| --- |
| license: apache-2.0 |
| language: |
| - en |
| base_model: |
| - distilbert/distilbert-base-uncased |
| tags: |
| - token-classification |
| - ner |
| - privacy |
| - pii-detection |
| - distilbert |
| --- |
| |
| # PrivyShield NER Model |
|
|
| A fine-tuned DistilBERT model for detecting personally identifiable information (person names and addresses) in unstructured text; the core "separate model" component of **PrivyShield**, an on-device privacy protection tool built for OSDHack 2026. |
|
|
| ## Model Details |
|
|
| - **Base model:** `distilbert-base-uncased` |
| - **Task:** Token Classification (Named Entity Recognition, BIO tagging scheme) |
| - **Labels:** `O`, `B-PERSON_NAME`, `I-PERSON_NAME`, `B-ADDRESS`, `I-ADDRESS` |
| - **Training data:** 7,500 synthetically generated, auto-labeled sentences (template + entity-pool based generation; no manual annotation), covering Indian names, addresses, cities, and localities |
| - **Training:** 5 epochs, fine-tuned end-to-end on the labeled dataset |
| - **Formats provided:** |
| - `ner_model.onnx` : exported for fast local inference |
| - `ner_model/` : original PyTorch checkpoint + tokenizer files |
|
|
| ## Why this model exists |
|
|
| Regex can reliably catch structured PII (card numbers, emails, Aadhaar/PAN formats), but it cannot catch **names and addresses**, which don't follow a fixed pattern. This model fills that specific gap as part of PrivyShield's layered detection pipeline; regex handles structured formats, this model handles unstructured entities, and both run entirely on-device. |
|
|
| ## Intended use |
|
|
| Designed to run locally (via ONNX Runtime) as part of a real-time screen-content privacy scanner. Not intended as a general-purpose NER model; it's scoped specifically to `PERSON_NAME` and `ADDRESS` detection for this use case, trained primarily on Indian name/address patterns. |
|
|
| ## Limitations |
|
|
| - Trained on synthetic data; real-world OCR noise (misspellings, broken formatting) may reduce accuracy |
| - English-only |
| - Name/address pools are India-centric; may generalize less well to other regions' naming/address conventions |
| - Not evaluated for adversarial or out-of-distribution inputs |
|
|
| ## Usage |
|
|
| ```python |
| from huggingface_hub import hf_hub_download |
| import onnxruntime as ort |
| |
| model_path = hf_hub_download( |
| repo_id="aditrynacode/privyshield-ner", |
| filename="ner_model.onnx" |
| ) |
| session = ort.InferenceSession(model_path) |
| ``` |
|
|
| ## Project |
|
|
| Part of **PrivyShield**, submitted to [OSDHack 2026](https://hack.osdc.dev/) (Open Source Developers Community). |
|
|
| Main repository: [PrivyShield on GitHub](https://github.com/aditrynacode/PrivyShield) |
