| # Swipe Keyboard trained models, 2026-10-23 |
|
|
| Models trained and released by Destin Moss via youcoded.ai, 2026. |
| SPDX-FileCopyrightText: 2026 Destin Moss <https://youcoded.ai> |
|
|
| Four neural next-word models (English, Spanish, French, German) and one swipe-recognition model. They sit beside the |
| language files published in release `language-files-2026-10-22-open-words` and are checked by SHA-256 against |
| `packaging/assets.lock.json` in github.com/itsdestin/swipe-keyboard before use. A missing or damaged file only turns |
| that model off (suggestions fall back to the n-gram files; swipes fall back to the geometric matcher); it never |
| disables a language. |
|
|
| | File | What it is | Declared licence | |
| |---|---|---| |
| | `LANG.nwp.bin` (en, es, fr, de) | next-word model: a small recurrent network (one LSTM layer, integer weights, about 4.7 MB), used together with `LANG.ngram` | ODC-By-1.0 AND CC-BY-2.0-FR | |
| | `swipe-model.bin` | swipe recognition: small convolution + recurrent network (about 0.5 million weights, 2.1 MB) shared by all four languages | MIT | |
|
|
| ## How they were built |
|
|
| Scripts in github.com/itsdestin/swipe-keyboard (commit `14f9722` and its ancestors on branch `feature/179-engine-in-keyboard`): |
|
|
| - Next-word models: `scripts/lm/nwp/prepare.py` (tokenise, vocabulary), `train.py` (train), `tune.py` (blend with the |
| n-gram), `export.py` (write `LANG.nwp.bin`). Training text: the same FineWeb (English, sample-10BT) / FineWeb-2 |
| (Spanish, French, German) text and Tatoeba sentences as the n-gram files, with test sentences held out (Tatoeba ids |
| ending in 0, and sentences sharing an 8-word run with the test sets). |
| - Swipe model: `scripts/swipe/prep_data.py`, `assemble_futo.py`, `synth_swipes.py`, `train.py`. Training data: real |
| swipes from FUTO's swipe.futo.org dataset (MIT) plus synthetic swipes generated from our open word lists |
| (our own noise model, no FUTO code or model weights). Sessions used for validation or final scoring were held out. |
|
|
| ## Sources and licences |
|
|
| - **Web text** (next-word models) - FineWeb and FineWeb-2 by Hugging Face, Open Data Commons Attribution License v1.0 |
| (`ODC-By-1.0.txt`); built from Common Crawl, whose terms of use also apply. The models contain learned weights, never text. |
| - **Short everyday sentences** (next-word models) - Tatoeba (tatoeba.org), CC BY 2.0 FR (`TATOEBA-CC-BY-2.0-FR.txt`). |
| The contributors are credited in `en-CONTRIBUTORS.txt`, `es-CONTRIBUTORS.txt`, `fr-CONTRIBUTORS.txt`, |
| `de-CONTRIBUTORS.txt` in release `language-files-2026-10-22-open-words`. |
| - **Real swipes** (swipe model) - the swipe.futo.org dataset, https://huggingface.co/datasets/futo-org/swipe.futo.org, |
| MIT licence, reproduced below and in `FUTO-SWIPE-DATA-LICENSE.txt`: |
|
|
| MIT License |
| |
| Copyright (c) 2025 FUTO |
| |
| Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated |
| documentation files (the "Software"), to deal in the Software without restriction, including without limitation |
| the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and |
| to permit persons to whom the Software is furnished to do so, subject to the following conditions: |
| |
| The above copyright notice and this permission notice shall be included in all copies or substantial portions |
| of the Software. |
| |
| THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO |
| THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF |
| CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS |
| IN THE SOFTWARE. |
| |
| - **Synthetic swipes** (swipe model) - generated by us from the open word lists above; no third-party terms. |
|
|
| Changes made: text was split into sentences and reduced to letter words before training; swipes were resampled and |
| normalised to key positions; the trained weights were converted to the compact formats the keyboard reads. |
| These models contain no FUTO model weights and no FUTO code. https://github.com/itsdestin/swipe-keyboard |
|
|
