Treebanks

Sentences annotated in Universal Dependencies (CoNLL-U: words, lemmas, tags, morphological features, dependency tree), UD 2.18, EWT conventions.

File	Sentences	Texts	Annotation
`mova-school-agreed-en.conllu` (new in 0.4)	42,759	the lessons of Mova's primary school: 26 courses written for Mova (ours)	see below
`mova-silver-en.conllu`	348	Tatoeba sentences under CC0 (`# source` says which), example sentences from Mova's seeds, and 6 sentences from ACL Anthology papers under CC BY 4.0	parsed by Mova, checked and corrected by a large language model, with UD validation gates
`mova-bronze-en.conllu`	326	example sentences from Mova's seeds	parsed by Mova only, not checked

The agreed school

The sentences of the school's lessons (# course, # lesson, # sent_id = school:<course>:<lesson>:<n>), in the 26 primary courses (the middle-school courses are not included).

How made (2026-10-15):

Mova's own parser (en) drafts each sentence; a large language model (Claude Haiku 5.5, medium effort) patches the draft under UD validation gates (# annotation says so).
Three open parsers parse the same tokens: Stanza and Trankit (UD) and spaCy trf.
Written rules (tools/treeagree in the repository) change a word only on the evidence of the parsers that agree with each other, and never past the UD gates: 82.6 k of 1.17 M words changed (7.1%), each with its rule in MISC (Agree=R7-rel-agree, the teacher's earlier choice in Haiku=). Spot checks: the strongest rule about 80% right, the others about 90%.

How Mova uses it: the training data of Mova's en model (en.bin, with UD English EWT and ESLSpok) and of its GPU parser. A held-out part (every 10th lesson) is the school test of Mova's parsing numbers.

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/corpus/open/treebanks/