Generate realistic data.
Export to any format.

There are 152, grouped into 19 categories in the picker, from the ordinary ones like integer, email and city through to ICD-10 codes, IBANs and MIME types. Two controls matter more than the count. An allowed-values list you type wins over the field type's own values, whatever the type, so a column can stay typed as state and still be held to the four states your test covers. And a date column has a range mode: switch it from an absolute range to relative to today with an offset in days, and the window moves with every run instead of your fixtures aging past the dates you pinned in March.

Point them at the same reference data. Every state, city and postal code triple comes out of the GeoNames postal table, all 50 states plus DC, and there is no invented fallback: a combination the reference data does not hold comes back empty and the dataset says which column and what to change, rather than shipping a postal code that belongs to another city. The same shape covers sector to industry to sub-industry, and an ICD-10 code with its description, which are read off one row so they agree. For anything arithmetic there are rules: when status is refunded, refund_amount is total times 0.9. A rule set makes you say what happens when nothing matches, either a catch-all or an explicit null, so the column cannot quietly come back empty.

Set the column to normal, log normal, exponential or triangular, and give it the numbers in the units you already think in. A log normal salary column takes a mean of 140,000 and a spread of 60,000, not a mu and a sigma in log space; the conversion happens for you. The column's own min and max then clip whatever the distribution produced, which is the thing to know before you centre a normal near its minimum: the tail that would have gone below the bound arrives sitting exactly on it.

Switch a child table from a fixed row count to per parent row and give it a range, for example 8 to 12 patients per doctor. One multiplier is drawn from that range for the relationship rather than for each doctor, and it replaces whatever row count was typed for that table. Ranges resolve in foreign key order, so doctors to patients to visits cascades, and two parents feeding the same junction table add up. How the children spread across parents is a separate setting: the default picks a parent at random for each child row, which leaves some doctors with no patients at all, so set the foreign key's sampling strategy to uniform when every parent has to be referenced the same number of times.

Every dataset carries a seed. Send the same seed with the same schema and the same row counts and the rows come back the same, which is what makes generated data worth committing as a fixture; leave it blank and one is allocated and recorded, so you can repeat a run you did not plan to repeat. UUID columns are the exception and are fresh every time. The form also pre-fills from your last finished run for that schema, row counts, ranges, seed and locale together, and those live on the schema rather than in your browser, so they follow you to another machine.

One archive per dataset, split by format: a csv folder holding orders.csv and one file for every other table, a json folder holding orders.json or orders.jsonl, a parquet folder, and a single nested JSON file when you asked for embedded documents. Ask for CSV and you also get one loader script per SQL dialect, 7 of them, each dropping tables in reverse dependency order, creating them in dependency order, running that dialect's own bulk load command over the CSVs, and adding the foreign keys either inline or at the end, whichever the dialect prefers. The UNIQUE constraints in those scripts are decided by reading the CSV the script is about to load rather than by what the schema intended, so a script never declares a UNIQUE its own rows would break; where a column turned out not to be unique the constraint comes off and the dataset tells you which one.