What is synthetic data?

Two ways to generate synthetic data

Start from a request when you know the schema and scenarios you need but do not have a source dataset. Describe what one row represents, the columns, category rates, missing values and relationships. Ground designs a dataset and shows 25 sample rows for review before you export.

Start from an existing dataset when you need new rows that reflect observed patterns. Ground profiles the source and fits a synthesizer to its columns and relationships. You can generate more rows, rebalance a label or focus on a thin segment. The result is an approximation of the source, so evaluate it against the task you intend to use it for.

When is synthetic data useful?

For software testing, you can specify edge cases and generate enough records to exercise a pipeline. For model development, synthetic rows can help explore scarce classes or training scenarios. Research teams can use generated datasets to develop analysis code before access to real data is available.

A designed dataset reflects your assumptions; it does not establish facts about a real population. A source-based dataset can inherit gaps and biases from the source. Neither approach replaces evaluation on appropriate real data when the downstream result depends on real-world performance.

Create your first dataset in Ground

A useful request states the unit of observation, expected volume and the relationships that matter. Review values across the same row, rather than checking each column in isolation.

1. Describe the dataset or upload a supported source file.
2. Review 25 rows. Refine your request or point to a cell that needs a change.
3. Check label rates, missing values and relationships against your intended use.
4. Export as CSV, JSON Lines or Parquet. Ground supports up to 10 million rows per dataset, with lower limits for AI-written text.

Example request

Create 10,000 subscription accounts. Each row is one account with plan, monthly spend, signup date and churn status. Make 8% churned and let churn vary by plan. Include missing values in optional fields.

Is synthetic data automatically anonymous?

No. Generating new rows does not, by itself, establish anonymity or regulatory compliance. Ground replaces detected personal-data columns, suppresses rare categories and checks for rare copied records in its default synthesis workflow. The quality and privacy report helps assess remaining risk.

Optional differential privacy is a separate fitting mode with an explicit privacy budget. Choose controls based on the data, the recipients and the release purpose. Read the security page and review the dataset report before sharing outputs.

How is synthetic data different from random test data?

Independent random values can satisfy a schema while breaking relationships. Synthetic data can encode distributions, label rates and dependencies. You still need to verify that those relationships match your intended use.

Do I need to upload real data to use Ground?

No. You can describe a dataset in plain English and review 25 generated rows. Upload a source dataset when you want Ground to learn its observed patterns.