CausalPFN for ATE/CATE: inference and evaluation
An independent inference wrapper and fixed-test-set evaluation using the official pretrained CausalPFN checkpoint unchanged. We did not train, fine-tune, or calibrate its neural weights. This is not an official authors' release. Our contributions are inference packaging, numerical adaptation, benchmark construction, evaluation and reproducible results.
ATE = E[Y(1) - Y(0)]; CATE(x) = E[Y(1) - Y(0) | X=x].
Official model · Original CausalPFN paper · Official code · Evaluation · Report PDF
The linked paper describes the original model, not this independent release. There is no newly trained Adapter, error head or calibration checkpoint.
Release validation status
The downloaded inference and reproduction archives passed SHA-256 and extracted-file verification. All 8,000 reference results and metric summaries were independently recomputed; the downloaded CPU FP32 example also passed. See current release validation.
The original formal experiment ran on CUDA bf16. A separate GPU smoke test of the standalone packaging has not been performed; CPU validation does not replace it. The verification JSON embedded in the archives records their earlier packaging-time status; RELEASE_STATUS.json records the subsequent download acceptance. Scientific code, weights, data, results and downloadable archives are unchanged.
Checkpoints and downloads
| Artifact | Contents |
|---|---|
| Checkpoint | Unchanged official weights |
| Source and SHA-256 | Immutable upstream revision and attribution |
| Inference ZIP | Standalone runtime, checkpoint, licenses and independent observed-only example |
| Reproduction archive | DGPs, 2,000 tables, oracle, both baselines and 8,000 reference results |
| Archive checksums | SHA-256 checksums |
Quick Start
Download only the inference files for this Quick Start. The full experiment is
available separately as one reproduction archive. Use a recorded commit ID with
--revision for an immutable release.
hf download fuyao11/CausalPFN-ATE-CATE --include "arena_n500/*" --local-dir causal_ate_cate
cd causal_ate_cate
python -m venv .venv
source .venv/bin/activate
python -m pip install -r arena_n500/requirements.txt
python arena_n500/predict.py --verify-only
python arena_n500/quickstart.py
This Quick Start is an explicitly labeled CPU FP32 compatibility demo. It does not reproduce GPU bf16 timing or claim numerical identity with that precision. The saved example output comes from real execution and records its environment.
For CUDA bf16 inference on a supported GPU:
python arena_n500/predict.py --observed arena_n500/example/observed_n500.npz --device cuda:0 --neural-precision bf16
Initialize Predictor once and reuse it. Inputs are observed X, T, Y and a single
CATE query; the interface has no oracle access. See input/output contract.
Evaluation
One frozen collection: 13 DGP families, 1,000 parameterized mechanisms with two independent datasets each, n=500 and dimension 2-10. Every dataset is evaluated for ATE and CATE by both methods. No training-distribution membership is asserted.
| Target | Method | Bias | Variance | RMSE | Coverage | Mean length | Time (s) |
|---|---|---|---|---|---|---|---|
| ATE | CausalPFN | 0.0119 | 0.0097 | 0.1114 | 77.63% | 0.2599 | 1.089 |
| ATE | RF | -0.0167 | 0.0137 | 0.1284 | 96.75% | 0.5897 | 90.490 |
| CATE | CausalPFN | 0.0143 | 0.0480 | 0.2966 | 100.00% | 5.8471 | 0.219 |
| CATE | RF | -0.0139 | 0.1560 | 0.5522 | 84.84% | 1.2689 | 55.501 |
Rows use equal family weights (1/13), not dataset-count weights. Variance is within-mechanism sampling variance, not SE. All 8,000 estimates and intervals completed successfully. RF uses sklearn 1.7.2, 200 trees, min_samples_leaf=5, five-fold AIPW for ATE / T-learner ensemble for CATE, and B=50 whole-row bootstrap. CausalPFN uses native intervals with 10,000 draws, no external bootstrap or added cross-fitting. CUDA neural forwards use bf16 and statistics use float64.
ATE CausalPFN coverage remains below 95%. CATE 100% coverage accompanies wide intervals and does not imply precise nominal calibration. Time includes interval work, including RF's 50 refits. Four L40 GPUs and 216 single-thread RF workers were used; these are not paired same-hardware speedup measurements.
Reproduction and checks
The complete experiment code, 2,000 input tables and 8,000 individual results are inside the reproduction archive. They are not duplicated as thousands of loose repository files. The scientific code, data, results and both archive checksums are unchanged by this layout update.
Download and verify the archive:
hf download fuyao11/CausalPFN-ATE-CATE --include "releases/*" --local-dir causal_ate_cate
cd causal_ate_cate/releases
# macOS (use sha256sum -c SHA256SUMS on Linux):
shasum -a 256 -c SHA256SUMS
tar -xzf ATE-CATE-reproduction.tar.gz
cd ATE-CATE-reproduction
python checks/verify_release.py
# With NumPy installed:
python checks/audit_reference.py
Inside the extracted archive:
| Path | Contents |
|---|---|
reproduction/README.txt, reproduction/SERVER_RUN.txt |
Setup and execution instructions |
reproduction/src/, reproduction/configs/ |
Baseline implementations and experiment settings |
reproduction/data/, reproduction/oracle/ |
Frozen test data and truth |
reproduction/reference_results/ |
Original results, preserved for inspection |
checks/ |
Independent result audit and model comparison checks |
New experiment runs write to reproduction/results/. No full rerun is necessary
to inspect the report or recompute reference summaries. The separate standalone
CUDA smoke test is still pending; see release validation.
For a full download of this compact repository, python checks/verify_release.py
checks its root release manifest. The extracted archive has its own independent
manifest and verification entry point.
Attribution and limitations
See NOTICE, checkpoint license, and model card. The license metadata identifies included checkpoint terms; it does not relicense third-party components. Original authors retain credit. No affiliation or endorsement is implied. See EVALUATION.md for estimands, weighting, interval interpretation and runtime limitations.