Evaluation pipeline for metadata quality assessment and consider ensembling?

[PGijsbers]

We want to asses metadata quality, and answer multiple questions about their quality (see also #16). For this, we preferably also want an evaluation pipeline -- though it may be much harder as labels aren't clearly defined. It should at least be explored. The pipeline should also easily allow for ensembles of models. Since metadata quality is assessed only roughly once per dataset, it is okay if this pipeline uses more resources.