NuBerea Features

Pre-analytical feature extraction outputs for the NuBerea corpus: Stanza lemma/lexicon/syntax tables for patristic Greek and Latin, SPhilBerta/PhiloBERTa embeddings for patristic segments and biblical verses, and PhiloBERTa lexical alignment primitives. Consolidates configs formerly in NuBerea/patristics-features, NuBerea/concept-primitives, and NuBerea/greek-embeddings-primitives.

Configs

Patristic Greek / Latin — Syntax

Config	Rows	Key Columns
grc_syntax	78,209	segment_id, source, cts_urn, author_label, finite_verb_rate, max_dep_depth, subordinate_clause_rate
lat_syntax	145,559	segment_id, source, cts_urn, author_label, finite_verb_rate, max_dep_depth, subordinate_clause_rate

Relocated (not in this dataset): grc_lemmas, grc_lexicon, lat_lemmas, lat_lexicon were moved out of features by k-remediation (2026-06-23) and now live in the isolation repo NuBerea/quarantine-derivatives (license:other, do-not-redistribute), with a verified byte-identical copy in NuBerea/patristics-features (grc_* = 98,872 rows, lat_* = 145,559 rows). Their lemma_text / tokens reconstruct restricted catlib (PD-Migne PG/PL) source text, so they are not redistributed here. The *_syntax configs above ship only aggregate stats (no reconstructable text) and remain.

Patristic Embeddings

Config	Rows	Key Columns
patr_grc_embeddings_sp	40,176	segment_id, cts_urn, family_id, author_label, work_label, embedding (SPhilBerta 768-dim)
philoberta_alignment	20,992	lang_pair, strong_a, strong_b, lemma_a, lemma_b, angular_sim, low_freq_flag

Biblical Verse Embeddings

Config	Rows	Model	Key Columns
deut_verse_embeddings_sp	5,857	SPhilBerta	osis_ref, book_id, verse_text, embedding
lxx_verse_embeddings	29,659	Krahn shlm-grc-en	osis_ref, verse_text, lxx_substitution_score, embedding, token_count, was_truncated
lxx_verse_embeddings_sp	29,659	SPhilBerta	osis_ref, verse_text, lxx_substitution_score, embedding
nt_verse_embeddings	7,939	Krahn shlm-grc-en	osis_ref, verse_text, embedding, token_count, was_truncated
nt_verse_embeddings_sp	7,939	SPhilBerta	osis_ref, verse_text, embedding
pseu_verse_embeddings_sp	15,043	SPhilBerta	osis_ref, book_id, verse_text, language, embedding
vulgate_verse_embeddings	30,948	Krahn shlm-grc-en	osis_ref, verse_text, stratum, embedding

NHC LN Bridge

Config	Rows	Model	Key Columns
gnostic_ln_bridge	~2,675	SPhilBerta KNN-5	segment_id, codex, tractate_num, work_slug, work_title, primary_ln, vote_confidence, nn_osis_refs, nn_cosines, embedding (768-dim), model_id, build_date

LN domain codes transferred from NT verse embeddings to Nag Hammadi Corpus segments via cosine KNN (k=5) in SPhilBerta embedding space. vote_confidence = top-voted LN weight / total weight.

Rabbinic Hebrew — Morphology (union)

rabbinica_morphology — one row per word, Stanza he morphology (tokenize, mwt, pos, lemma, depparse), the additive union of mishnah_features, tosefta_features and tosefta_vilna_kodashim_tahorot_features, which stay published. Every row carries corpus, edition and k so each part's provenance and licence is declared on the row itself. Licence composition (the config's licence is the composite of its parts; the most restrictive part, CC BY-SA 4.0, governs redistribution of the whole table):

corpus	edition (text source)	rows	licence
mishnah	Mishnah ed. Romm, Vilna 1913 — NuBerea/rabbinica:mishnah	313,412	Public Domain
tosefta	Tosefta ed. Lieberman, JTS 2001 — NuBerea/rabbinica:tosefta_lieberman	273,128	CC BY 4.0
tosefta_vilna	Tosefta Kodashim + Tahorot, he.wikisource Vilna pass — NuBerea/rabbinica:tosefta_vilna_kodashim_tahorot	151,520	CC BY-SA 4.0

The two Tosefta arms are different editions and overlap in Kodashim/Tahorot; they are kept as distinct rows, never merged or deduplicated. Filter on k to take only the permissive parts. The morphology itself is machine-derived (Stanza) and carries no editorial apparatus from any edition.

Source Repos (Deprecated)

Former Repo	Configs Absorbed
NuBerea/patristics-features	grc_syntax, lat_syntax (absorbed) — grc_lemmas, grc_lexicon, lat_lemmas, lat_lexicon were relocated to NuBerea/quarantine-derivatives, not absorbed (see Configs note)
NuBerea/concept-primitives	patr_grc_embeddings_sp, philoberta_alignment
NuBerea/greek-embeddings-primitives	deut_verse_embeddings_sp, lxx_verse_embeddings, lxx_verse_embeddings_sp, nt_verse_embeddings, nt_verse_embeddings_sp, pseu_verse_embeddings_sp, vulgate_verse_embeddings

Usage

from datasets import load_dataset

# Patristic GRC syntax
ds = load_dataset("NuBerea/features", "grc_syntax")

# NT verse embeddings (SPhilBerta)
ds = load_dataset("NuBerea/features", "nt_verse_embeddings_sp")

# PhiloBERTa cross-language alignment
ds = load_dataset("NuBerea/features", "philoberta_alignment")

NuBerea project. Licensed CC BY 4.0.

Methodology — patristic SPhilBerta embeddings (patr_grc_embeddings_sp, patr_lat_embeddings_sp)

Both configs embed patristic segments with bowphs/SPhilBerta (Riemenschneider & Frank 2023, arXiv:2308.13116), a sentence-transformer trained jointly on Latin, Ancient Greek, and English. The same encoder is used for both languages by design: it places Greek and Latin segments in a single shared vector space, so cross-language concept comparison (e.g. a Cyril of Alexandria Greek passage vs. an Augustine Latin passage) is meaningful. Using a language-specific encoder for either side would break that comparability.

Config	Lang	Rows	Source configs (patristics-pta/csel/canonical/catlib split repos)	Dim
patr_grc_embeddings_sp	grc	40,176	patristic Greek segments	768
patr_lat_embeddings_sp	lat	141,137	lat_segments, catlib_lat_segments, catlib_orig_lat	768

Build (Latin, 2026-06):

1. Pull Latin patristic segments from the three source configs on the patristics-* split repos (pta/csel/canonical/catlib; patristics-texts was retired 2026-06-24).
2. Unicode-NFC normalize; keep segments with Latin-script purity ≥ 0.80 (drops language-mislabels, critical-apparatus contamination, stray non-Latin characters). 146,293 → 141,137 segments.
3. Join reconciliation metadata (cts_urn, author_label, work_label) from the canonical author/work registry by segment_id.
4. Encode with SPhilBerta (batch 64) on Apple-Silicon MPS; 768-dim sentence vectors, un-normalized.

Schema (both configs): segment_id, cts_urn, cts_group_urn, family_id, author_label, work_label, text, lang, embedding (768-d float), model_id.

Parity note: patr_lat_embeddings_sp is the Latin counterpart built under the project's Greek/Latin parity rule — every patristic feature built for one language is built for the other. Latin coverage (141k) exceeds Greek (40k) because the Latin patristic corpus (Augustine, Ambrose, Tertullian, Cyprian, Jerome, …) is larger in segment count.

Rabbinica-derived morphology configs (2026-08-30) — license envelope note

The repo license tag moved cc-by-4.0 → cc-by-sa-4.0: it is the most-restrictive ENVELOPE of the configs now present, per the estate's rabbinica precedent. True per-config provinces:

config	rows	k (true province)	source text
mishnah_features	313,412	public-domain	rabbinica/mishnah (Romm Vilna 1913)
tosefta_features	273,128	cc-by-4.0	rabbinica/tosefta_lieberman
tosefta_vilna_kodashim_tahorot_features	151,520	cc-by-sa-4.0	rabbinica/tosefta_vilna_kodashim_tahorot (he.wikisource)
targum_arc_lexicon	527,019	cc-by-sa-4.0	rabbinica six targum configs (text_arc)
targum_arc_lexicon_v2	527,019	cc-by-sa-4.0	rabbinica six targum configs (text_arc)
mishnah_ln_bridge_rabbinica	4,192	cc-by-4.0	rabbinica/mishnah_en (Kulp CC-BY) — clean LN bridge, C1 2026-09-01
yerushalmi_ln_bridge_rabbinica	12,243	cc-by-4.0	rabbinica/yerushalmi (Guggenheimer CC-BY en leg) — first clean Yerushalmi LN bridge, C1 2026-09-01

All four are original-language (Hebrew/Aramaic) surfaces; zero English. ShareAlike modification notice: SA-province source text is tokenized/annotated into row-per-word parquet; attribution and verbatim license ride the source rows in NuBerea/rabbinica. ⚠ targum_arc_lexicon reproduces the incumbent's NULL annotation surface faithfully (all rows oov — the TAHOT-keying producer defect is documented in build/features/RECEIPT_C1_2026-08-30.md); do not wire anything assuming it carries MACULA annotation. Pre-existing configs' provinces are unchanged by the tag move.