extending the ner with a new entity

I am running the following:
prodigy ner.batch-train 3_ent_bis_dataset en_core_web_lg --output bis-model --label "BIS-PRODUCT, BIS-CONCEPT, BIS-OTHER" --eval-split 0.2 --n-iter 6 --batch-size 8

and I get the following output:


the accuracy is very poot (29%), what am I doing wrong?

note that before the ner.batch-train action, I have run the following:

prodigy ner.manual 3_ent_bis_dataset en_core_web_lg data/db2_bis_data.txt --label "BIS-PRODUCT, BIS-CONCEPT, BIS-OTHER" &

What's BIS-PRODUCT, BIS-CONCEPT? Maybe the distinction is difficult to learn. Also, you probably don't want to be starting off with a pre-trained model here, because your new labels are possibly conflicting with the labels already present in the model, like PRODUCT or something else. So you're constantly fighting existing weights and your model isn't learning anything.