v1.4.0 + nabu-data v1.0.0 — the loop closes

1 August 2026 · Nabu news

Two releases, cut together, because they are one story.

nabu-data v1.0.0 is the library’s publishing arm made real: twelve datasets — Sanskrit form-to-lemma from DCS gold, the Tibetan Wylie fold and verb paradigms, a segmented Classical Tibetan slice with its boundary F1 published in-band, Greek metrical scansions anchored to Perseus CTS, the Han and kyūjitai orthography folds, the Aozora and Kanripo gaiji censuses, the Sabellic loanword table, the cuneiform value-to-sign table flattened from the Oracc Sign List, and — the release’s centerpiece — the first re-publication: stable anchors for ACTib’s segmented eKangyur, 461,301 rows tying every Derge Kangyur passage to its ACTib line by citation and content fingerprint, with the measured match census (99.51% letter-exact) in-band and the 2,245 divergent lines published side-by-side as a proofreading worksheet. Every dataset is plain CSV with a Frictionless Data Package manifest naming the exact upstream version, the producing code version, and a re-runnable recipe. CC BY 4.0, three share-alike carve-outs stated per dataset.

Nabu v1.4.0 is everything since v1.3.0 — six phases in six days:

The Elephantine and Tibetan phases (their own posts below): the island’s thirteen-language documentary record, and the Tibetan desk whole — the Derge Kangyur and Tengyur, the Old Tibetan editions, the gold treebanks and lexica, folio-paired parallel reading.
The signs desk: nabu signs resolves ATF transliteration — a catalog urn or raw pasted text — token by token through the Oracc Sign List into sign identities: value → sign name → Unicode codepoints, with one honest status per token and every ambiguity listed rather than guessed. Built for Edubba, the scribal-school sister project, whose reading panels consume the frozen --json contract; the twelfth MCP tool serves the same payload.
The consumers: the two-way loop closed in public. The published segmentation dataset now trains the library’s own segmenter — show --segmented renders Tibetan word-split, search --words keeps only word-boundary-aligned hits — and the published verb table feeds dictionary lookup, so ཕྱིན finds འགྲོ (suppletion only the table knows). Three of the twelve datasets are consumed back from the published files; the other nine are one-way by design.
The granted sources: two shelves that exist because coordinators answered letters. The complete secular Galician-Portuguese lyric — 1,676 cantigas at verse grain, under Projeto Littera’s written any-use grant — and the Ras Shamra Tablet Inventory: 5,075 inventory cards with the KTU/CTA concordances the field actually cites, published editions at line grain, under U. Chicago CORPUS/OCHRE’s NC-SA grant, with the project’s credit rendered on every surface. Sefaria’s classical midrash (the ten Rabbah collections and the Midrash Halakhah shelf) joined the same week.

The census at the tags: 974,972 documents / 68,381,456 passages in 125 language codes across 108 sources, 1.4 million dictionary entries, twenty-three research desks — and, for the first time, a data repository going the other way.