OpenCausaLab/causcale-data

CauScale Dataset

Dataset for CauScale: Neural Causal Discovery at Scale (ICML 2026).

Overview

This repository contains test data for CauScale, a neural architecture for causal graph discovery that scales to graphs with up to 1000 nodes.

Files

File	Description
data_test.zip	Test data for synthetic and SERGIO benchmarks

Data Format

Synthetic Data

Generated using linear, neural network, and polynomial structural equation models on Erdős–Rényi (ER) and Scale-Free (SF) graphs.

Each instance folder contains:

• DAG{i}.npy — ground-truth adjacency matrix (d, d), boolean, entry (i, j) = True means i → j
• data_interv{i}.npy — concatenated observational and interventional data (N, d)
• intervention{i}.csv — intervened node index per row (empty = observational)
• regime{i}.csv — environment index per row (0 = observational, 1..K = interventional)

Test: 10–1000 nodes, 5 DAGs each, 1000 data points per DAG

SERGIO Gene Expression Data

Simulated single-cell gene expression data using the SERGIO simulator.

Each instance folder contains:

• DAG.npy — ground-truth adjacency matrix (d, d), boolean
• data_intv.npy — interventional data (N_int, d)
• intv.npy — intervention mask (N_int, d), boolean
• data.npy / clean.npy / dropout.npy — observational data variants (150, d)
• info.json — metadata

Test: Nodes 100–200, ER graphs, expected degrees 2 and 4, 5 graphs each, 20000 interventional points

Usage

from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="OpenCausaLab/causcale-data",
    filename="data_test.zip",
    repo_type="dataset",
)

See the CauScale code repository for full training and inference instructions.

Citation

@article{peng2026causcale,
  title={CauScale: Neural Causal Discovery at Scale},
  author={Peng, Bo and Chen, Sirui and Tian, Jiaguo and Qiao, Yu and Lu, Chaochao},
  journal={arXiv preprint arXiv:2602.08629},
  year={2026}
}