| license: other | |
| license_name: parda-mixed | |
| tags: [onnx, pii, redaction, ocr, gliner, yolo, easyocr, india] | |
| # Parda web models | |
| ONNX models used by the Parda web app, which redacts personal data in Indian documents | |
| (English, Hindi, Kannada) **entirely in the browser**. Nothing is uploaded. | |
| | File | What | Licence | | |
| |---|---|---| | |
| | `gliner/model_int8_embed.onnx` | GLiNER PII model fine-tuned on synthetic Indian documents (embedding table 8-bit) | Apache-2.0 (base: urchade/gliner_multi_pii-v1, mDeBERTa-v3 MIT) | | |
| | `yolo/parda-yolo.onnx` | YOLO11n for faces, signatures, QR codes, stamps | **AGPL-3.0** (Ultralytics YOLO11); faces trained with **CelebA: non-commercial research use only** | | |
| | `easyocr/*.onnx` | EasyOCR CRAFT detector + en / hi+en / kn+en recognizers, converted to ONNX | Apache-2.0 (EasyOCR) | | |
| **Use: free, non-commercial research and demonstration only.** Commercial use would require retraining the face | |
| detector without CelebA and complying with AGPL-3.0. Trained only on synthetic documents; always review the output | |
| before sharing a document. Code and benchmarks: https://github.com/chetanjagan/Parda | |