updates:
- [github.com/astral-sh/ruff-pre-commit: v0.1.13 → v0.1.14](astral-sh/ruff-pre-commit@v0.1.13...v0.1.14)

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* read file in read mode

* cast parameters to expected types

* Following PGijsbers proposal to ensure that avoid_duplicate_runs is a boolean after reading it from config_file

* Add a test, move parsing of avoid_duplicate_runs

---------

Co-authored-by: PGijsbers <[email protected]>

* Update 'sparse' parameter for OHE for sklearn >= 1.4

* Add compatability or skips for sklearn >= 1.4

* Change 'auto' to 'sqrt' for sklearn>1.3 as 'auto' is deprecated

* Skip flaky test

It is unclear how a condition where the test is supposed to pass
is created. Even after running the test suite 2-3 times, it does
not yet seem to pass.

* Fix typo

* Ignore description comparison for newer scikit-learn

There are some minor changes to the docstrings. I do not know
that it is useful to keep testing it this way, so for now I will
disable the test on newer versions.

* Adjust for scikit-learn 1.3

The loss has been renamed.
The performance of the model also seems to have changed slightly
for the same seed. So I decided to compare with the lower fidelity
that was already used on Windows systems.

* Remove timeout and reruns to better investigate CI failures

* Fix typo in parametername

* Add jobs for more recent scikit-learns

* Expand the matrix with all scikit-learn 1.x versions

* Fix for numpy2.0 compatibility (#1341)

Numpy2.0 cleaned up their namespace.

* Rewrite matrix and update numpy compatibility

* Move comment in-line

* Stringify name of new step to see if that prevented the action

* Fix unspecified os for included jobs

* Fix typo in version pinning for numpy

* Fix version specification for sklearn skips

* Output final list of installed packages for debugging purposes

* Cap scipy version for older versions of scikit-learn

There is a breaking change to the way 'mode' works, that
breaks scikit-learn internals.

* Update parameter base_estimator to estimator for sklearn>=1.4

* Account for changes to sklearn interface in 1.4 and 1.5

* Non-strict reinstantiation requires different scikit-learn version

* Parameters were already changed in 1.4

* Fix race condition (I think)

It seems to me that run.evaluations is set only when the run is
fetched. Whether it has evaluations depends on server state.
So if the server has resolved the traces between the initial
fetch and the trace-check, you could be checking
len(run.evaluations) where evaluations is None.

* Use latest patch version of each minor release

* Convert numpy types back to builtin types

Scikit-learn or numpy changed the typing of the parameters
(seen in a masked array, not sure if also outside of that).
Convert these values back to Python builtins.

* Specify versions with * instead to allow for specific patches

* Flow_exists does not return None but False is the flow does not exist

* Update new version definitions also installation step

* Fix bug introduced in refactoring for np.generic support

We don't want to serialize as the value np.nan, we want to
include the nan directly. It is an indication that the
parameter was left unset.

* Add back the single-test timeout of 600s

* [skip ci] Add note to changelog

* Check that evaluations are present with None-check instead

The default behavior if no evaluation is present is for it
to be None. So it makes sense to check for that instead.
As far as I can tell, run.evaluations should always contain
some items if it is not None. But I added an assert just in
case.

* Remove timeouts again

I suspect they "crash" workers. This of course introduces the risk of hanging processes... But I cannot reproduce the issue locally.

* Add HTTP headers to all requests

This allows us to better understand the traffic we see to our API.
It is not identifiable to a person.

* Update unit test to pass even with user-agent in header

* Add packaging dependency

* Change use of distutils to packaging

* Update missed usage of distutils to packaging

* Inline comparison to clear up confusion

* Prefer parquet over arff, do not load arff if not needed

* Only download arff if needed

* Test arff file is not set when downloading parquet from prod

* Add progress bar to downloading minio files

* Do not redownload cached files

There is now a way to force a cache clear, so always redownloading
is not useful anymore.

* Set typed values on dictionary to avoid TypeError from Config

* Add regression test for parsing booleans

* Towards lazy-by-default for dataset loading

* Isolate lazy behavior to pytest function outside of class

* Solve concurrency issue where test would use same cache

* Ensure metadata is downloaded to verify dataset is processed

* Clean up to reflect new defaults and tests

* Fix oversight from 1335

* Download data as was 0.14 behavior

* Restore test

* Formatting

* Test obsolete, replaced by test_get_dataset_lazy_behavior

Sometime between 3.9 and 3.12 the stringification of
ordered dicts changed from using a list of tuples to
a dictionary.

* Remove archive after it is extracted to save disk space

* Leave a marker after removing archive to avoid redownload

* Automatic refresh if expected marker is absent

* Be consistent about syntax use for path construction

* Pass kwargs through task to ```get_dataset```

Allows to follow the directions in the warning ```Starting from Version 0.15 `download_data`, `download_qualities`, and `download_features_meta_data` will all be ``False`` instead of ``True`` by default to enable lazy loading.```

* docs: explain that ```task.get_dataset``` passes kwargs

* Update openml/tasks/task.py

Remove Py3.8+ feature for backwards compatibility

---------

Co-authored-by: Pieter Gijsbers <[email protected]>

* Change defaults for `get_task`

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Fix linting errors

* Add missing type annotation

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* Expand 0.15.0 changelog with other PRs not yet added

* Bump version number

* Add newer Python versions since we are compatible

* Revert "Add newer Python versions since we are compatible"

This reverts commit 5088c80.

* Add newer compatible versions of Python