OPENER-ZS — zero-shot open-world NER

This repo hosts the contrastively fine-tuned Matryoshka embedder of OPENER, used in its zero-shot operating point (OPENER-ZS). It needs no target labels: you give it text and the candidate type names.

OPENER is a three-stage pipeline: GLiNER-L (frozen detector) → this fine-tuned embedder → label-name prototypes (+ transductive refine + detector fusion).

Usage (turnkey)

pip install opener-ner

from opener import OpenerZS

m = OpenerZS.from_pretrained("Thibault-GAREL/opener-zs")   # + auto-downloads GLiNER-L
ents = m.predict(
    "Marie Curie discovered radium at the University of Paris.",
    labels=["person", "discovery", "organization", "location"],
)

Each detected mention is typed by cosine similarity to the nearest label-name prototype, transductively refined on the unlabelled inputs, and fused with GLiNER's own zero-shot label (score += β·detector_score, β = 0.05).

How it was trained

• Base: nomic-ai/nomic-embed-text-v1.5 (Matryoshka, Apache-2.0).
• Stage 1: Triplet margin loss (margin 1) on CoNLL-2003 training spans.
• Stage 2: error-driven hard-negative mining (8000 triplets, 65% hard, 3 epochs) on the type pairs the model most often confuses.
• Mentions are embedded in context ([ENT] … [/ENT], task prefix classification:).

Results (13-dataset benchmark)

OPENER-ZS reaches 39.4 end-to-end AMI, the best of the compared zero-shot systems (ahead of GLiNER-L 38.9 and a zero-shot OWNER 34.3), at ~180 ms / 1.7 Wh per sentence.

License & credits

MIT. Base embedder Apache-2.0 (nomic-ai/nomic-embed-text-v1.5); detector GLiNER (urchade/gliner_large-v2.1). From the OPENER research code (LyRIDS Symposium 2026).