dagzoo 0.14.0

Synthetic tabular data generator for causal modeling

dagzoo

High-throughput synthetic tabular data generation built around causal structure. Use it to generate, benchmark, and stress-test tabular datasets with deterministic seed behavior.

flowchart LR
    %% Class Definitions
    classDef setup fill:#e1f5fe,stroke:#01579b,stroke-width:2px,color:#01579b
    classDef core fill:#fff3e0,stroke:#e65100,stroke-width:2px,color:#e65100
    classDef gate fill:#f1f8e9,stroke:#33691e,stroke-width:2px,color:#33691e
    classDef out fill:#f3e5f5,stroke:#4a148c,stroke-width:2px,color:#4a148c

    Seed([Root Seed]) --> RNG[Deterministic Seeding]
    RNG --> Layout[Layout & DAG Sampling]
    Layout --> Mechanisms[Random Functional Mechanisms]
    Mechanisms --> Converters[Feature/Target Converters]
    Converters --> Filter[Learnability Filter]
    Filter --> Bundle[[DatasetBundle: X, y, Metadata]]

    %% Assign Classes
    class Seed,RNG setup
    class Layout,Mechanisms,Converters core
    class Filter gate
    class Bundle out

From Latent DAG to Tabular Data

Unlike many generators that treat each column as an independent noise source, dagzoo generates data from a latent causal structure. A single node in the causal graph can branch into multiple observable features, preserving complex dependency patterns.

flowchart LR
    %% Class Definitions
    classDef latent fill:#e1f5fe,stroke:#01579b,stroke-width:2px,color:#01579b,stroke-dasharray: 5 5
    classDef observable fill:#f5f5f5,stroke:#212121,stroke-width:2px,color:#212121

    subgraph LatentSpace [Latent Causal DAG]
        NodeA((Node A)) --> NodeB((Node B))
    end

    subgraph ObservableSpace [Tabular Dataset Layout]
        Feat1[Feature 1: Numeric]
        Feat2[Feature 2: Categorical]
        Feat3[Feature 3: Numeric]
        Target[Target Variable]
    end

    %% Mapping connections
    NodeA -. mapping .-> Feat1
    NodeA -. mapping .-> Feat2
    NodeB -. mapping .-> Feat3
    NodeB -. mapping .-> Target

    %% Assign Classes
    class NodeA,NodeB latent
    class Feat1,Feat2,Feat3,Target observable

    style LatentSpace fill:#f0faff,stroke:#01579b,stroke-dasharray: 5 5
    style ObservableSpace fill:#fafafa,stroke:#212121

Why dagzoo

Tabular foundation models rely on synthetic data priors whose quality directly determines downstream performance. The design of the synthetic prior — what graph structures are sampled, what functional relationships are used, what noise and missingness patterns are injected — determines the effective diversity of the training corpus: the breadth of meta-feature space the model sees during pretraining. Higher effective diversity means the model's in-context learning covers more of the real-world task space.

dagzoo provides independent control over graph structure, mechanism families, noise distributions, distribution shift, missingness, and learnability filtering. It is designed for researchers who need to systematically control and measure the axes of their synthetic prior.

dagzoo is for situations where you need synthetic tabular data that is:

• Causally structured: datasets are generated from a sampled latent DAG, not independent column noise.
• Reproducible: deterministic seed fan-out and effective-config trace artifacts make runs auditable.
• Stress-testable: shift, noise, missingness, and deferred filter controls let you probe model robustness under controlled distribution changes.
• Operationally scalable: canonical fixed-layout generation and benchmark guardrails support repeatable high-throughput workflows.

Quick Start

Examples in this README assume a repo checkout (so configs/*.yaml is available):

./scripts/dev bootstrap
source .venv/bin/activate
./scripts/dev doctor all

Install the packaged CLI globally when you do not need repo presets/config files:

uv tool install dagzoo

Generate a default batch from the repo:

dagzoo generate --config configs/default.yaml --num-datasets 10 --out data/run1

Each generate run writes effective_config.yaml and effective_config_trace.yaml in the resolved output directory. dagzoo generate samples one internal fixed-layout plan per run, so all datasets emitted in the same run share one sampled layout/execution plan. Run dagzoo filter as a separate stage for acceptance decisions. Deferred filtering now replays strictly from embedded shard metadata; generated artifacts must include metadata.config.dataset.task and metadata.config.filter. Generate configs must not include runtime.worker_count or runtime.worker_index.

Run deferred filtering on generated shards:

dagzoo filter --in data/run1 --out data/run1_filter

dagzoo filter now applies the small-shot ease filter as a separate replay stage, while benchmark smoke presets may override lineage-veto behavior when the goal is throughput measurement rather than structural rejection.

Run a downstream handoff workflow from generate:

dagzoo generate --config configs/default.yaml --num-datasets 10 --handoff-root handoffs/run1 --device cpu --hardware-policy none

dagzoo generate --handoff-root writes one stable handoff root with:

• handoff_manifest.json as the downstream machine-readable entrypoint
• generated/ for raw shard outputs plus effective-config artifacts

Run a smoke benchmark:

dagzoo benchmark --suite smoke --preset cpu --out-dir benchmarks/results/smoke_cpu

--device is a single-preset benchmark override. For multi-preset benchmark runs, set the device in each preset/config instead of passing one shared CLI override.

Inspect detected hardware tier:

dagzoo hardware

View help and available options for commands:

dagzoo --help
dagzoo generate --help
dagzoo filter --help
dagzoo benchmark --help

Local repo workflow before review:

./scripts/dev review-base
./scripts/dev ready

For focused local analysis outside the pre-review flow:

./scripts/dev impact
./scripts/dev verify quick

Documentation

Primary docs site:

• https://bensonlee5.github.io/dagzoo/docs/

Start here for end-user workflows and contracts:

• How It Works: System flow and terminology.
• Transforms (Math Reference): Formal transform math, notation, and operator definitions.
• Usage Guide: Primary workflow hub.
• Output Format: Output schema and artifacts.
• Feature Guides: Diagnostics, missingness, many-class, shift, noise, and benchmark guardrails.

If you are integrating dagzoo downstream, treat these as the stable references:

• Handoff workflow and CLI usage: Usage Guide
• Generated artifacts and handoff manifest schema: Output Format

License

dagzoo is released under the Apache License 2.0. See LICENSE for the full license text, NOTICE for the project notice, and THIRD_PARTY_NOTICES.md for repo-level third-party disclosure guidance.