The language, machine-readable.

Everything below is generated from the language repository when this site is built, so it cannot drift from what the pages say. All of it is CC BY-SA 4.0: use it, keep the licence, say where it came from.

• /corpus.tsvAmadunia–English sentence pairs · 104 KB
• /corpus.jsonlThe same pairs, one JSON object per line · 139 KB
• /dictionary.jsonEvery root — 300 of them · 51 KB
• /dictionary.csvThe same roots as a spreadsheet · 26 KB
• /index-english.jsonEnglish into Amadunia — 322 entries · 30 KB
• /llms-full.txtThe whole language as one document · 585 KB

/corpus.tsv

1321 rows, three columns: the Amadunia, the English, and the file each came from.

/corpus.jsonl

For translation and language-model pipelines, which read JSONL rather than TSV.

/dictionary.json

Word, meaning, thematic group, and the languages each root came from.

/dictionary.csv

Imports straight into a flashcard deck. Generated upstream and checked against the Markdown on every push.

/index-english.json

The index that runs the other way to the dictionary, for writing rather than reading. What the translate page looks words up in.

/llms-full.txt

Every grammar page, lesson, text and the dictionary, concatenated. For a model or a person who wants all of it at once.

What it is, and what it is not.

Every pair comes from a lesson, a grammar page or an original text, where an Amadunia sentence stands beside its English. Nothing was translated to make the corpus; the translations were already there, because the material teaches by glossing.

It is small, and honest about that. 1321 pairs will not train a translation system on their own — a language pair normally needs hundreds of thousands. What it is good for is a language that has no data at all: a seed, a test set, or the evidence that a constructed language has been used rather than only specified. It grows with every lesson and every text, and it has never contained a sentence that breaks a rule.

How to cite.