frontal-labs/causal-1-training-data

Causal-1 Training Data

Synthetic structural causal models (SCMs) with exact ground-truth interventional and counterfactual answers, used to train frontal-labs/causal-1.

Every row is one complete causal task: a DAG, an observational sample from it, a treatment/outcome query, and the true average and individual treatment effects. Because the worlds are simulated, the counterfactuals are not estimated — they are computed by re-running the same structural equations with the same exogenous noise and one variable pinned.

Splits

split	rows	variables	purpose
train	2,000	3–20	curriculum mixture
validation	256	3–20	same distribution, disjoint worlds
unseen_structures	128	21–27	larger than anything in training

Splits share no worlds: their seed ranges are disjoint, enforced by the generator rather than by filtering after the fact.

Fields

field	meaning
variable_names, variable_types, n_categories, observed	per-node schema; observed=false marks a latent confounder
edge_source, edge_target, edge_strength, edge_type	the DAG, as indices into variable_names
treatment, outcome, x0, x1	the query: the effect of moving treatment from x0 to x1
ate, ite	ground truth, computed from the simulator
identified, adjustment	whether a valid backdoor set exists in the observed graph, and a minimal one
observational, observed_mask	[observational_rows, n_variables], flattened row-major
factual, counterfactual	[counterfactual_rows, n_variables], flattened; same noise under x0 and x1
outcome_scale, n_latent	the outcome's observational spread, and the number of hidden variables

Reshape a flattened field with its companion row count:

import numpy as np
obs = np.array(row["observational"]).reshape(
    row["observational_rows"], row["n_variables"]
)

Regenerating it

This is a fixed sample of an unbounded stream. manifest.json records each split's spec and seed offset, so the same generator version reproduces these rows exactly and can produce far more:

from causal.curriculum import matched_validation_split
from causal.datasets import build_tasks

tasks = build_tasks(matched_validation_split(20), count=100_000, start=0)

Cross-process reproducibility is covered by the project's test suite.

What is in the distribution

Linear and nonlinear mechanisms; continuous, binary and categorical variables; latent confounders; latent mediators (the observed graph is silent about a real effect); selection variables that are colliders on the treatment–outcome pair; heteroscedastic noise; heavy-tailed (Student-t) noise; monotone marginal warps; missing values; and eight topology families — chain, fork, collider, confounded, mediator, latent-mediator, hub, and random DAGs.

Known biases — read before training on this

Measured properties of the generator, and the leading explanation for why a model trained here transfers poorly to real observational studies:

• Wide graphs have weak treatments. Tasks whose outcome has eight or more parents average a normalised |ATE| of 0.639, against 1.543 overall. A model trained on this learns that a long covariate list implies a modest treatment effect. Real studies routinely violate that — IHDP's treatment moves its outcome by 1.8 standard deviations across 26 covariates.
• Every node is calibrated to unit variance. This deliberately removes marginal-variance ordering as a spurious orientation cue, but it also means a single parent cannot dominate a long parent list the way it can in real data.
• Fixed sample size. Every task carries 256 observational rows, so a model trained here is never taught that more data should sharpen its estimate.
• One homogeneous SCM per task. No cell-population mixtures, no feedback loops. Real biological data such as protein signalling has both, and a DAG cannot represent a cycle at all.

Citation

Generated by the causal project. Licensed Apache-2.0.