Order of labels impacts detection confidence scores

[saibhamidipati]

I'm using the PII model using the following script:

model = GLiNER.from_pretrained("urchade/gliner_multi_pii-v1")
model.eval()
entities = model.predict_entities(text, labels)

for entity in entities:
    print(entity["text"], "=>", entity["label"], "=>", entity["score"] )

If my order of labels changes, I get a different score for the detected entities. Why is this happening?

[urchade]

Hi, it is because the model uses positional encoding
Bi-encode version of gliner of @Ingvarstep should not have this effect

[saibhamidipati]

Thanks for your response, but I don't find bi-encoder version for gliner_multi_pii-v1. Is it available yet?