panmiqi/spaCy-DeBERTa

This model is a spaCy NER pipeline built on top of microsoft/deberta-v3, trained on the CoNLL-2003 English NER dataset.

It predicts the four standard CoNLL entity types:

• PER
• ORG
• LOC
• MISC

Training

• Base model: DeBERTa v3
• Dataset: CoNLL-2003 (train + dev)
• Framework: spaCy 3.x with spacy-transformers
• Hardware: A800 GPU

Performance (Dev Set)

The model reaches performance in the expected range for transformer-based CoNLL systems:

• F1 ≈ 96.1
• Precision ≈ 95.6
• Recall ≈ 96.6

Usage (Linux and Mac)

Dependencies (Bash):

pip install "spacy>=3.8.11,<3.9.0" "numpy<2" spacy-transformers huggingface_hub

Installation & Test (Python):

python - << 'PY'
import spacy
from huggingface_hub import snapshot_download

def load_panmiqi_spacy_deberta():
    repo_id = "panmiqi/spacy-deberta"
    local_dir = snapshot_download(repo_id)
    return spacy.load(local_dir)

nlp = load_panmiqi_spacy_deberta()

text = "U.N. official Patrick heads for Beijing."
doc = nlp(text)

print([(ent.text, ent.label_) for ent in doc.ents])
PY