Hi,
I created a model to read training and testing data from 2 respective cosmos db tables and used an sql transformation block to rename the column names and used it for training and in score model as shown in snapshot (modela.jpg) . The Predicted parameter (scored label) is very much different and incorrect (refer onlysqltrans.jpg)
When I use a "convert to csv" block connected to output of sql transformation block (refer modelb.jpg for model diagram ) and then use it in training , I get expected results ( refer withcsvblock.jpg) .
The Mean Absolute error in first case was 51.2 while in modelb was only 0.09
I used a convert to dataset block after sql transform block and used that to connect to training model
but that too gave the same result as modela output
In case you want to see what sql transformation i used



select "['CombiTimeTable.y[1]']" as av1 ,"['CombiTimeTable.y[2]']" as av2,"['CombiTimeTable.y[3]']" as av3,"['CombiTimeTable.y[4]']" as av4,"['CombiTimeTable.y[5]']" as av5,"['CombiTimeTable.y[6]']" as av6,"['CombiTimeTable.y[7]']" as av7,"['CombiTimeTable.y[8]']" as av8,"['CombiTimeTable.y[9]']" as av9 from t1
Thanks