Model gives output even for SEP token

Models
1

Hi ,

So I trained a Ner model using Bert.
Just wanted to understand if any of our labels would be predicted for [CLS] or [SEP] token ?
Is it normal or is there something I might have gone wrong with ?

While training I had assigned a -100 label to them so as they would be ignored during training .

A sample o/p when I run my prediction on a sentence , so CLS and SEP being encoded into 101 and 102 while encoding .

    tok  preds
0	101	    0
1	1045	0
2	2215	0
3	2000	0
4	4965	0
5	1037	0
6	2417	0
7	18059	0
8	2007	0
9	16380	1
10	1021	2
11	102	    2

Any inputs would be helpful for my understanding .