Thanks for the report. But from your description it's not clear to me what you're intending to change. It seems to me writing a custom kernel is exactly designed for these cases.

So closing as out of scope, but happy to reopen if more context is given.

Could a scikit-learn hacker provide an implementation of the Tanimoto kernel?
Since scikit-learn's kernel template requires much more than just implementing the K method (i.e. score a pair of instances), this is super annoying for potential users who are not numerical maths researchers familiar w/ the scikit-learn codebase.

The Tanimoto kernel is for instances which are bitstrings.
The more general version of this kernel is called the Jaccard index (cf. https://en.wikipedia.org/wiki/Jaccard_index).
The Jaccard index can handle bitstrings but also vectors of positive integers, so it might be potentially useful to more scikit-learn users.

Please reopen this issue.

There are some interesting kernels implemented in there:
https://github.com/AustinT/sklearn_tanimoto_gp/blob/main/sklearn_tanimoto_gp/kernels.py