TabDGP

A data-generating process (DGP) for grouped and flat tables: grouping structure and typed covariates with the quirks of real data.

What and why

• Measured against real tables. The default hyperprior is matched to mixedbench's registry of real grouped tables, then widened; two tests prove it: generated tables are indistinguishable from real ones on their summary statistics, and every real table lies inside the generated support.
• White-box. Every dataset comes with its generating graph, its latent nodes and its column metadata (type, levels, role).
• Composable. One structural causal model whose nodes are grouping factors, group-level and within-group covariates, time and treatment; a flat table is the case without grouping factors.
• The quirks i.i.d. simulators lack. Group confounding, empty cells, rare levels, ties, bounds, irregular time, near dependence, leverage points: a catalogue, each quirk measured on real tables before it gets an axis.
• Covariates only. The outcome belongs to whoever fits a model; an optional outcome node builds white-box misspecified datasets.
• Overridable. Every quantity has a default distribution; pass a point value or another distribution for any of them.

Install

pip install tabdgp

From source:

git clone https://github.com/adkipnis/tabdgp
cd tabdgp
uv sync

Quickstart

The first call arrives with roadmap item 1; see ROADMAP.md.

Contents

• The package tabdgp under src/.
• CONTEXT.md: the vocabulary; docs/adr/: the design decisions.

Contributing

Interested in becoming part of TabDGP? Check out CONTRIBUTING.md.

Licence

MIT.