraining.

All three land in the same place: a validated, versioned dataset.

Good habits

  • Name datasets after what they contain, not after today's date — the version history carries the dates for you.
  • Deprecate, don't delete. A deprecated dataset stops appearing in pickers but keeps every promise it made to the models trained on it.
  • Check the column statistics on the detail page before training: a column that is 60% empty is easier to find now than after a disappointing model.

Next