Closed spozdn closed 1 week ago
I think the basic issue here is not one of documentation, it's more that the default parameters shouldn't be parameters that get you (close to) state of the art accuracy, but they should be parameters that provide a good potential with decent speed (both training and inference) in practice. Then we can have a section of the docs to explain how to change parameters to get better accuracy (pretty much as in the other architectures we have)
Let's try to come up with a better set of defaults in this PR
Given the discussion with @frostedoyster, we agreed to separate the issues of incomplete documentation and a demand for a more lightweight and fast model. The latter is discussed here - https://github.com/lab-cosmo/metatrain/issues/273, and, thus, is uncoupled from this pr.
Adding an additional clarification to PET documentation.
📚 Documentation preview 📚: https://metatrain--268.org.readthedocs.build/en/268/