Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
|
@@ -1,3 +1,36 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
# NERPA — Fine-Tuned GLiNER2 for PII Anonymisation
|
| 2 |
|
| 3 |
A fine-tuned [GLiNER2 Large](https://huggingface.co/fastino/gliner2-large-v1) (340M params) model trained to detect Personally Identifiable Information (PII) in text. Built as a flexible, self-hosted replacement for AWS Comprehend at [Overmind](https://overmindai.com).
|
|
@@ -143,8 +176,33 @@ The inference pipeline in `anonymise.py`:
|
|
| 143 |
- **GLiNER2 version:** Requires `gliner2>=1.2.4`. Earlier versions had a bug where entity character offsets mapped to token positions instead of character positions; this is fixed in 1.2.4+.
|
| 144 |
- **Device:** Automatically uses CUDA > MPS > CPU.
|
| 145 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 146 |
## Citation
|
| 147 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 148 |
Built by [Akhat Rakishev](https://github.com/workhat) at [Overmind](https://overmindai.com).
|
| 149 |
|
| 150 |
-
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language:
|
| 3 |
+
- en
|
| 4 |
+
license: apache-2.0
|
| 5 |
+
library_name: gliner2
|
| 6 |
+
tags:
|
| 7 |
+
- named-entity-recognition
|
| 8 |
+
- ner
|
| 9 |
+
- pii
|
| 10 |
+
- anonymisation
|
| 11 |
+
- gliner
|
| 12 |
+
- gliner2
|
| 13 |
+
- token-classification
|
| 14 |
+
- privacy
|
| 15 |
+
datasets:
|
| 16 |
+
- synthetic
|
| 17 |
+
base_model: fastino/gliner2-large-v1
|
| 18 |
+
model-index:
|
| 19 |
+
- name: NERPA
|
| 20 |
+
results:
|
| 21 |
+
- task:
|
| 22 |
+
type: token-classification
|
| 23 |
+
name: Named Entity Recognition
|
| 24 |
+
metrics:
|
| 25 |
+
- type: precision
|
| 26 |
+
value: 0.93
|
| 27 |
+
name: Micro-Precision
|
| 28 |
+
- type: recall
|
| 29 |
+
value: 0.90
|
| 30 |
+
name: Micro-Recall
|
| 31 |
+
pipeline_tag: token-classification
|
| 32 |
+
---
|
| 33 |
+
|
| 34 |
# NERPA — Fine-Tuned GLiNER2 for PII Anonymisation
|
| 35 |
|
| 36 |
A fine-tuned [GLiNER2 Large](https://huggingface.co/fastino/gliner2-large-v1) (340M params) model trained to detect Personally Identifiable Information (PII) in text. Built as a flexible, self-hosted replacement for AWS Comprehend at [Overmind](https://overmindai.com).
|
|
|
|
| 176 |
- **GLiNER2 version:** Requires `gliner2>=1.2.4`. Earlier versions had a bug where entity character offsets mapped to token positions instead of character positions; this is fixed in 1.2.4+.
|
| 177 |
- **Device:** Automatically uses CUDA > MPS > CPU.
|
| 178 |
|
| 179 |
+
## Acknowledgements
|
| 180 |
+
|
| 181 |
+
This model is a fine-tuned version of [GLiNER2 Large](https://huggingface.co/fastino/gliner2-large-v1) by [Fastino AI](https://fastino.ai). We thank the GLiNER2 authors for making their model and library openly available.
|
| 182 |
+
|
| 183 |
## Citation
|
| 184 |
|
| 185 |
+
If you use NERPA, please cite both this model and the original GLiNER2 paper:
|
| 186 |
+
|
| 187 |
+
```bibtex
|
| 188 |
+
@misc{nerpa2025,
|
| 189 |
+
title={NERPA: Fine-Tuned GLiNER2 for PII Anonymisation},
|
| 190 |
+
author={Akhat Rakishev},
|
| 191 |
+
year={2025},
|
| 192 |
+
url={https://huggingface.co/OvermindLab/nerpa},
|
| 193 |
+
}
|
| 194 |
+
|
| 195 |
+
@misc{zaratiana2025gliner2efficientmultitaskinformation,
|
| 196 |
+
title={GLiNER2: An Efficient Multi-Task Information Extraction System with Schema-Driven Interface},
|
| 197 |
+
author={Urchade Zaratiana and Gil Pasternak and Oliver Boyd and George Hurn-Maloney and Ash Lewis},
|
| 198 |
+
year={2025},
|
| 199 |
+
eprint={2507.18546},
|
| 200 |
+
archivePrefix={arXiv},
|
| 201 |
+
primaryClass={cs.CL},
|
| 202 |
+
url={https://arxiv.org/abs/2507.18546},
|
| 203 |
+
}
|
| 204 |
+
```
|
| 205 |
+
|
| 206 |
Built by [Akhat Rakishev](https://github.com/workhat) at [Overmind](https://overmindai.com).
|
| 207 |
|
| 208 |
+
Overmind is infrastructure to make agents more reliable. Learn more at [overmindai.com](https://overmindai.com).
|