Orchive dataset

[nkundiushuti]

the annotated part of orchive is quite small (I was not aware)
they provide, however, a quite large unannotated dataset with cropped calls

[david-rx]

Hey @nkundiushuti @GaganNarula @mil-ad - what do we think of including this in the end? For vocal repertoire, we don't need a training set, clustering/retrieval are the primary evals for many repertoire datasets anyways. And, this would be our hardest, most diverse repertoire task and only marine one. Maybe we could just add a split for "all data" since we will probably not train on it and we want to use everything, and maybe a split with all data but having at least 3 examples per class?

Lmk what you think and happy to help with that if you agree

EDIT - hold that thought, I think there is a much better one