SpaCy model training and eval metrics

[solpunit]

I am using the SpaCy NER model for training. I have around 20K samples in total for 7-8 labels. I train the model but my eval set contains words which is never seen during the model training. I see the model results are bad, my F1 and recall are in 20. Is this something which is expected? Does the NER model perform bad at samples it have never seen? I am using the lg model and doing fine-tuning.