Template configurations

The dagsampler.templates module ships helper functions that build CausalDataGenerator configs for the structural patterns that recur in CI-test benchmarks — chains, forks, colliders, and no-edge (independent) variable sets. Each helper handles node-spec normalisation, seeding, and reasonable defaults so a benchmark fixture can be assembled in a few lines.

All templates are re-exported from the package root:

from dagsampler import (
    CausalDataGenerator,
    chain_config,
    fork_config,
    collider_config,
    indep_config,
    independence_config,
)

Common arguments

Most templates accept a mechanism argument selecting the structural form applied to endogenous continuous and binary nodes:

• "linear" — weighted sum of parents with additive Gaussian noise.

• "sigmoid" — tanh of a weighted parent sum, plus additive Gaussian noise.

• "stratum_means" — required when any parent is categorical; the simulator falls back to "linear" automatically when no categorical parent is present.

Categorical endogenous nodes always use categorical_model = {"name": "logistic"} regardless of the mechanism argument.

The seed argument can be:

• an int — sets both seed_structure and seed_data to the same value;

• a dict like {"structure": 1, "data": 2} — sets each stream independently;

• None — leaves both streams unseeded.

The optional post_transform argument, when set, applies the named post-nonlinear transform (any key from the post-transform registry — see Explanation: SCM formulation) to every endogenous continuous node.

independence_config / indep_config

A no-edge config: every node is exogenous. Useful for null scenarios in CI benchmarks.

cfg = independence_config(
    var_specs=[
        {"name": "X", "type": "continuous"},
        {"name": "B", "type": "binary"},
        {"name": "C", "type": "categorical", "cardinality": 4},
    ],
    n_samples=300,
    seed=7,
    force_uniform=True,
)

indep_config is a backwards-compatible shorthand: if var_specs is omitted, it generates n_vars variables of a single node_type named {prefix}0..{prefix}{n_vars-1}.

cfg = indep_config(n_vars=5, node_type="binary", n_samples=300, seed=7)

When force_uniform=True (default), the config sets simulation_params.force_uniform_marginals = True so binary nodes get an exact 50/50 split and categorical nodes get equal class counts.

chain_config

A chain var_specs[0] → var_specs[1] → ... → var_specs[-1]. The first node is exogenous; each subsequent node is endogenous with the previous as its single parent.

cfg = chain_config(
    var_specs=[
        {"name": "X", "type": "continuous"},
        {"name": "M", "type": "continuous"},
        {"name": "Y", "type": "continuous"},
    ],
    mechanism="linear",
    n_samples=400,
    seed={"structure": 11, "data": 12},
    post_transform="tanh",
)

fork_config

A fork root → left, root → right (a common-cause / confounder pattern). var_specs is a dict with keys "root", "left", "right".

cfg = fork_config(
    var_specs={
        "root":  {"name": "Z", "type": "continuous"},
        "left":  {"name": "X", "type": "continuous"},
        "right": {"name": "Y", "type": "continuous"},
    },
    mechanism="linear",
    n_samples=300,
    seed=42,
)

collider_config

A v-structure left → collider, right → collider. left and right are exogenous; collider has both as parents.

cfg = collider_config(
    var_specs={
        "left":     {"name": "X", "type": "continuous"},
        "right":    {"name": "Y", "type": "continuous"},
        "collider": {"name": "Z", "type": "continuous"},
    },
    mechanism="linear",
    n_samples=300,
    seed=42,
)

Putting it together

Each template returns a plain dict ready to hand to CausalDataGenerator:

from dagsampler import CausalDataGenerator, chain_config

cfg = chain_config(
    var_specs=[
        {"name": "X", "type": "continuous"},
        {"name": "Y", "type": "continuous"},
    ],
    mechanism="linear",
    n_samples=200,
    seed=0,
)
result = CausalDataGenerator(cfg).simulate()