File size: 11,009 Bytes
14cd859 86b2247 14cd859 86b2247 14cd859 86b2247 14cd859 86b2247 14cd859 86b2247 14cd859 86b2247 14cd859 86b2247 14cd859 86b2247 14cd859 86b2247 14cd859 86b2247 14cd859 86b2247 14cd859 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 | ---
license: apache-2.0
language:
- en
- de
- fr
- es
- nl
- pt
- it
pipeline_tag: token-classification
tags:
- pii
- privacy
- ner
- redaction
- anonymization
- multilingual
- on-device
- browser
- typescript
datasets:
- wikimedia/wikipedia
- nvidia/Nemotron-PII
- OpenAssistant/oasst2
- AmazonScience/massive
- google-research-datasets/taskmaster1
metrics:
- f1
model-index:
- name: veil-0.3.1
results:
- task:
type: token-classification
name: Person name detection
dataset:
type: eriktks/conll2003
name: CoNLL-2003 English
split: test
metrics:
- type: f1
value: 73.1
name: F1, PERSON, exact span
- task:
type: token-classification
name: Person name detection
dataset:
type: leondz/wnut_17
name: WNUT-17
split: test
metrics:
- type: f1
value: 51.5
name: F1, PERSON, exact span
- task:
type: token-classification
name: Person name detection
dataset:
type: eriktks/conll2002
name: CoNLL-2002 Spanish
config: es
split: test
metrics:
- type: f1
value: 75.0
name: F1, PERSON, exact span
- task:
type: token-classification
name: Person name detection
dataset:
type: eriktks/conll2002
name: CoNLL-2002 Dutch
config: nl
split: test
metrics:
- type: f1
value: 70.5
name: F1, PERSON, exact span
- task:
type: token-classification
name: Person name detection
dataset:
type: ildpil/text-anonymization-benchmark
name: Text Anonymization Benchmark
split: test
metrics:
- type: f1
value: 88.1
name: F1, PERSON, leading title ignored
---
# Veil
Veil finds names, addresses, phone numbers and other personal data in text, so you can mask them before the text is
stored, logged or sent to another service. The model is 1.18 MB. It runs in the browser or in Node, with no
dependencies and no network calls.
**[Try the demo](https://tars-technologies.github.io/veil/).** The model runs in your browser, so the text you type
stays on your machine. The library, training code and evaluation scripts are on [GitHub](https://github.com/Tars-Technologies/veil).
```ts
import { createVeil } from "@tars-technologies/veil"
import { loadModel } from "@tars-technologies/veil/model"
const veil = createVeil({ model: loadModel() })
await veil.mask({
text: "I'm Laura Bianchi, call me on +33 6 39 98 12 34 or write to laura.bianchi@example.com",
types: ["PERSON", "PHONE", "EMAIL"]
})
// "I'm [NAME], call me on [PHONE] or write to [EMAIL]"
```
## Install
```sh
npm install github:Tars-Technologies/veil#v0.3.1
```
The package ships TypeScript source. Use it with a bundler or runtime that compiles TypeScript, such as Vite,
esbuild, tsx or Bun. Node's built-in TypeScript support does not compile packages in `node_modules`, so in Node run
your code with tsx, as an ES module (`"type": "module"` or a `.mts` file).
To load the model straight from this page:
```ts
const veil = createVeil({
model: "https://huggingface.co/tars-technologies/veil/resolve/main/veil-0.3.1.bin"
})
```
## Usage
`detect` returns the spans and `mask` returns the text with each span replaced. Both are async.
```ts
await veil.mask({ text: "my manager Daniel Okafor told me to call" })
// "my manager [NAME] told me to call"
await veil.detect({ text: "call Anna Schmidt" })
// [{ start: 5, end: 17, type: "PERSON" }]
await veil.mask({ text: "Summer", context: "What is your name?", expectName: true })
// "[NAME]"
```
`model` takes the model bytes or a URL to `veil-0.3.1.bin`. `loadModel()` returns the bytes from inside the package,
which suits servers and serverless functions. It adds about 1.6 MB of base64 to a bundle, so in a browser load the
file by URL instead, through `@tars-technologies/veil/browser`. That entry point has the same API and runs the model
in a Web Worker, so the page does not wait for it. With Vite:
```ts
import { createVeil } from "@tars-technologies/veil/browser"
import modelUrl from "@tars-technologies/veil/veil-0.3.1.bin?url"
const veil = createVeil({ model: modelUrl })
```
| Option | Effect |
|---|---|
| `types` | Labels to find. Default `["PERSON"]`. |
| `threshold` | Between 0 and 1, not including 0 or 1. Default 0.5. A lower value masks more names, and more words that are not names. |
| `context` | The text that came before, for example the question this text answers. |
| `expectName` | The text answers a name question. A short answer is masked as a whole. |
| `always` | Names to mask wherever they occur as whole words, for example a name the user gave earlier. The spans have the label `PERSON`. |
| `never` | Terms to leave alone, for example your brand, product or assistant names. |
| `replacement` | One string that replaces every span. By default each label has its own: `[NAME]`, `[ADDRESS]`, `[DATE OF BIRTH]`, `[ID]`, `[FINANCIAL]`, `[PHONE]`, `[EMAIL]`. |
Pass options to `detect` or `mask` per call, or once as `createVeil({ model, defaults: { ... } })`. A call's options
override the defaults.
One call takes up to 20,000 characters and 3,000 tokens (words and punctuation marks), `context` included. For text
with more tokens, use
`detectChunked(veil, text, options)` from `@tars-technologies/veil/chunked`. It splits the text into overlapping calls
and returns offsets into the full text. The 20,000-character limit still applies. All offsets are UTF-16.
Veil throws on errors and never returns partial results. For a privacy feature, the safe response to an error is
usually to not store the text.
### React
```tsx
import { useMask, VeilProvider } from "@tars-technologies/veil/react"
<VeilProvider veil={veil}>…</VeilProvider>
const { data, spans, isPending, error } = useMask({ text: draft, types: ["PERSON", "EMAIL"] })
```
`useMask` masks the text again each time its options change, which suits a live preview. To mask on send, call
`useVeil().mask` in the handler.
## Model
| | |
|---|---|
| Author | Vaibhav Kubre, Tars Technologies |
| Labels | `PERSON`, `ADDRESS`, `DOB`, `ID`, `FINANCIAL`, `PHONE`, `EMAIL` |
| Architecture | Hashed character n-gram embeddings (32,768 buckets × 64, 2-bit codes), word-shape features and three word-list flags, into a 2-layer bidirectional LSTM with 96 units. LSTM and output weights are int8 with one scale per row. |
| Size | 1,177,096 bytes (`veil-0.3.1.bin`) |
| Languages | Latin script. Most training text is English, German, French, Spanish, Dutch, Portuguese or Italian. |
| Input | Windows of 256 tokens with a 64-token overlap. Up to 300 characters of `context`. |
| Licence | Apache-2.0 for code and model. `NOTICE` credits the data sources. The word lists keep their source licences (Wikidata CC0, GeoNames CC BY 4.0). |
The model file carries its name, version and format in a header and ends with a checksum. The library reads format 3
and rejects a file whose checksum does not match. Each new model gets a new file name.
## Evaluation
We measured `PERSON` on five public test sets. None of them is in the training data. The model that labelled the
Wikipedia rows was trained on the CoNLL-2003 English and CoNLL-2002 Spanish and Dutch training splits; their test
splits were not used. All models use their default settings. F1 on exact span match, except that the TAB row ignores
a leading title (see below):
| Test set | Names | Veil 0.3.1 (1.2 MB) | spaCy sm (15 MB) | newmean nano (2.6 MB) | newmean small (7.1 MB) | newmean large (20.7 MB) |
|---|---|---|---|---|---|---|
| CoNLL-2003 English, news | 1,617 | **73.1** | 66.9 | 43.2 | 45.1 | 50.7 |
| WNUT-17, social media | 429 | **51.5** | 27.0 | 22.1 | 31.0 | 35.0 |
| CoNLL-2002 Spanish, news | 735 | **75.0** | 68.2 | 51.2 | 57.1 | 64.6 |
| CoNLL-2002 Dutch, news | 1,098 | **70.5** | 62.4 | 53.6 | 50.2 | 63.0 |
| TAB, court judgments | 1,018 | **88.1** | 52.1 | 42.9 | 58.3 | 71.0 |
TAB includes titles in its names ("Mrs Karin Schieck") and Veil does not, so the TAB row ignores a leading title.
With strict matching, Veil scores 10.1 and spaCy 29.7.
F1 does not tell the whole story for masking. A missed name is a leak, while an extra mask costs little. This table
counts the names that had no character masked at all (lower is better):
| Test set | Veil 0.3.1 | spaCy sm | newmean large |
|---|---|---|---|
| CoNLL-2003 English | 526 | 543 | **487** |
| WNUT-17 | 250 | 259 | **246** |
| CoNLL-2002 Spanish | 202 | **130** | 139 |
| CoNLL-2002 Dutch | 415 | 393 | **227** |
| TAB | 93 | 199 | **64** |
newmean large misses the fewest names on four of the five sets, but masks many words that are not names. Of the characters Veil masks, 84 to
93 % are part of a name. For spaCy that figure is 29 to 80 %, and for newmean 27 to 85 %. Versions: spaCy 3.8 with
the `*_core_*_sm` 3.8.0 models, and `@newmean-ai/pii` 1.6.0 with model 1.1.0, counting `PERSON` and `PERSON_L2`.
[`evaluation/`](https://github.com/Tars-Technologies/veil/blob/main/evaluation/README.md) has the full results and the scripts to reproduce them.
## Training data
The training data includes these public datasets:
| Source | Rows | Licence |
|---|---|---|
| Wikipedia articles (DE, FR, EN), labelled by a larger model | 181,689 | CC BY-SA 4.0 |
| Wikipedia talk pages (FR, NL, ES, PT, DE, IT) | 40,371 | CC BY-SA 4.0 |
| ABCD customer-service dialogues | 74,792 | MIT |
| Taskmaster dialogues | 39,723 | CC BY 4.0 |
| OASST2 assistant conversations | 35,793 | Apache-2.0 |
| MASSIVE utterances | 43,504 | CC BY 4.0 |
| Stack Exchange posts | 8,455 | CC BY-SA 4.0 |
| French public-service texts | 7,206 | Licence Ouverte 2.0 |
| NVIDIA Nemotron-PII | 96,050 | CC BY 4.0 |
The word-list flags come from three lists in `training/assets/gazetteer.json.gz`: given names and surnames,
places from GeoNames, and common words from the training text. The training set itself is not
published. [`training/`](https://github.com/Tars-Technologies/veil/blob/main/training/README.md) has the trainer, the checkpoint and the export script, which rebuilds
`veil-0.3.1.bin` byte for byte.
## Limitations
- Veil misses names. In the news test sets, 27 to 38 % of names have no character masked, and social media is
worse (lower-case handles, lone surnames). Treat it as a way to reduce exposure, not as anonymisation.
- Latin script only. Names in Greek, Cyrillic, Arabic and other scripts are not found.
- The most common false masks are brand, product and place names. Add the ones you know to `never`.
- Titles such as Mr or Dr are not part of the span.
- Only `PERSON` has been measured on public data. The other six labels have no published evaluation yet.
## Citation
```bibtex
@software{veil_2026,
title = {Veil: a 1 MB on-device model for masking personal data},
author = {Kubre, Vaibhav},
organization = {Tars Technologies},
year = {2026},
version = {0.3.1},
url = {https://github.com/Tars-Technologies/veil}
}
```
|