Hugging Face's logo

Datasets:
ciscoriordan
/
grapho-data

Languages:
Greek
Size:
100K<n<1M
Tags:
greek
modern-greek
monotonic
word-list
spell-checking
n-gram
License:
Dataset card Files
xet
 Community
The Dataset Viewer has been disabled on this dataset.

Grapho keyboard data

The monotonic Modern Greek word list and next-word model that Grapho, a Greek keyboard for iOS, ships. These are the exact files of data version 1.3.18-234585a: word list version 1.3.18 and next-word model version 1.3.18, built by Dilemma at Dilemma commit 234585a.

Grapho bundles the word list as published here, compiled at build time into a binary lookup table with the same entries, and the model compressed with LZMA; here the model is published unpacked, byte for byte the file the app installs. manifest.json lists every data file with its size and SHA-256.

Versions

Each data version is a tag of this dataset, named from the files' .version data. The app links the tag of the data it ships, here 1.3.18-234585a. The main branch holds the latest version. Older tags are never moved or deleted. To download one version:

from huggingface_hub import snapshot_download
snapshot_download("ciscoriordan/grapho-data", repo_type="dataset", revision="1.3.18-234585a")

Files

File Contents Size
word-list/el_monotonic.dic The word list: 637,221 spellings, each with a frequency rating 20.2 MB
word-list/el_monotonic.aff The Hunspell-style affix header (no affix rules) 360 bytes
word-list/el_monotonic.version The word list's provenance: version, Dilemma commit, counts, and the SHA-256 of every input and output 1.8 kB
word-list/el_monotonic_abbreviations.txt 220 abbreviations with their periods, one per line 2.2 kB
word-list/el_monotonic_acronyms.txt 1,734 all-capital acronyms, one per line 15.7 kB
next-word-model/el_ngram.bin The next-word model, a trigram model in Dilemma's GNLM version 2 binary format: 155,791-word vocabulary, 153,141 bigram and 1,269,045 trigram contexts 61.3 MB
next-word-model/el_ngram.version The model's provenance (JSON): version, Dilemma commit, training settings and counts per source 2.7 kB
attribution.json The sources below, as Grapho's Sources screen reads them 6.8 kB
manifest.json Every file above with its size and SHA-256

License

These files are released under the Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0) license. They are adaptations of the sources listed below, several of which are under CC BY-SA. Material under CC BY-SA 3.0 is adapted under CC BY-SA 4.0, which version 3.0 permits for adaptations. Each source keeps its own license. When you reuse these files, credit the sources below, link the license, say what you changed, and share your adaptations under CC BY-SA 4.0.

To credit this dataset: Grapho keyboard data, version 1.3.18-234585a, built with Dilemma from the sources below, https://huggingface.co/datasets/ciscoriordan/grapho-data, licensed under CC BY-SA 4.0.

The files are provided as they are, without warranties of any kind.

Sources

Built with

  • Dilemma. Open Greek. License: MIT License. Used for the word list and the next-word model.

Words and their forms

  • Wiktionary. English Wiktionary and Greek Wiktionary (Wikimedia community). License: CC BY-SA 3.0. Used for the word list.

How often words are used

  • FrequencyWords. Hermit Dave, Modern Greek full word list (el_full.txt, 1.49M forms), built from the OpenSubtitles2018 corpus of OPUS. License: CC BY-SA 4.0. Used for the word list.

Greek text

Dilemma's NOTICE gives the full credit for each source.

How the files differ from the sources

Dilemma, an open-source project under the MIT License, built these files at Dilemma commit 234585a. They hold words and counts computed from the sources, not the sources' text: no sentence, definition or example of any source is included.

  • The word list holds the inflected forms of Wiktionary's Modern Greek entries (from the inflection tables, as extracted by kaikki.org), with function words, names, elided forms and words attested in the corpora, in monotonic spelling and Unicode NFC. Each form is rated by how often it occurs in the FrequencyWords counts (word counts of the OpenSubtitles2018 subtitles) and in the corpora; no subtitle text was read. Dilemma describes how the list is built.
  • The next-word model holds counts of one-, two- and three-word sequences, computed from the normalized, tokenized sentences of the corpora's training split (221,240,505 tokens), pruned to the most frequent continuations of each context and stored as log probabilities. Dilemma describes how the model is built.
  • The abbreviation and acronym lists come from the same build of the word list.

Formats

  • el_monotonic.dic is UTF-8 text. The first line is the number of entries; then one entry per line, sorted by UTF-8 bytes: the form, a tab, and fr:C, fr:M, fr:R or fr:X, the frequency rating. C is a form with at least 37 occurrences per million words, M at least 3.7, R a rarer attested form, and X a form no frequency source has, which also carries mg:generated. After the rating, separated by spaces, a line may carry spelling notes: ac:FORM, a commoner spelling of the same letters with another accent; sp:FORM, a preferred spelling with other letters; and us:LABEL, a usage label in lowercase words joined by underscores (rare, dialectal, archaic_spelling). A note does not make the marked form wrong. Lines with each note: 82 ac:, 1 sp:, 2 us:. There are no affix flags. Names keep their capital, a final sigma is ς, and elision uses U+2019.
  • el_monotonic.aff is the matching Hunspell affix file header: SET UTF-8, LANG el, FLAG num and a TRY line, with no affix rules.
  • The .version files are provenance records: key: value lines for the word list, JSON for the model.
  • el_ngram.bin is little-endian binary, made to be memory-mapped: a 128-byte header, a sorted UTF-8 vocabulary, a unigram count per word, and three tables of the most likely next words (overall, after one word and after two words of context), each entry a 32-bit word id and a 16-bit fixed-point natural-log probability (divide by 1024). A table of counts for words outside the vocabulary follows, found through a 16-byte footer at the end of the file. Dilemma documents the format, and the docstring of its exporter gives the full layout.
Downloads last month
90
Total file size:
81.5 MB