Ground, SDV and MOSTLY AI: reading the synthetic data benchmark

The published run compares source-based synthesizers on six public datasets. Ground has a higher overall score than the listed alternatives on five datasets; MOSTLY AI has the higher score on Adult. These are Ground's own measurements from one run, with limited training budgets. They do not establish a general ranking of the products.

Written by the Ground team · Source run: 2026-10-20 · Training cap: 30,000 rows

Compare the recorded results

Every number below comes directly from the same results file used by the full benchmark page. This walkthrough does not report a new experiment. Overall combines column shapes, column relationships and detection; usefulness is a separate train-on-synthetic, test-on-real measurement.

All recorded synthesizer results by dataset
Dataset	Recorded method	Overall / 100	Real-trained utility	Synthetic-trained utility	Fit seconds
adult	Ground	94	0.904 AUC	0.897 AUC	100.9
adult	Ground (DP ε=3)	77	0.904 AUC	0.874 AUC	2.3
adult	Ground (DP ε=3, MST)	75	0.904 AUC	0.884 AUC	0.4
adult	MOSTLY AI (TabularARGN, local)	95	0.904 AUC	0.897 AUC	235.4
adult	SDV CTGAN (50 epochs)	67	0.904 AUC	0.865 AUC	168.9
adult	SDV GaussianCopula	65	0.904 AUC	0.867 AUC	6.9
bank	Ground	90	0.919 AUC	0.913 AUC	110.8
bank	Ground (DP ε=3)	64	0.917 AUC	0.869 AUC	4.1
bank	Ground (DP ε=3, MST)	65	0.919 AUC	0.880 AUC	0.6
bank	MOSTLY AI (TabularARGN, local)	65	0.919 AUC	0.896 AUC	231.9
bank	SDV CTGAN (50 epochs)	60	0.919 AUC	0.722 AUC	190.6
bank	SDV GaussianCopula	59	0.919 AUC	0.882 AUC	9.9
credit	Ground	91	0.833 AUC	0.827 AUC	32.5
credit	Ground (DP ε=3)	63	0.835 AUC	0.755 AUC	0.7
credit	Ground (DP ε=3, MST)	66	0.835 AUC	0.776 AUC	0.1
credit	MOSTLY AI (TabularARGN, local)	79	0.833 AUC	0.823 AUC	153.5
credit	SDV CTGAN (50 epochs)	61	0.833 AUC	0.456 AUC	20.7
credit	SDV GaussianCopula	63	0.833 AUC	0.801 AUC	2.5
telco	Ground	92	0.833 AUC	0.829 AUC	34.3
telco	Ground (DP ε=3)	64	0.833 AUC	0.804 AUC	0.5
telco	Ground (DP ε=3, MST)	69	0.833 AUC	0.810 AUC	0.1
telco	MOSTLY AI (TabularARGN, local)	90	0.833 AUC	0.821 AUC	139.6
telco	SDV CTGAN (50 epochs)	64	0.833 AUC	0.610 AUC	1350.5
telco	SDV GaussianCopula	65	0.833 AUC	0.810 AUC	2.5
housing	Ground	95	0.688 R²	0.628 R²	53.9
housing	Ground (DP ε=3)	75	0.688 R²	0.533 R²	0.7
housing	Ground (DP ε=3, MST)	76	0.688 R²	0.397 R²	0.1
housing	MOSTLY AI (TabularARGN, local)	93	0.688 R²	0.607 R²	145.7
housing	SDV CTGAN (50 epochs)	53	0.688 R²	0.058 R²	73.4
housing	SDV GaussianCopula	60	0.688 R²	0.436 R²	2.4
insurance	Ground	98	0.871 R²	0.813 R²	12.3
insurance	Ground (DP ε=3)	80	0.871 R²	0.643 R²	0
insurance	Ground (DP ε=3, MST)	82	0.871 R²	0.726 R²	0
insurance	MOSTLY AI (TabularARGN, local)	77	0.871 R²	0.019 R²	47
insurance	SDV CTGAN (50 epochs)	68	0.871 R²	0.088 R²	7.4
insurance	SDV GaussianCopula	82	0.871 R²	0.507 R²	0.5

Where another method scores higher

On Adult, MOSTLY AI scores 95 overall and Ground scores 94. Both record synthetic-trained utility of 0.897 AUC against the real-trained baseline of 0.904. The one-point overall difference is within the run's stated noise. This is a useful reminder to inspect task utility rather than infer a universal winner from one summary number.

What the setup controls

For each dataset, the harness uses a fixed-seed 80/20 split. Fitting uses the training portion, capped at 30,000 rows, and utility is evaluated on the held-out real portion. Methods are scored using Ground's report. Re-run the benchmark for the exact settings and data you intend to use.

What limits the comparison

The report is Ground's own. Overall is 35% column shapes, 35% column pairs and 30% detection; it doesn't include Rules, which checks the rules Ground detects and enforces, so Rules favors Ground by design.
SDV CTGAN was trained for 50 epochs (its default is 300) to keep the run within a few hours on one CPU; more epochs would likely improve it. SDV GaussianCopula used its defaults. Both ran on SDV 2 with metadata detected from the training split.
Differential privacy at ε = 3 keeps column shapes and the label mix, but a classifier can still tell DP rows from real ones (detection AUC 0.75–1.0): the noise breaks fine-grained structure, and rules can't be learned from the data without spending budget.
Single runs with fixed seeds; differences of a point or two in Overall are within noise.
MOSTLY AI ran locally from its open-source SDK (TabularARGN) with a 5-minute training cap per dataset; Synthcity ran with its plugin defaults. Gretel isn't included: it trains in Gretel's cloud, which would mean uploading the data to a third party.
Where MOSTLY AI scores low we checked the rows, not just the numbers: on bank, 35% of its euribor3m values are off the 308-value grid the real rate sits on, which a classifier spots; on insurance it lost the smoker effect on charges (smokers average $14.2k vs $31.4k real), after 47 seconds of training, well under its 5-minute cap.

In particular, CTGAN used 50 epochs rather than its documented default of 300, and MOSTLY AI used a five-minute cap. The default and differentially private Ground variants solve different privacy/utility tradeoffs. Do not use their scores interchangeably or infer compliance from an empirical attack metric.

What is missing from this snapshot

The recorded file contains no environment provenance block with a commit, machine and external package versions, and the current reproduction harness is in Ground's private repository. This snapshot is therefore not an independently reproducible public benchmark. A stronger next comparison should publish the runnable harness and provenance, repeated seeds, confidence intervals, a matched tuning budget and an independent evaluation.

Inspect the evidence

Evaluate Ground on your own task

Start with a representative dataset and a real holdout, define acceptance criteria and compare the workflow end to end. Create a Ground account or inspect the downloadable examples before uploading your own data.