unstale

pypi / flaml

flaml

What changed in flaml 2, from its official release notes, upgrade guides and advisories. Sources checked 2026-10-19.

5 new APIs and options21 fixes and other

See which of these apply to your code, ranked, in your agent:

npx @unstale/cli init

@unstale/cli on npm

flaml 226 changes

New APIs and options 5

  • Adds an optional resampler keyword argument for per-fold class-imbalance handling.

    2.7.0release notes

  • Generates timestamps for time series predictions when test data is not provided.

    2.6.0release notes

  • Adds configurable label overlap handling for the classification holdout strategy.

    2.5.0release notes

  • The LGBMEstimator search space adds the objective parameter.

    2.5.0release notes

  • Task-level and estimator-level preprocessors are exposed as a public API.

    2.5.0release notes

Fixes and other changes 21

  • Preserves anonymous tuning metrics whose value is zero.

    2.7.0release notes

  • Categorical codes are fixed at fit time, and unseen values trigger a warning.

    2.7.0release notes

  • Float sampling now honors unit quantization.

    2.7.0release notes

  • Handles mixed-type conditional parameters.

    2.7.0release notes

  • Forwards free_mem_ratio from compute_estimator to the evaluation.

    2.7.0release notes

  • Fixes crashes and misaligned sample weights when preparing data with non-shuffled or group holdout splits.

    2.7.0release notes

  • SGD, logistic regression, random forest, and XGBoost estimators now honor random_seed for reproducible results.

    2.7.0release notes

  • Supports recent scikit-learn changes, including LRL1/LRL2 deprecations and fitted-state detection when reloading MLflow autologged models.

    2.7.0release notes

  • Naive forecasters support statsmodels 0.15, and future forecast horizons are corrected.

    2.7.0release notes

  • Trial averages are computed from per-metric report counts.

    2.7.0release notes

  • Fixes compatibility with pandas 3.0, including StringDtype, datetime resolution, deprecated APIs, and pyspark.pandas imports.

    2.6.0release notes

  • Adds a build-system section and migrates packaging metadata to PEP 621 to unblock editable installs.

    2.6.0release notes

  • Fixes BaseEstimator type detection for ensembles to support scikit-learn 1.7+.

    2.6.0release notes

  • Pins setuptools below version 82 to prevent a pkg_resources not found error.

    2.6.0release notes

  • Selects parameters according to the Transformers version to support v4.26+ and v5.0+.

    2.6.0release notes

  • Preserves FLAML_sample_size in best_config_per_estimator.

    2.5.0release notes

  • Fixes an OptunaSearch warning for non-hierarchical spaces with Ray Tune domains.

    2.5.0release notes

  • Adds validation and clear error messages for the custom_metric parameter.

    2.5.0release notes

  • Fixes eval_set preprocessing for XGBoost estimators when categorical features are present.

    2.5.0release notes

  • FLAML adds support for Python 3.13.

    2.5.0release notes

  • Fixes nested dictionary merging that could lose sampled hyperparameters.

    2.5.0release notes