It is hard to tell because “what is expected” depends largely on the domains of your datasets. We advise that you inspect these accounts to gauge these percentages and determine how to interpret them.

Comparing the percentages (or better, the distributions) of the different groups is IMHO much more meaningful than labeling individual accounts. You may not be able to say “dataset 3 has high bot activity”, but you should be able to say “dataset 3 displays higher levels of automation than dataset 1”.

Finally, the score threshold of 3 is somewhat arbitrary. Consider setting a threshold on the CAP, so that it is easier to interpret (as a p-value) as explained in the FAQ on the Botometer website.

Hope this helps,
-Botometer Team