MLJAR AutoML 1.4.0rc6

mljar-supervised 1.4.0rc6

One AutoML for classic machine learning, neural networks, and tabular foundation
models. This release candidate brings the changes tested in local beta builds
into the public 1.4 release series. Requires Python 3.11+.

python -m pip install "mljar-supervised==1.4.0rc6"

Choose your PyTorch build before adding [realmlp], [tabm], or [tfm] for
RealMLP, TabM mini, Kumo, and TabICLv2. [all] adds all optional model backends;
[apps] adds the runtime for exported prediction apps. See the
installation guide.

Validation and training

• Score cross-validation models and ensembles by the arithmetic mean of their
  fold scores across all repeats. Averaging repeated OOF predictions no longer
  substitutes for averaging the validation metrics used for model selection.
• Select a shared RealMLP epoch across CV folds and repeats by default. Evaluate
  checkpoints at epochs 10, 20, …, 250, 256 for the default training schedule,
  then restore each fold's own weights at the chosen epoch. Save the selection
  curve, verify restored predictions, and clean up temporary checkpoints.
• Use mean fold scores for shared XGBoost, LightGBM, and CatBoost round selection.
• For eligible small datasets, automatic validation uses five folds and three
  repeats, with stacking disabled. Dataset size, class counts, time budget, and
  explicit validation settings still determine eligibility.
• Reject incomplete candidates from final model selection and ensembling.
  Use completed-fold costs to decide whether to continue a candidate and reduce
  wasted training. Keep routine continuation messages out of the console.

Compete defaults and saved models

• Keep golden features enabled. Disable KMeans features, feature discovery, and
  boost-on-errors by default. Each remains available through its explicit option.
• Give generated features separate semantic identities, storage names, and
  display labels; preserve their fitted meaning when replaying saved models.
• Store KMeans artifacts separately for each CV fold and repeat, fixing changed
  predictions after reloading repeated-CV models.
• Use fold models to evaluate Kumo and TabICLv2, and one separately prepared
  full-training-data context for inference. Retain their predictions as ensemble
  and stacking inputs without fitting new TFM models on stacked features.
• Improve resource admission, expected budget-skip reporting, and PyTorch input
  handling. Keep failed candidates out of the ensemble.
• Include Python 3.11+ installation extras, exported-app dependencies, native
  compatibility checks, and documentation updates developed during the betas.

Known limitations

This is a release candidate. Validation remains part of adaptive model and
feature selection; it is not an unbiased estimate of final test performance.
The deferred validation-boundary work for legacy golden features, feature
selection, and optional boost-on-errors is not resolved by this release. Use a
separate held-out test set for final evaluation. No general accuracy improvement
is claimed from the partial local benchmark comparison.

Release checks

• Main tests passed on Python 3.11 and 3.12, along with the neural-model, SDM, native-compatibility, feature-discovery, exported-app, and documentation workflows.
• The release wheel and source package passed clean installation, dependency, training, ensemble-prediction, and reload checks.
• RealMLP, TabM, Kumo Large, and TabICLv2 passed GPU training and per-model reload checks.
• Shared-checkpoint, budget-skip, repeated-CV, and package tests passed against the final release artifacts.

Release CI. Download checksums are included in SHA256SUMS.