Metadata: More flexible specification of metadata type

[neubig]

Currently, string-based meta-data is treated as nominal if there are 20 or fewer examples, and text if there are 21 or more examples:

zeno/zeno/util.py

Lines 46 to 47 in 808f4b2

	if len(unique) < 21:
	return MetadataType.NOMINAL

This can be limiting. For example, I have a use case where I want to run slice finder on clusters found from a text clustering algorithm. As-is, this means that I am limited to 20 or fewer clusters, which is probably not granular enough to make these clusters meaningful.

One possible design would be if distill functions could (optionally) specify the metadata type like this:

return DistillReturn(distill_output=document_clusters, metadata_type=MetadataType.NOMINAL)

If no type is specified we could fall back to the current behavior (but that also should potentially be documented somewhere).