Views
No views yet
apg / arabic-pii-py, a local
Arabic/GCC PII gateway. They are redistributed quantizations of existing
open models — no new training, same weights. onnxruntime
runs them with no PyTorch / Transformers at inference time, which is why
apg stays lightweight.apg uses them as the sole authority for PERSON / ORGANIZATION detection
(and LOC → ADDRESS); it has no name/org regex.| folder | language | entities | from | size |
|---|---|---|---|---|
arabic-ner/ | Arabic | PERSON, ORGANIZATION, LOCATION, … | hatmimoha/arabic-ner | ~105 MB |
en-ner/ | English | PER, ORG, LOC, MISC | dslim/bert-base-NER | ~104 MB |
model_quantized.onnx + tokenizer.json + config.json.arabic-ner/)en-ner/)bert-base-cased, fine-tuned on the CoNLL-2003 English NER dataset.model_quantized.onnx variant): identical weights, ~4× smaller
on disk, fast on CPU via onnxruntime. Nothing was re-trained or fine-tuned..onnx files:arabic-ner/model_quantized.onnx 8e1c68e5d812835b8d0770e0693764d8d96276971585d77e9b59c97c5217ded9
en-ner/model_quantized.onnx b324e829f1fad3b897f926d1a1d1372803c6d04546831a3a2ed103652b916adfpip install "apg[ner]"apg auto-downloads these on first use (cached under ~/.cache/huggingface).
You can also load them directly with onnxruntime + tokenizers.1@misc{hatmimoha_arabic_ner, title={Arabic Named Entity Recognition Model}, author={Hatim Mohamed}, howpublished={\url{https://huggingface.co/hatmimoha/arabic-ner}}}
2@inproceedings{tjong-kim-sang-de-meulder-2003-conll, title={Introduction to the {C}o{NLL}-2003 Shared Task}, author={Tjong Kim Sang, Erik F. and De Meulder, Fien}, booktitle={CoNLL}, year={2003}}