OpenCausaLab/causcale-model

CauScale

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

Checkpoints

File	Trained on	AUPRC
synthetic/auprc=0.905_migrated.ckpt	Synthetic data (10–500 nodes)	0.905
sergio/auprc=0.703_migrated.ckpt	SERGIO gene expression data (10–200 nodes)	0.703

Usage

Download and place under checkpoints/:

from huggingface_hub import hf_hub_download

hf_hub_download(
    repo_id="OpenCausaLab/causcale-model",
    filename="synthetic/auprc=0.905_migrated.ckpt",
    repo_type="model",
    local_dir="checkpoints",
)
hf_hub_download(
    repo_id="OpenCausaLab/causcale-model",
    filename="sergio/auprc=0.703_migrated.ckpt",
    repo_type="model",
    local_dir="checkpoints",
)

Then run inference:

bash bash/inference-synthetic.sh   # synthetic data
bash bash/inference-sergio.sh      # SERGIO gene expression data

See the CauScale code repository for full 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}
}