Hi Dominik
Creating the csv and json file will require some manual implementation from your side. This thread does give some hint on how to export to CSV.
There is some ongoing work to have an easier way to do this in future release.
I did not test but it might also be possible to build an info table with the logged data and pass it to the data infotable of datasetRef in the training Thing CreateJob() service. Similarly create a info table with the metada and pass it to the metadata infotable of datasetRef.
I have done that with scoring so I am thinking it could work with training too, though I do not know how well this would scale. For large amount of data the csv option is probably better.
Regarding how much data you need for training, there is not a simple answer.
This depends on the number of fields and the correlation between them.
For example if you try to predict z, which is related to x and y with z=x+2y , then 5 to 10 points would be enough.
If you have a lot of fields with no obvious relations, then you do need a much larger number of fields to establish . https://www.ptc.com/en/support/article?n=CS255070 indicate at least 30 records per field, but with machine learning the more records the better.
Hope this helps
Christophe
