mdeberta-ner-ontonotes5-onnx

mDeBERTa NER OntoNotes5 (ONNX INT8)

ONNX INT8 quantized export of rustemgareev/mdeberta-ner-ontonotes5 for browser-based inference via Transformers.js or ONNX Runtime Web.

Key Features

• Multilingual (mDeBERTa-v3-base-lite encoder)
• Detects lowercase entities ("amsterdam", "paris" without capitalization)
• OntoNotes5 labels: GPE, LOCATION, FACILITY, ORGANIZATION, PERSON, DATE, EVENT, etc.
• Standard token-classification — works with pipeline('token-classification') in Transformers.js

Model Details

• Parameters: 211M
• Size: 258 MB (INT8 quantized), 805 MB (fp32)
• Labels: GPE (cities/countries), LOCATION (natural features), FACILITY (buildings/landmarks), ORGANIZATION (companies/hotels), PERSON, DATE, EVENT, etc.
• Training: OntoNotes5 English + multilingual transfer

Usage with Transformers.js

import { pipeline } from '@huggingface/transformers';

const ner = await pipeline('token-classification', 'Berk/mdeberta-ner-ontonotes5-onnx', {
  quantized: true,
});

const result = await ner('I want to book best western in amsterdam');
// [{ word: 'best western', entity_group: 'ORGANIZATION' }, { word: 'amsterdam', entity_group: 'GPE' }]

Performance

License

MIT (same as original model)