Rakshya 0.2: card, code and licences
Browse files- LICENSE.md +14 -0
- NOTICE.md +54 -0
- README.md +72 -0
LICENSE.md
ADDED
|
@@ -0,0 +1,14 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Licences
|
| 2 |
+
|
| 3 |
+
The files `onnx/model.onnx`, `model.safetensors`, `config.json`, `tokenizer.json`, `tokenizer_config.json` and
|
| 4 |
+
`tagger.json` are the server model of Rakshya. Copyright (c) 2026 Fiste AI. They are licensed under the Creative
|
| 5 |
+
Commons Attribution-NonCommercial 4.0 International licence (CC BY-NC 4.0):
|
| 6 |
+
https://creativecommons.org/licenses/by-nc/4.0/legalcode
|
| 7 |
+
|
| 8 |
+
In short, and without replacing the licence text: you may use, share and adapt the model for non-commercial
|
| 9 |
+
purposes, with credit to "Rakshya by Fiste AI" and a link to https://huggingface.co/fiste-ai/rakshya-server.
|
| 10 |
+
|
| 11 |
+
The model learned from data, and is fine-tuned from a model, that keep their own terms. See `NOTICE.md`. Some of
|
| 12 |
+
that data is for non-commercial or research use only, so there is no commercial licence for this version.
|
| 13 |
+
|
| 14 |
+
The code that runs the model is at https://github.com/fiste-ai/rakshya under the MIT License.
|
NOTICE.md
ADDED
|
@@ -0,0 +1,54 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Third-party notices
|
| 2 |
+
|
| 3 |
+
## The model
|
| 4 |
+
|
| 5 |
+
The server model of Rakshya is fine-tuned from XLM-RoBERTa base by Facebook AI (https://huggingface.co/FacebookAI/xlm-roberta-base). That model is released under the MIT License, and the notice below stays with copies and substantial portions of it.
|
| 6 |
+
|
| 7 |
+
MIT License
|
| 8 |
+
|
| 9 |
+
Copyright (c) Facebook, Inc. and its affiliates.
|
| 10 |
+
|
| 11 |
+
Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated
|
| 12 |
+
documentation files (the "Software"), to deal in the Software without restriction, including without limitation
|
| 13 |
+
the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and
|
| 14 |
+
to permit persons to whom the Software is furnished to do so, subject to the following conditions:
|
| 15 |
+
|
| 16 |
+
The above copyright notice and this permission notice shall be included in all copies or substantial portions of
|
| 17 |
+
the Software.
|
| 18 |
+
|
| 19 |
+
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO
|
| 20 |
+
THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
| 21 |
+
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF
|
| 22 |
+
CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
|
| 23 |
+
IN THE SOFTWARE.
|
| 24 |
+
|
| 25 |
+
## Training data
|
| 26 |
+
|
| 27 |
+
Rakshya was trained on text made up by a generator of ours, in which every person, number and address is invented,
|
| 28 |
+
and on the public sets below. None of them is redistributed here. Two are for non-commercial use and one is for
|
| 29 |
+
research use, which is why the Rakshya weights are non-commercial.
|
| 30 |
+
|
| 31 |
+
| Set | Language | Terms |
|
| 32 |
+
|---|---|---|
|
| 33 |
+
| EverestNER (Niraula and Chapagain, 2022), https://github.com/nowalab/everest-ner | Nepali | non-commercial; commercial use needs the permission of its authors |
|
| 34 |
+
| DanfeNER (Niraula and Chapagain, 2023) | Nepali | non-commercial |
|
| 35 |
+
| IndicPHI, https://huggingface.co/datasets/Sidharth1743/indicphi | Nepali, Hindi, English | MIT |
|
| 36 |
+
| HiNER (Murthy and others, 2022), https://huggingface.co/datasets/cfilt/HiNER-original | Hindi | CC BY-SA 4.0 |
|
| 37 |
+
| CoNLL-2003 (Tjong Kim Sang and De Meulder, 2003) | English | Reuters text, for research use |
|
| 38 |
+
| Few-NERD (Ding and others, 2021) | English | CC BY-SA 4.0 |
|
| 39 |
+
| WNUT-17 (Derczynski and others, 2017) | English | CC BY 4.0 |
|
| 40 |
+
| Broad Twitter Corpus (Derczynski and others, 2016) | English | CC BY 4.0 |
|
| 41 |
+
| Text Anonymization Benchmark (Pilán and others, 2022) | English | MIT |
|
| 42 |
+
| Nemotron-PII, https://huggingface.co/datasets/nvidia/Nemotron-PII | English | CC BY 4.0 |
|
| 43 |
+
| Maskara, https://huggingface.co/datasets/somukandula/maskara-indian-pii-200k | English, Hindi | MIT |
|
| 44 |
+
|
| 45 |
+
Sentences of EverestNER, DanfeNER and HiNER were also used a second time with the names of foreign people and
|
| 46 |
+
places put in, taken from ParaNames (https://github.com/bltlab/paranames, built from Wikidata), and sentences of
|
| 47 |
+
EverestNER and DanfeNER a third time rewritten in Latin letters.
|
| 48 |
+
|
| 49 |
+
The generator takes the names of provinces, districts and local levels from nepali-canonical
|
| 50 |
+
(https://github.com/ritushishir/nepali-canonical), licensed under CC BY 4.0.
|
| 51 |
+
|
| 52 |
+
## Code
|
| 53 |
+
|
| 54 |
+
The code that runs the model is at https://github.com/fiste-ai/rakshya under the MIT License; its notices are there.
|
README.md
ADDED
|
@@ -0,0 +1,72 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language:
|
| 3 |
+
- ne
|
| 4 |
+
- en
|
| 5 |
+
- hi
|
| 6 |
+
license: cc-by-nc-4.0
|
| 7 |
+
library_name: onnx
|
| 8 |
+
pipeline_tag: token-classification
|
| 9 |
+
base_model: FacebookAI/xlm-roberta-base
|
| 10 |
+
tags:
|
| 11 |
+
- token-classification
|
| 12 |
+
- named-entity-recognition
|
| 13 |
+
- pii
|
| 14 |
+
- pii-detection
|
| 15 |
+
- privacy
|
| 16 |
+
- redaction
|
| 17 |
+
- anonymization
|
| 18 |
+
- nepali
|
| 19 |
+
- romanized-nepali
|
| 20 |
+
- devanagari
|
| 21 |
+
- english
|
| 22 |
+
- onnx
|
| 23 |
+
- rakshya
|
| 24 |
+
- fiste-ai
|
| 25 |
+
metrics:
|
| 26 |
+
- f1
|
| 27 |
+
---
|
| 28 |
+
|
| 29 |
+
# Rakshya server (रक्षा)
|
| 30 |
+
|
| 31 |
+
**The large model of [Rakshya](https://huggingface.co/fiste-ai/rakshya): personal data found in Nepali and English text, for servers. By [Fiste AI](https://fiste-ai.github.io/rakshya/).** It has the rules, the 21 labels and the code of the small model, with 278 million weights in place of 40 million. Use the small one in a web page or on a phone, and this one where 1.1 GB and a few tens of milliseconds a sentence are no trouble.
|
| 32 |
+
|
| 33 |
+
## Benchmark
|
| 34 |
+
|
| 35 |
+
882 real texts in seven test sets that no system had seen: news, web pages, government pages and comments. The number is the F1 of exact spans over every kind of personal data in a set. Higher is better.
|
| 36 |
+
|
| 37 |
+
| System | Weights | News (ne) | Web (ne) | Government (ne) | Comments (ne, Latin) | News (en) | Web (en) | News (hi) | Mean |
|
| 38 |
+
|---|---|---|---|---|---|---|---|---|---|
|
| 39 |
+
| **[Rakshya 0.2](https://huggingface.co/fiste-ai/rakshya)** | 40M | **0.878** | 0.802 | 0.731 | **0.529** | 0.803 | 0.703 | **0.804** | 0.750 |
|
| 40 |
+
| **[Rakshya 0.2 server](https://huggingface.co/fiste-ai/rakshya-server)** | 278M | 0.870 | **0.815** | **0.740** | 0.520 | **0.813** | **0.740** | 0.800 | **0.757** |
|
| 41 |
+
| [Gopya 0.1](https://huggingface.co/fiste-ai/rakshya/tree/v0.1) | 34M | 0.855 | 0.683 | 0.636 | 0.233 | 0.729 | 0.563 | 0.725 | 0.632 |
|
| 42 |
+
| [EU PII anonymisation](https://huggingface.co/bardsai/eu-pii-anonimization-multilang) (bards.ai) | 278M | 0.647 | 0.601 | 0.409 | 0.253 | 0.787 | 0.530 | 0.727 | 0.565 |
|
| 43 |
+
| [XLM-R base NER HRL](https://huggingface.co/Davlan/xlm-roberta-base-ner-hrl) (Davlan) | 277M | 0.660 | 0.617 | 0.267 | 0.302 | 0.772 | 0.498 | 0.712 | 0.547 |
|
| 44 |
+
| [XLM-R large, CoNLL-2003](https://huggingface.co/FacebookAI/xlm-roberta-large-finetuned-conll03-english) (FacebookAI) | 559M | 0.651 | 0.574 | 0.200 | 0.265 | 0.783 | 0.481 | 0.760 | 0.531 |
|
| 45 |
+
| [EU PII anonymisation v3](https://huggingface.co/bardsai/eu-pii-anonimization-multilang-v3) (bards.ai) | 308M | 0.144 | 0.704 | 0.328 | 0.162 | 0.605 | 0.709 | 0.528 | 0.454 |
|
| 46 |
+
| [Nym PII multilingual](https://huggingface.co/Wismut/nym-pii-multilingual) (Wismut) | 308M | 0.427 | 0.420 | 0.301 | 0.252 | 0.655 | 0.478 | 0.595 | 0.447 |
|
| 47 |
+
|
| 48 |
+
None of the 25 published systems we ran scores higher than Rakshya 0.2 server on any of the seven sets.
|
| 49 |
+
|
| 50 |
+
All 25 systems, the scores by kind, the false alarms and the method are in the [full benchmark](https://github.com/fiste-ai/rakshya/blob/main/BENCHMARK.md).
|
| 51 |
+
|
| 52 |
+
## Use it
|
| 53 |
+
|
| 54 |
+
```python
|
| 55 |
+
# pip install git+https://github.com/fiste-ai/rakshya
|
| 56 |
+
from rakshya import Rakshya
|
| 57 |
+
|
| 58 |
+
pii = Rakshya(model="server") # fetches this model once, 1.1 GB
|
| 59 |
+
pii.redact("मेरो नाम सीता थापा हो, फोन ९८४१२३४५६७।")
|
| 60 |
+
```
|
| 61 |
+
|
| 62 |
+
From a shell: `rakshya --server "some text"`. On a processor it reads a sentence of about 21 tokens in 40 ms on one thread and 15 ms on four (a laptop processor, Intel Core i5-13420H). With `onnxruntime-gpu` installed, `Rakshya(model="server", providers=["CUDAExecutionProvider"])` runs it on a graphics card.
|
| 63 |
+
|
| 64 |
+
The repository also loads as a plain token classifier in `transformers`. Read that way the model gives its tags alone: the rules and the cut at the case marker are in the package.
|
| 65 |
+
|
| 66 |
+
The scores above were measured with these weights on a graphics card. The ONNX file here holds the same weights and gives the same tag on 10,259 of 10,259 tokens of a check.
|
| 67 |
+
|
| 68 |
+
## Limits
|
| 69 |
+
|
| 70 |
+
The limits of the small model hold here too: it is a helper and not a guarantee, Nepali in Latin letters is the hardest, Hindi is experimental, and identity numbers are tested mostly on made-up text. See the [card of the small model](https://huggingface.co/fiste-ai/rakshya#limits).
|
| 71 |
+
|
| 72 |
+
Weights CC BY-NC 4.0, code MIT; see `LICENSE.md` and `NOTICE.md`. Part of the training data is for non-commercial use, so there is no commercial licence for this version. Questions: [contact us](https://fiste-ai.github.io/contact/).
|