I released synloc 1.1.0, and the headline feature is model export: fit a synthetic-data model once, export only its parameters, and generate synthetic data in a completely separate environment — without the original records.

This is the workflow I kept wanting for data custodians: fit inside a secure environment, hand over a small parameter file, and let researchers draw synthetic samples elsewhere. The existing fit() workflow is unchanged; export is an optional addition.

pip install --upgrade synloc

Two ways to export

Cluster models — clusterCov can now fit a reusable Gaussian mixture without generating any synthetic rows:

from synloc import clusterCov, ClusterCovSampler

model = clusterCov(data, n_clusters=20, size_min=20).fit_model()
diagnostics = model.disclosure_diagnostics()  # inspect before export
model.export_model("cluster-model.npz")

# In a separate environment, with just the .npz file:
sampler = ClusterCovSampler.load("cluster-model.npz")
synthetic = sampler.sample(1000, random_state=123)

kNN models — LocalCov.prepare_export() selects anchors using the full dataset, estimates their local distributions, and returns a fixed proposal you can inspect before exporting:

from synloc import LocalCov

model = LocalCov(data, K=100, normalize=True)
proposal = model.prepare_export(n_components=50, random_state=42)
diagnostics = proposal.disclosure_diagnostics()
proposal.export_model("local-model.npz")

n_components=None uses every row as an anchor; a smaller count selects distinct anchors uniformly. K is the neighbourhood size, independent of how many components you export, and random_state seeds anchor selection, not later sampling.

What the file contains

The version-1 NPZ stores only version, column_names, weights, means, and covariances. No source rows, no cluster memberships, no contributor counts, no original extrema, no pickled estimator. Means and covariances are in original units — normalization only affects neighbour or cluster selection, so nothing needs rescaling when loading. Loading validates the format with pickle disabled.

One honest caveat: exported sampling does not clip to source bounds, so generated values can fall outside the original range.

Inspecting a model before release

disclosure_diagnostics() returns one row per component, in the same order as the fitted means:

  • count: source rows in the component.
  • rank: rank of the component’s centred rows — a low rank means the component is close to a few records.
  • constant_columns: variables with one shared value in the component.
  • top1_share_<column> / top2_share_<column>: the largest one and two values as shares of the component total (NaN for negative values or a zero total).

These diagnostics are computed from the source rows, so keep them confidential; they are not written to the exported file. They apply no thresholds and do not establish privacy — apply your own disclosure rules before releasing a model.

The privacy fine print

I want to be clear about what this is and is not. The export is a tool for the disclosure-control workflow, not a substitute for it. Loading validates the numerical format, not privacy: means, covariances, weights, overlapping neighbourhoods, and repeated exports can still disclose information. Counts refer to rows, not verified unique people.

Other 1.1 changes

  • LocalCov accepts an optional random_state, so kNN fitting is reproducible across worker counts; omitting it keeps the legacy behaviour.
  • Fitted mixture parameters are read-only, so sampling, export, and diagnostics can’t silently desynchronize.
  • Tighter input validation (complex values rejected, NumPy integer counts accepted) and covariance checks that can no longer be masked by large variances elsewhere.

Have a nice day!