Causilo 1.0.3

Changes

• Add Python 3.13 and 3.14 support. Causilo now supports Python 3.10–3.14.
• Expand CI coverage to all five supported Python versions and update the installation requirements in the README.
• No inference logic or dependency constraints changed.

Addresses #12 via #13.

Validation

The offline test suite passed on Python 3.10–3.14 with PyTorch 2.13.0 CPU: 140 passed, 7 skipped, and 66 pretrained tests deselected per version. Package build and distribution checks passed, and the publication workflow verified the installed wheel before uploading it to PyPI. GPU and pretrained-model tests were not run.

Installation

pip install --upgrade causilo==1.0.3

PyPI package · Full changelog

Causilo 1.0.2

Causilo 1.0.2 packages the latest classification and memory-handling improvements.

• Support classification with more than 10 classes through deterministic error-correcting output codes (ECOC).
• Reuse query preprocessing across ECOC code rows and deduplicate shared training data in serialized fitted state.
• Handle large-input out-of-memory conditions with tiled column-summary recomputation.
• Preserve power-normalization overflows as NaN for downstream imputation.

Saved estimators require matching Causilo and dependency versions. Refit with 1.0.2 when upgrading from an older saved estimator.

Install or upgrade with pip install --upgrade causilo==1.0.2.

Full changelog: v1.0.1...v1.0.2

Causilo 1.0.1

Causilo 1.0.1 adds median and quantile regression predictions through predict.

• Use output_type="median" for the ensemble's 0.5 quantile.
• Use output_type="quantiles", quantiles=[0.05, 0.5, 0.95] to request specific quantiles, with linear interpolation and exponential tail extrapolation.
• Use output_type="raw" to retrieve all 999 native ensemble quantiles in target units.
• The default remains mean prediction. Existing checkpoints are supported.

The inference guide includes output shapes and examples. Validation covers cached execution, serialization, quantile interpolation and tails, and median/0.5-quantile agreement.