Synthetic data for research labs and data teams
Synthetic data that behaves like the real thing
Ground is a synthetic data generator for research labs and data teams. Describe a dataset or extend one you already have. Review 25 sample rows, refine, then export realistic data as CSV, JSON Lines or Parquet.
| # | timestamp | amount_usd | merchant | merchant_category | channel | state | is_fraud |
|---|---|---|---|---|---|---|---|
| 1 | 2025-04-09T17:36:34Z | 754.65 | Crona Group | grocery | card_present | MI | false |
| 2 | 2025-02-21T00:07:47Z | 713.67 | Anderson-Lubowitz, Nolan and Aufderhar | electronics | online | CA | false |
| 3 | 2025-06-19T12:14:40Z | 531.94 | Prohaska, Quigley and Abernathy | other | card_present | PA | false |
| 4 | 2025-03-13T18:53:16Z | 4000.00 | Reichert - Watsica | online_retail | card_present | NC | false |
| 5 | 2025-06-01T09:39:15Z | 173.39 | Stroman, Raynor and Vandervort | online_retail | online | NY | false |
| 6 | 2025-02-05T10:15:44Z | 128.14 | Kreiger - Keeling | grocery | card_present | WA | false |
| 7 | 2025-02-27T09:26:54Z | 4000.00 | Weissnat - Sawayn | restaurants | card_present | MN | false |
| 8 | 2025-04-09T12:21:04Z | 924.49 | Grant, Conn and Larkin | other | online | IL | false |
| 9 | 2025-01-26T04:50:48Z | 925.61 | Ryan Group | online_retail | online | FL | false |
For teams who can't wait for, can't share or can't collect the real data
Start with the job you need to do
For product & engineering teams
Test an application
Build reproducible test fixtures for ingestion, dashboards and application workflows. Start with 1,000 orders, inspect the rules, then adapt the schema to your product.
Explore the workflowFor ML & data teams
Exercise an evaluation workflow
Test splitting, metrics and ingestion with a reproducible rare-label fixture. Inspect 10,000 transactions with explicit channel-dependent probabilities before adapting the scenario.
Explore the workflowFor teams with an existing dataset
Extend training data
Upload a table, inspect what Ground learns and generate additional synthetic rows. Explore the workflow first with a deterministic sample of 3,000 fictional loan applications.
Explore the workflowSee it before you generate it
Describe it
Write your request in plain English: what each row is, the label rates you need, the quirks of the real data. Ground designs every column.
Review 25 rows, then refine
Edit the request or click any cell to say what's off. The preview updates, and every revision can be undone.
“Patient presents with chest pain…”
Generate at size
Up to 10 million rows run in the background. Your file starts with the exact rows you approved, and you download it as CSV or JSON Lines.
72%
Measured, not claimed
93/100 on six public datasets
Every method trains on the same 80% of each table and is scored on the same unseen 20%: column shapes, relationships and a classifier trying to tell real from synthetic. MOSTLY AI averages 83 and SDV 66.
Every number, method and caveatOverall score, 0–100, from Ground's own report; independent scoring is on our roadmap. MOSTLY AI wins on adult; differences of a point or two are within noise.
Extend your data
Already have data? Get more of it.
Upload a CSV or Parquet file. Ground learns each column and how the columns relate, per class of your label, then generates new rows: more of the same, more of a rare class or segment, or the months after your data ends.
Extend a datasetTen real rows, ten generated, shuffled. If you can't tell them apart, neither can a reviewer.
Raise a 3% label to 25%, add rows for a thin segment, or extend six months forward, in one click.
Detected personal data is replaced, rare record copies are checked, and the report measures remaining privacy risk.
Fidelity, a real-vs-synthetic classifier test, train-on-synthetic results and privacy metrics, as a shareable data card.
Realistic enough to train on
Coherent rows
Names, emails, employers and locations agree within each row, and AI-written text matches the values next to it.
Real-world mess
Long tails, skewed categories, rare labels at believable rates and missing values where real systems have gaps.
Reproducible
Every dataset is a saved spec and a seed. Regenerate the same rows next quarter, or draw a fresh sample.
Built for size
Up to 10 million rows per dataset, generated in the background and streamed straight to a file.
Every column your data needs
Mix rule-based columns, which are fast and free, with AI columns for text that has to make sense.
Names, emails, companies and places that agree with each other in every row.
Whole numbers in a range.
Uniform, normal or long-tailed distributions for prices, scores and revenue.
Flags with the exact rate you need, down to rare 2% fraud labels.
Weighted values, so your plan mix and order statuses look real.
Dates and timestamps across any window.
Unique identifiers, stable for a given seed.
Readable IDs like ORD-100231.
Formulas built from other columns, like emails from names.
Free text written by Claude that matches the rest of the row.
Start from a template
Customer questions with intents, difficulty and reference answers.
Orders with skewed prices, statuses and optional reviews.
Inbound tickets with realistic subjects, messages and priority.
Payments with a rare fraud label for model training.
De-identified-style clinical encounters with notes.
HRIS-style roster with salaries, tenure and reviews.
Put synthetic data to work
Practical guides for your first dataset, testing pipelines, training models and checking the results.
What is synthetic data?
Learn what synthetic data is, when to generate it from a prompt or existing data, and how to evaluate realism, utility and privacy in Ground.
Read guide →How to generate synthetic test data
Build realistic test datasets for pipelines and applications with reproducible seeds, missing values, rare cases and related tables in Ground.
Read guide →Using synthetic data for machine learning
Extend training datasets, rebalance rare labels and evaluate synthetic data against real holdout data with Ground's quality and privacy report.
Read guide →How to evaluate synthetic data quality
Evaluate synthetic datasets with column fidelity, relationships, detection AUC, TSTR and privacy checks. Learn what Ground's report measures and what it cannot prove.
Read guide →