Mova's own corpus: silver UD for English

Mova's corpus holds random English sentences, annotated in a loop:

1. a large model annotates the sentence;
2. the Rust annotator en annotates the same tokens;
3. the large model compares the two annotations and gives the final one;
4. Rust gates check the result.

Where it lives

The corpus is kept outside the repo, in the data directory. This file only describes it; older versions of the corpus are in the git history.

• The DuckDB annotation database db/mova.duckdb holds every layer and source over the same sentences: treebank gold, silver, bronze, the large-model and Rust runs, and the tales. Mova annotation database (DuckDB) describes the schema and gives example queries; the tool is tools/db.
• The CoNLL-U files corpus/mova-silver-en.conllu and corpus/mova-bronze-en.conllu: annot silver appends to them, db load loads them into the database, and db export writes them back byte for byte.
• The school (2026-10-15): corpus/school/school-silver.conllu holds the teacher's trees, 44 011 sentences. Next to it, school-agreed.conllu is the agreed layer: the teacher plus Stanza, Trankit and spaCy-trf, combined by the written rules of tools/treeagree, whose README has the description and measurements.

Silver

mova-silver-en.conllu holds the sentences annotated by the annot silver loop above. The Rust gates are:

• closed lists of UPOS, PTB tags, FEATS and relations;
• the UD 2.18 registry for English (en/data/ud-registry-en.tsv);
• one root, no cycles, forms equal to the tokens.

Format: CoNLL-U, UD 2.18, EWT conventions. The columns are ID, FORM, LEMMA, UPOS, XPOS (PTB), FEATS, HEAD and DEPREL with subtypes; DEPS and MISC are «_». Every sentence has:

• # source: the source and licence of the text;
• # annotation: who annotated it, at what effort, and the run folder;
• # text and # sent_id.

State on 2026-09-26: 348 silver sentences.

• 20 sentences from Tatoeba and from papers, annotated in two passes at maximum effort.
• 328 examples from the seeds, raised from bronze in edit mode at high effort: 171 sentences of 5–30 tokens (49.5 k output tokens) and 157 shorter ones (36 k). Long, complex sentences are kept along with the short ones; nothing is filtered by length.

annot — the Mova annotator subproject explains, with its calibration, why edit mode was chosen.

Bronze

mova-bronze-en.conllu holds sentences annotated by the Rust annotator en alone, without the large model (annot bronze). The gates are the same as for silver.

Bronze is a queue: a large-model pass can raise its sentences to silver, and they also test the rules of the expert system (An expert system for English grammar: seeds and rules). On 2026-09-26 it held 326 sentences from the examples in the seeds en/seeds/; the gate stopped two more.

Sources of the sentences

The author and title of each paper are in the library's card for it.

The corpus licence is CC BY 4.0. The annotation is ours; the texts are CC0 or CC BY, with attribution in # source.

Checks

• A third opinion. UDPipe 2 (the EWT model) is used only for comparison in the run reports. UDPipe models are under a non-commercial licence, so their output never goes into the corpus or into training.
• Quality. Calibration on EWT sentences that have a gold standard: annot calib, with reports in the data directory under annot/runs/. Only independent gold counts as a reference, never our own final annotation.

Licence

The corpus itself is CC BY 4.0, as stated above: the texts are CC0 or CC BY, and the annotation is ours. Mova's code and models are under the Elamur Cosmic License (Elamur Cosmic License).

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