ci_transformer: Causal data generation + CI classifiers

Unified Python toolkit for:

• Generating diverse continuous SCMs and datasets
• Training and benchmarking conditional-independence (CI) classifiers
• Streaming, curriculum-based CI training workflows

---

Install

Recommended (avoids PEP 668 issues):

python3 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -e .

Requires Python>=3.9 plus numpy, torch (installed via setup).

---

Quickstart

Generate an SCM and data

from ci_transformer.causal_data_gen import CausalSCM, RandomDAGConfig

cfg = RandomDAGConfig(d=12, edge_prob=0.22, max_parents=4, ensure_connected=True)
scm = CausalSCM.random(cfg, seed=0)

X = scm.sample(n=5000, seed=1)
A = scm.adjacency()
print(scm.is_ci_true(i=3, j=7, S=[0, 2]))

Save a dataset

from ci_transformer.causal_data_gen import generate_dataset, RandomDAGConfig

ds, scm = generate_dataset(n=20000, cfg=RandomDAGConfig(d=10, edge_prob=0.25, max_parents=3), seed=123)
ds.save_npz("toy_scm_dataset.npz")

Train a CI classifier (script)

python scripts/train_ci.py --out runs/ci_demo --steps 50000

Checkpoints/logs land in runs/ci_demo/.

Use the streaming dataloader directly

from ci_transformer.training import CIStreamingDataset, make_dataloader, Curriculum

ds = CIStreamingDataset(n_rows=500, m_max=5, curriculum=Curriculum())
dl = make_dataloader(ds, batch_size=64)
batch = next(iter(dl))  # contains x, y, z, z_mask, label

---

Package layout

• ci_transformer.causal_data_gen: SCMs, DAG sampling, mechanisms, dataset I/O
• ci_transformer.ci_models: PyTorch CI classifier + configs
• ci_transformer.training: Streaming CI dataset + training loop utilities
• scripts/train_ci.py: CLI entry for CI training
• tests/: unit tests (e.g., CI model invariances)

---

Testing

pytest
# or narrow:
pytest tests/ci_transformer/ci_models/test_invariants.py

---

Notes

• Observational data only (no interventions/environments yet).
• Mechanism/noise families are pluggable (see ci_transformer/causal_data_gen/mechanisms.py).
• Imports should all be via ci_transformer.* after editable install.