OPENER-Sup — supervised open-world NER

This repo hosts the contrastively fine-tuned Matryoshka embedder of OPENER, used in its supervised operating point (OPENER-Sup): the most accurate setting, where a tiny balanced linear head is fitted on your labelled spans.

Pipeline: GLiNER-L (frozen detector) → this fine-tuned embedder → LinearSVC (balanced). The embedder is identical to Thibault-GAREL/opener-zs; only the typing head differs (a trained probe vs. label-name prototypes).

Usage (turnkey)

pip install opener-ner

from opener import OpenerSup

m = OpenerSup.from_pretrained("Thibault-GAREL/opener-sup")   # + auto-downloads GLiNER-L

# fit the typing head on YOUR data: (start_char, end_char, label) per sentence
texts  = ["Marie Curie discovered radium."]
annots = [[(0, 11, "person"), (23, 29, "element")]]
m.fit(texts, annots)

ents = m.predict("Albert Einstein formulated relativity.")
m.save_head("opener_sup_head.joblib")   # reuse with m.load_head("opener_sup_head.joblib")

How it was trained

Same embedder as OPENER-ZS:

• Base nomic-ai/nomic-embed-text-v1.5; Triplet contrastive (CoNLL-2003) + hard-negative mining.
• The typing head is a LinearSVC(class_weight="balanced") fitted one-vs-rest on the target's labelled spans (in this embedding space).

Results (13-dataset benchmark)

OPENER-Sup is the most accurate system overall: 40.2 end-to-end AMI and 62.5 on gold mentions (vs a zero-shot OWNER 43.0), while staying frugal (~143 ms / 1.6 Wh).

License & credits

MIT. Base embedder Apache-2.0; detector GLiNER. From the OPENER research code (LyRIDS Symposium 2026).