Causal Structure Uncertainty Benchmark – Synthetic Linear SEM Suite

It consists of independent and identically distributed samples under a variety of conditions, including different numbers of observed variables, graph densities, varying direct effect strengths, measurement noise levels, and finite sample sizes.

All samples are produced from explicitly defined causal skeletons and complete orientation information, thereby providing known ground-truth causal structures while retaining the structural inference challenges that arise from finite-sample variability. The dataset can serve as a public benchmark for testing resampling-based diagnostic tools, evaluating methods that quantify structural uncertainty, and comparing the stability performance of different causal discovery algorithms.