File size: 1,696 Bytes
a35b223 6d86249 a35b223 6d86249 55f28f8 6d86249 7411d48 6d86249 7411d48 6d86249 55f28f8 6d86249 | 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 | ---
title: README
emoji: 🐦
colorFrom: purple
colorTo: indigo
sdk: static
pinned: false
---
**Proofbird** — a writing assistant for Windows that checks spelling and grammar as you type, in the apps you already use.
The grammar models run **on your own computer**: your text doesn't leave the machine for a grammar check. They're small,
fast and open — published here under open licences (Apache-2.0; Spanish CC BY 4.0, matching its base model), with their
training data, measured results and known limits on each card.
| Model | Language | Version | Gold set | False alarms on everyday text |
|---|---|---|---|---|
| [gector-bert-en](https://huggingface.co/proofbird/gector-bert-en) | English | 1.0.0 | 258 fixed · 0 made worse | 0.05% |
| [gector-bert-it](https://huggingface.co/proofbird/gector-bert-it) | Italian | 1.0.0 | 512 fixed · 0 made worse | 0.10% |
| [gector-bert-es](https://huggingface.co/proofbird/gector-bert-es) | Spanish | 1.1.0 | 426 fixed · 0 made worse | 0.50% |
| [gector-bert-fr](https://huggingface.co/proofbird/gector-bert-fr) | French | 1.0.0 | 472 fixed · 0 made worse | 0.43% |
German follows. Every model card has a quick start: one Python file, or one click to run it in Google Colab.
**How they're made:** synthetic errors over Wikipedia and Tatoeba sentences — no learner essays, no non-commercial training
data, no user text. Precision first: a missed error is acceptable, a wrong correction is not.
Built by Patrick Andrew Hartnett as personal work.
[proofbird.app](https://proofbird.app) · [LinkedIn](https://www.linkedin.com/company/145373519) ·
[Bluesky](https://bsky.app/profile/proofbird.bsky.social) · [X](https://x.com/proofbirdapp)
|