ner-pk: named entity recognition for Pakistani news
A fine-tuned XLM-RoBERTa-large model that finds 20 kinds of entities in Pakistani news text: people, government bodies, political parties, laws, CNIC numbers, money and more.
Training data came from two LLM labellers with agreement filtering, plus synthetic pattern sentences, trained in two stages. A regex rules layer and a document-level consistency pass sit on top, and it ships as a CPU-ready command-line tool that reads text, JSONL or CoNLL.
PyTorch / Transformers / LLM labelling / regex / pytest
Try a browser version below →| Micro-F1 | 0.918 |
|---|---|
| Macro-F1 | 0.903 |
| Precision | 0.916 |
| Recall | 0.919 |
| DATE F1 | 0.962 |
| COUNTRY F1 | 0.953 |