Synthetic data

Training data without the patient records.

Phantomset generates stand-in datasets that behave like the real thing: same distributions, same correlations, same ugly edge cases. What they don't contain is a single personal record. Your models learn the shape of the data without ever touching the people in it.

Request early access How it works

Pre-launch · for teams blocked on data access, not on models

The bottleneck moved from modelling to permission.

01 / The gap

Legal review is the critical path

Access to production data now takes longer than the modelling work it enables. Roadmaps stall on signatures, not on GPUs.

The hard cases are the rare ones

Rare conditions, unusual combinations, extreme values — exactly the rows you most need are the ones that fewest people can share.

Vendor data carries provenance risk

Third-party datasets arrive without a defensible lineage. When a model is challenged, nobody can trace what it was trained on.

Learn the structure, generate the rows.

02 / The method

Profile a source you're allowed to see

Schemas, type constraints, distributions and inter-column relationships. Where real data isn't shareable at all, we fit to a documented specification instead.

Generate a phantom set

Relational integrity is preserved across tables, so joins behave the way your pipelines expect. Categorical mixes and tail behaviour are matched, not smoothed away.

Prove it can't be traced back

Re-identification and nearest-neighbour tests against the source, with the results attached to the dataset as evidence rather than a claim in an email.

A dataset you can hand to a contractor, a regulator and a model — without a meeting in between.

Built for the questions that get asked after the breach.

03 / Capabilities

Distribution fidelity

Marginal and joint statistics compared against the source, reported per column so you can see where the fit is weak.

Relational integrity

Multi-table generation that keeps keys, cardinalities and referential structure intact across the whole schema.

Privacy testing

Membership-inference and nearest-neighbour checks run against every generated set, with pass thresholds you set.

DP budgets where needed

Optional differential privacy with an explicit epsilon, tracked per release rather than per pipeline run.

Provenance record

Every dataset ships with its generator version, fitting parameters, privacy results and the decisions made along the way.

No re-identification surface

Direct identifiers are never carried through, and quasi-identifier combinations are tested rather than assumed safe.

Early access

We're working with a small number of teams whose data access is the constraint. Send us the schema shape and the regulatory frame you operate in, and we'll tell you whether a phantom set is viable for it.

Or email hello@phantomset.com directly.