I don't think this fixes thins as for all cases:

The issue is this seldon core v2 huggingface demo explainer is calling a pipeline so we can't get the metadata, so the blackbox explainer adds content-type='str' by default which no longer works as we need content_type="hf".

if content_type is provided in payload it always has precedene over any in model-settings.json.

The metadata content_type also needs updating: https://github.com/SeldonIO/MLServer/blob/master/runtimes/huggingface/mlserver_huggingface/metadata.py from str to hf everywhere

As far as I can see, that metadata only applies to the input-level metadata but leaves the request-level metadata unset - so we can leave it as it is.