Base ModelCausal (discovery & effects)
CDFM · Causal Discovery Foundation Model
Which variable causes which, from a table · one forward pass
Qiao, Cai, Li et al. 2026 · CDFMApache-2.0
Use for a fast first causal map of a cross-sectional table — rows are independent samples, not a time series. Inference-only. For time-series causal discovery with lags use Ceptron XS-discovery or TCPFN; for a trainable causal model use Do-PFN or TCPFN.
Params
~9.66M
Priors
15 families
synthetic SCM mechanisms
Output
Graph
D×D, one forward pass
Run it
CLI
git clone https://github.com/profitopsai/pfnstudio && cd pfnstudio
pip install -e packages/core -e packages/cli
pfnstudio run cdfm-basePulls the pretrained weights from Hugging Face and runs the model's own held-out sample, printing the tasks it supports, a lock digest pinning exactly what was pulled, and a live result. Pass --input your.json to run your data.
Or in the studio: install this base model → then Predict.