As discussed in #12464 (comment) and #15358 (comment) in some specific use-cases negative sample_weight may be meaningful, however in most cases they should never happen.

So we may want to

• add force_positive=None parameter to _check_sample_weights.
• add a assume_positive_sample_weights=True config parameter to sklearn.set_config / get_config.

By default, force_positive=None would error on negative sample weight, but this check could be disabled globally with sklearn.set_config(assume_positive_sample_weights=False).

With _check_sample_weights(.., force_positive=True) the check would always be done irrespective of the config parameter.

If there are no objections to this, please tag this issue as "Help Wanted".