points
Ideally the hold out set is randomly sampled. You need a hold out set of Inklings pretraining run.
Should we trust that the dataset of each company is big and diverse enough to be representative? I don’t hope too much…