NOTICE

CleverKeys
Copyright (C) tribixbite and contributors
Licensed under the GNU General Public License v3.0 (see LICENSE).

This product includes third-party data and components:

================================================================================
Word frequencies — overview
================================================================================
CleverKeys' word lists come from three CC BY-SA sources, selected per list:

  List                                              Source
  ------------------------------------------------  -----------------------------
  bundled dictionaries (src/main/assets/            wordfreq (+ AOSP LatinIME
    dictionaries/) and langpack-<lang>.zip for        inclusion oracle where noted
    en, es, fr, de, it, nl, pt, sv, ru, el, tr, he,   below)
    uk, bg, mk, id, ms, tl; langpack-en-wordfreq.zip
  langpack-en-opensubtitles.zip,                    FrequencyWords (OpenSubtitles)
    langpack-en-opensubtitles-50k.zip
  langpack-sw.zip                                   Swwiki (Swahili Wikipedia)

Every list is modified: words were selected, filtered and ranked, and their
frequencies quantised into CleverKeys' binary dictionary format ("converted to
a CleverKeys frequency list"). Each langpack-*.zip carries its own NOTICE.txt
and "license" / "attribution" / "source" manifest keys, so the attribution
travels with a pack downloaded on its own (scripts/build_langpack.py;
docs/audit/2026-09-26-data-licensing-audit.md).

CC BY-SA 4.0 adaptations may be licensed under GPLv3
(https://creativecommons.org/compatible-licenses/; conditions at
https://wiki.creativecommons.org/wiki/ShareAlike_compatibility:_GPLv3 — keep
the attribution, licence notice and link, indicate changes). CleverKeys
distributes these adapted lists under GPL-3.0-only, with the attribution below.

================================================================================
Word frequencies — wordfreq
================================================================================
  wordfreq — Robyn Speer et al.
  https://github.com/rspeer/wordfreq
  Code: Apache License 2.0 (no wordfreq code is distributed with CleverKeys)
  Data: Creative Commons Attribution-ShareAlike 4.0 (CC BY-SA 4.0)
  https://creativecommons.org/licenses/by-sa/4.0/

wordfreq's frequency data combines, among others:

  * SUBTLEX word lists (SUBTLEX-US, SUBTLEX-UK, SUBTLEX-CH, SUBTLEX-DE,
    SUBTLEX-NL), created by Marc Brysbaert et al. and freely available at
    http://crr.ugent.be/programs-data/subtitle-frequencies. SUBTLEX is freely
    available data; its authors are credited here as its terms require.
  * OPUS OpenSubtitles 2018 (http://opus.nlpl.eu/OpenSubtitles.php), whose data
    originates from OpenSubtitles.org (http://www.opensubtitles.org/) — see the
    OpenSubtitles section below.
  * Google Books Ngram Viewer data. Acknowledgement requested by Google:
    source Google Books Ngram Viewer, http://books.google.com/ngrams
  * Wikipedia (http://www.wikipedia.org), the Leeds Internet Corpus
    (University of Leeds Centre for Translation Studies) and ParaCrawl
    (https://paracrawl.eu), under Creative Commons licences.

================================================================================
Word frequencies — FrequencyWords / OpenSubtitles
================================================================================
langpack-en-opensubtitles.zip (25k) and langpack-en-opensubtitles-50k.zip are
built from the English lists of FrequencyWords (2018 release):

  FrequencyWords — Hermit Dave
  https://github.com/hermitdave/FrequencyWords
  License: "MIT License for code. CC-by-sa-4.0 for content." The word lists
  are content: CC BY-SA 4.0, https://creativecommons.org/licenses/by-sa/4.0/

FrequencyWords is computed from the OPUS OpenSubtitles corpus, whose data
originates from OpenSubtitles.org (http://www.opensubtitles.org/):

  P. Lison and J. Tiedemann (2016). OpenSubtitles2016: Extracting Large
  Parallel Corpora from Movie and TV Subtitles. In Proceedings of the 10th
  International Conference on Language Resources and Evaluation (LREC 2016).
  OPUS: http://opus.nlpl.eu/OpenSubtitles.php

================================================================================
Word frequencies — Swwiki (Swahili)
================================================================================
langpack-sw.zip is built from Swwiki, a word-frequency list made from the text
of Swahili Wikipedia (https://sw.wikipedia.org/, Wikipedia contributors):

  Swwiki — Kevin Donnelly
  https://kevindonnelly.org.uk/swahili/swwiki/
  License: Creative Commons Attribution-ShareAlike 3.0 Unported (CC BY-SA 3.0)
  https://creativecommons.org/licenses/by-sa/3.0/

Licence chain: CC BY-SA 3.0 is not itself GPL-compatible, but its §4(b) lets an
adaptation be distributed under a later version of the licence with the same
licence elements — CC BY-SA 4.0 — and CC BY-SA 4.0 adaptations may be licensed
under GPLv3 (https://creativecommons.org/compatible-licenses/). The Swahili
list is therefore adapted under CC BY-SA 4.0 and distributed under
GPL-3.0-only, with this attribution preserved.

================================================================================
AOSP LatinIME word lists
================================================================================
The dictionary build pipeline (scripts/build_wordlist.py) uses the Android
Open Source Project LatinIME word lists as build-time inclusion oracles for
en, es, fr, de, it, nl, pt (pt_BR + pt_PT), sv, ru, el, tr and he (published
upstream as "iw"); per-snapshot provenance is recorded in
scripts/data/PROVENANCE.md. The oracle only decides which words are kept; no
AOSP frequencies are distributed. (The prebuilt langpack-en.zip and
langpack-en-wordfreq.zip predate the oracle and do not use it.)

  AOSP LatinIME dictionaries — The Android Open Source Project
  https://android.googlesource.com/platform/packages/inputmethods/LatinIME/
  License: Apache License 2.0, https://www.apache.org/licenses/LICENSE-2.0

Apache-2.0 is one-way compatible with GPLv3; the derived word selections
distributed with CleverKeys are redistributed under GPL-3.0, with attribution
preserved here.

================================================================================
FUTO swipe gesture corpus (swipe.futo.org)
================================================================================
The bundled CTC swipe-emission encoder
(src/main/assets/models/ctc_swipe_encoder.onnx) was trained from scratch by the
CleverKeys project on gesture data derived from the FUTO swipe corpus. No FUTO
model weights and no FUTO model outputs were used at any point in training; the
bundled weights are original CleverKeys work distributed under GPL-3.0 with the
app. (The FUTO Model Weights License therefore does not apply to this asset;
the decode algorithms in src/main/kotlin/.../swipe/ctc/ are a clean-room port of
FUTO's GPL-3.0 swipe-library, license-compatible with this project.)

  swipe.futo.org dataset — FUTO
  https://huggingface.co/datasets/futo-org/swipe.futo.org
  License: MIT

================================================================================
How-We-Swipe shape-writing dataset
================================================================================
Training of the CTC swipe-emission encoder additionally used the How-We-Swipe
shape-writing dataset.

  How-We-Swipe — Luis A. Leiva, Sunjun Kim, Wenzhe Cui, Xiaojun Bi,
  Antti Oulasvirta (OSF node sj67f)
  https://osf.io/sj67f/
  License: MIT (Copyright (c) 2021 the authors above, as declared on the OSF node)

MIT is GPLv3-compatible; the trained weights distributed with CleverKeys are
redistributed under GPL-3.0, with attribution preserved here.

================================================================================
Shipped CTC script models — per-model provenance
================================================================================
CleverKeys ships one CTC swipe-emission encoder per script. Every one of them is
original CleverKeys work, trained from scratch, distributed under GPL-3.0 with
the app. This list is exhaustive; add a line here whenever a script model is
added.

Two delivery routes, both of them "shipped":

  * The Latin encoder is an APK asset (src/main/assets/models/).
  * Since 2026-09-10 the six per-script encoders travel INSIDE their language
    packs (scripts/dictionaries/langpack-<code>.zip, member model.onnx) rather
    than in the APK. Every one of those six languages needs its pack to decode
    at all, so the model reaches exactly the users who can use it — and nobody
    else carries 3.1 MB they cannot. The licence position is unchanged: the
    bytes are the same GPL-3.0 CleverKeys artifacts, distributed with the app's
    own language packs from the app's own repository.

  src/main/assets/models/ctc_swipe_encoder.onnx — Latin (en/fr/de/es/…).
    Trained on gesture data derived from the FUTO swipe corpus (MIT) and the
    How-We-Swipe shape-writing dataset (MIT); see the two sections above.

  langpack-ru.zip!model.onnx = ru_synth_v3_ch80_fp16w.onnx — Cyrillic (ru).
    Trained on 100% SYNTHESIZED gesture data: a learned trace generator fitted
    only to the same two MIT corpora above (FUTO + How-We-Swipe) emits Russian
    traces from a Russian word list and the on-screen layout geometry. No real
    Cyrillic gesture row appears anywhere in its pipeline — not in training, not
    in the validation split used for checkpoint selection. The Russian corpus
    used to *report* this model's accuracy (Yandex Cup 2023) carries no license
    grant, was therefore treated as research/held-out evaluation only, and
    contributed no bytes to any shipped artifact; the generator track that
    touched it is permanently unshippable. See
    docs/specs/ctc-architecture-and-multiscript-guide.md §3.3 and the licence
    seal note for the full record.

  langpack-el.zip!model.onnx = el_synth_v3_ch80_fp16w.onnx — Greek (el).
    Trained on 100% SYNTHESIZED gesture data from the same learned trace
    generator and MIT-only FUTO + How-We-Swipe source corpora described above,
    using a Greek word list and the shipped Greek layout geometry. No real Greek
    gesture row appears in training or checkpoint selection. Greek has no
    real-swipe evaluation corpus at any evidence tier; synthesis-holdout levels
    are not reported as accuracy.

  langpack-uk.zip!model.onnx = uk_synth_v3_ch80_fp16w.onnx — Cyrillic (uk).
    Trained on 100% SYNTHESIZED gesture data from the same learned trace
    generator and MIT-only FUTO + How-We-Swipe source corpora described above,
    using a Ukrainian word list and the shipped Ukrainian layout geometry. No
    real Ukrainian gesture row appears in training or checkpoint selection.
    Ukrainian has no real-swipe evaluation corpus at any evidence tier;
    synthesis-holdout levels are not reported as accuracy.

  langpack-bg.zip!model.onnx = bg_synth_v3_ch80_fp16w.onnx — Cyrillic (bg).
    Trained on 100% SYNTHESIZED gesture data from the same learned trace
    generator and MIT-only FUTO + How-We-Swipe source corpora described above,
    using a Bulgarian word list and the shipped Bulgarian layout geometry. No
    real Bulgarian gesture row appears in training or checkpoint selection.
    Bulgarian has no real-swipe evaluation corpus at any evidence tier;
    synthesis-holdout levels are not reported as accuracy.

  langpack-mk.zip!model.onnx = mk_synth_v3_ch80_fp16w.onnx — Cyrillic (mk).
    Trained on 100% SYNTHESIZED gesture data from the same learned trace
    generator and MIT-only FUTO + How-We-Swipe source corpora described above,
    using a Macedonian word list and the shipped Macedonian layout geometry. No
    real Macedonian gesture row appears in training or checkpoint selection.
    Macedonian has no real-swipe evaluation corpus at any evidence tier;
    synthesis-holdout levels are not reported as accuracy.

  langpack-he.zip!model.onnx = he_synth_v3_ch80_fp16w.onnx — Hebrew (he).
    Trained on 100% SYNTHESIZED gesture data from the same learned trace
    generator and MIT-only FUTO + How-We-Swipe source corpora described above,
    using a Hebrew word list and the shipped Hebrew layout geometry. No real
    Hebrew gesture row appears in training or checkpoint selection. Hebrew has
    no real-swipe evaluation corpus at any evidence tier; synthesis-holdout
    levels are not reported as accuracy.

================================================================================
Static context language model — Leipzig Corpora Collection
================================================================================
The bundled context models (src/main/assets/lm/<lang>.cklm, built by
scripts/build_static_lm.py) are derived from word-pair counts over these
Leipzig Corpora Collection downloads (web and news text):

  English     (lm/en.cklm)  eng-com_web-public_2018_300K
  German      (lm/de.cklm)  deu-de_web_2021_300K
  French      (lm/fr.cklm)  fra-fr_web_2013_300K
  Italian     (lm/it.cklm)  ita-it_web-public_2019_300K
  Portuguese  (lm/pt.cklm)  por-pt_web_2015_300K (weight 0.5) +
                            por-br_newscrawl_2011_300K (weight 0.5)
  Swedish     (lm/sv.cklm)  swe_news_2023_300K

Only pruned, quantised word-pair statistics are distributed; no sentence text
is shipped.

  Leipzig Corpora Collection — Universität Leipzig
  https://wortschatz.uni-leipzig.de/en/download
  Terms: https://wortschatz.uni-leipzig.de/en/usage
  License: Creative Commons Attribution 4.0 (CC BY 4.0)

  D. Goldhahn, T. Eckart & U. Quasthoff: Building Large Monolingual
  Dictionaries at the Leipzig Corpora Collection: From 100 to 200 Languages.
  In: Proceedings of the 8th International Language Resources and Evaluation
  Conference (LREC'12), 2012.

CC BY 4.0 is compatible with GPLv3 redistribution; attribution is preserved
here. Source hashes are pinned in scripts/data/PROVENANCE.md.

================================================================================
Static context language model — Tatoeba
================================================================================
The same models additionally count sentences in their language from the
Tatoeba collaborative sentence database, weighted against the Leipzig counts
(English 0.1, German 0.25, French 0.05, Italian 0.1, Swedish 2.0; the
Portuguese model counts no Tatoeba sentence; the per-snapshot weight and hash
are recorded in src/main/assets/lm/<lang>.json).

  Tatoeba — https://tatoeba.org
  License: Creative Commons Attribution 2.0 France (CC BY 2.0 FR)
  https://creativecommons.org/licenses/by/2.0/fr/

The contributors whose sentences were counted are listed, per language, in
files generated with each model:

  scripts/data/tatoeba-contributors-en.txt
  scripts/data/tatoeba-contributors-de.txt
  scripts/data/tatoeba-contributors-fr.txt
  scripts/data/tatoeba-contributors-it.txt
  scripts/data/tatoeba-contributors-sv.txt