bcv-commons/hebrew-clause-layout

Hebrew Bible Clause Layout

The Hebrew Bible, with its Aramaic portions, laid out one clause segment per line, chapter by chapter. Each line carries its indentation in the text: which clause it depends on, across verses, and how far it is indented under it. The layout is built from the MACULA Hebrew syntax trees over the unfoldingWord Hebrew Bible (UHB) text. It is meant for translators, and for the tools and agents that help them, to read a chapter clause by clause, with each subordinate clause, speech and law case indented under the clause it belongs to.

It is a share-alike alternative to clause layouts derived from the ETCBC BHSA (CC BY-NC 4.0). It contains no data under a non-commercial licence.

Release	v0.37 (rules 0.37), built 2026-10-21
Segmentation	8977039d6ddb5383: changes only when a segment boundary moves
Levels	0.37+2862201eca6c2125: the rules version and the hash of the settings the levels were computed with
Size	39 books, 929 chapters, 23,213 verses, 89,940 clause segments

Since v0.35 (levels only; the segmentation and every book's sha256 are unchanged): every speech is a level, from the Hebrew Bible Speaker Layer, with Moses as Deuteronomy's frame (see Speeches); MACULA's tree sets the level for more kinds of attachment inside a verse. Not yet scored against BHSA (see How the text-level hierarchy was checked).

Files

hebrew-clause-layout-v0.37.tar.gz   the pinned download: everything below except data/, in one archive
clauses/<BOOK>/<BOOK>_<CCC>.jsonl     one JSON object per clause segment, one file per chapter (in the archive)
layout/<BOOK>/<BOOK>_<CCC>.txt        the same chapter as indented plain text (in the archive)
data/clauses-<BOOK>.parquet           the same rows, one file per book, for the dataset viewer and datasets.load_dataset
_clauses_source.json                  the version stamp (see "Versions")
build.json                            the build record: sources, versions, relation sets, settings
clause_layout.toml                    the rules and settings the layout was built with, each rule documented with examples
schema/clause_segment.schema.json     JSON Schema of a row
RELEASE_NOTES.md                      what is in this release, the rules and their reach, checksums
SHA256SUMS                            SHA-256 of the archive and the data files

• <BOOK> is the USFM book code (GEN, RUT, PSA, …), and <CCC> the zero-padded chapter (001).
• To pin this release, download https://huggingface.co/datasets/bcv-commons/hebrew-clause-layout/resolve/v0.37/hebrew-clause-layout-v0.37.tar.gz and check it against its SHA-256 in SHA256SUMS.

What a clause segment is

A clause segment is a maximal run of whole words, inside one verse, that belongs to one clause.

• A clause interrupted by an embedded clause gives two segments, one before and one after the embedded clause, as BHSA's clause atoms do.
• A clause that runs over a verse boundary gives one segment per verse.
• A word is assigned as a whole, including its prefixes (וְ, בְּ, לְ, הַ, …).

The segments are MACULA's clauses, cut at word and verse boundaries. Every row has the MACULA word keys of its words, so a row can be joined to anything keyed on MACULA words.

The levels

MACULA has no clause-to-clause relations above the verse: a MACULA sentence is one verse, and the main clauses in a verse are siblings. The text-level hierarchy is therefore derived by rules from MACULA's features: verb forms, the clause's first words, verbs of saying, subordinators, person, and the MACULA tree. Every clause gets a discourse_parent and a relation.

• text_depth: the indentation shown in layout/. It counts the links up the discourse_parent chain that add a level: a speech, a subordinate clause, an embedded clause. Chain links (continuation, coordination, asyndetic, parallel, …; the set is build.json relations.flat) stay on the same level. 0 is the main line. Each chapter starts at the margin, unless a speech goes on over the chapter break (see speeches below). Conditionals are laid out as BHSA lays them out: a protasis before its apodosis stands one level in, the apodosis stands level with it, and what goes on from the apodosis stays there; a later case of a series steps back and in again. (Before 0.35 the apodosis came back to the line; that layout is the profile indented_protasis.)
• text_depth_law: the same, with the conventions of BHSA's law on top: the later cases of a law step in one level each, and a new law stands where the law before it stands. It matters in the law books and casuistic prophecy.
• tree_depth: every link counted except parallel, chain links included, so each clause of a chain nests under the one before it, as in BHSA's tree.
• depth: the number of clause ancestors in the MACULA tree, within the verse.

Speeches (since 0.36; rule speech_levels). Every quotation of the Hebrew Bible Speaker Layer (release pinned in clause_layout.toml and build.json) is a level, in text_depth and text_depth_law. A speech stands one level in from the line that introduces it, and while it lasts none of its lines goes further out than that: a line may go deeper and come back to the speech's level, never above it. A speech inside a speech gets a level of its own one further in. When a speech ends, the next line goes back to the level of the speech (or the narrative) around it. A speech that goes on over a chapter break keeps its level, so a chapter inside a speech does not start at the margin (Lev 7:1 goes on God's message to Aaron of Lev 6:18). Where the speaker layer starts or ends a speech, the layout follows it, its mistakes included. One exception (since 0.37; [speech] frames): in Deuteronomy, Moses' speeches are the book's frame, as the psalmist is a psalm's and the prophet a prophetic book's, so they add no level; the main line of his speeches stands at the margin and the quotations inside them (God's words) count from there. In bsb-add-books' scoring 39% of the lines of Moses' speeches stand at or out of their inquit in BHSA, against 11% in other books; Job's speeches are levels (94% stay in from the inquit). Because the speech levels are applied last, text_depth is no longer the count of links up the discourse_parent chain inside a speech; discourse_parent, relation and tree_depth are as the rules give them.

The tree keeps a protasis under its apodosis. A client that wants BHSA's direction for conditionals in the tree inverts those links as build.json relations.conditional_inversion describes. For a flat view, use text_depth_law.

The rules (121, each with its scope and examples; a few are trials, switched off) are in clause_layout.toml, and the code in bcv-commons/bcv-query (shoresh/macula/discourse_rules.py). Settings files called profiles can change the conventions. This release is built with the default profile: chapters start at the margin, except inside a speech, which keeps its level over a chapter break. Inside a verse, MACULA's syntax tree sets the level for 17 kinds of attachment (clause_layout.toml [tree] trusted; tree_attachments.json: which MACULA attachments carry a level, figures from the comparison below): phrases, clauses under cjpCLx, a "כִּי" clause under ClCl, and since 0.36 the kinds where the tree was right at least twice as often as the rules in bsb-add-books' scoring of a trial that used the tree everywhere (a fronted noun and its verb, אוּלַי and its clause, a subject and its participle, an עַד phrase); since 0.37 not where such a kind opens with a word that lost in their scoring of those kinds ([tree] untrusted: the question "הֲלֹא", a fronted circumstantial clause, the Aramaic "דִּי"). It stays off for clauses on the verse's own group, at a speech's start and after an inquit.

layout/: plain text

RUT 1
1:0 וַיְהִ֗י בִּימֵי֙
1:1 שְׁפֹ֣ט הַשֹּׁפְטִ֔ים
1:0 וַיְהִ֥י רָעָ֖ב בָּאָ֑רֶץ
1:0 וַיֵּ֨לֶךְ אִ֜ישׁ מִבֵּ֧ית לֶ֣חֶם יְהוּדָ֗ה
1:1 לָגוּר֙ בִּשְׂדֵ֣י מוֹאָ֔ב

• The first line is <BOOK> <chapter>.
• Every other line is <verse>:<text_depth> <text>. A reader shows the number as indentation: " " * n + text.
• The text is the UHB surface text with vowels and accents. Words are separated by single spaces, maqaf ־ and paseq ׀ are kept, and sof pasuq ׃ ends the verse.
• The line shows the vocalised (qere) reading; the clauses/ row has both readings.

clauses/ and data/: one row per segment

Ruth 1:1, fifth segment:

{"book": "RUT", "chapter": 1, "verse": 1, "seg": 5, "verse_ref_en": "RUT 1:1",
 "clause_key": "080010010131", "parent_key": "080010010081", "clause": 119513, "parent": 119508, "sentence": 119498,
 "depth": 1, "discourse_parent": "080010010081", "relation": "purpose", "discourse_attach": null, "discourse_rule": "embedded",
 "text_depth": 1, "text_depth_law": 1, "tree_depth": 3, "kind": "VC", "rule": "V-PP", "language": "hbo",
 "first": "080010010131", "last": "080010010151",
 "words": ["080010010131", "080010010132", "080010010141", "080010010142", "080010010151"],
 "text": "לָגוּר֙ בִּשְׂדֵ֣י מוֹאָ֔ב", "ketiv_qere": []}

Field	Meaning
book, chapter, verse	USFM code, chapter and verse, in Hebrew numbering
seg	position of the segment in the chapter, from 1
verse_ref_en	the same verse in English numbering (see "Versification")
clause_key	stable clause id: the first word key of the clause's first segment, shared by all segments of one clause
parent_key	clause_key of the nearest enclosing MACULA clause with words of its own, or null
clause, parent, sentence	row ids of the tree build; they change when it is rebuilt, so never key on them
depth	embedding depth in the MACULA tree, within the verse
discourse_parent	clause_key of the clause this one depends on in the text-level hierarchy, also across verses; null where nothing comes before it on its line (the first clause of a chapter or a Psalm, a fresh start after a paragraph break)
relation	the kind of link to discourse_parent: continuation, coordination, asyndetic, parallel, speech, inquit, vocative, causal, purpose, temporal, conditional, relative, complement, circumstantial, consequence, adversative, embedded; null where discourse_parent is null
discourse_attach	word key of the word inside the discourse_parent clause that this clause depends on (a relative clause's antecedent), or null
discourse_rule	the rule that set discourse_parent (for information; relation is the field to use)
text_depth	the indentation (flat view), as in layout/
text_depth_law	the indentation with conditionals laid out as in BHSA's law
tree_depth	the number of links up the discourse_parent chain, except parallel ones (chain links included)
kind, rule	MACULA clause kind (VC verbal, NC nominal, …) and constituent rule (V-S-PP)
language	hbo Hebrew or arc Aramaic
first, last, words	the MACULA word keys of the segment (BBCCCVVVWWW plus a part digit), in order
text	surface text of the segment, as in layout/
ketiv_qere	{key, ketiv, qere} for the words of the segment that have a ketiv/qere (from UHB); ketiv or qere is null where only one exists

Versification

Verses use the Hebrew numbering of the source text, which differs from English Bibles in places: Gen 31:55 = Hebrew 32:1; Psalm superscriptions are verse 1; Joel 3–4; Malachi 3:19–24; and others. Every row carries verse_ref_en, the reference in English numbering, from the bcv-commons/bibles verse map org-to-eng (from STEPBible TVTMS, CC BY 4.0). Where English divides a Hebrew verse (Ps 13:6, Isa 63:19), verse_ref_en names the English verse the segment falls in.

Versions

The chapter files carry no version header. _clauses_source.json is the version stamp:

{"dataset": "bcv-commons/hebrew-clause-layout", "release": "v0.37", "segmentation": "8977039d6ddb5383", "levels": "0.37+2862201eca6c2125",
 "books": {"RUT": {"chapters": 4, "verses": 85, "segments": …, "sha256": "…"}, …}}

• segmentation changes only when a segment boundary moves anywhere. Each book's sha256 covers that book's segment boundaries (book, chapter, verse, first and last word key), so a rebuild that moves a boundary shows exactly which books changed.
• levels changes when the rules or their settings change. The relations, discourse_parent and the depths can then change, but the segments and their word keys don't.
• A levels-only release keeps segmentation and every book's sha256. Anything keyed on the segments (sidecars, alignments to clause segments) stays valid.

build.json records the sources: the MACULA Hebrew commit, the UHB tag, the TVTMS revision and the hashes of the versification files.

How the text-level hierarchy was checked

The hierarchy is derived by rules from MACULA Hebrew alone. The rules were developed and tuned with the help of a comparison with the ETCBC's BHSA (CC BY-NC 4.0), made privately by bcv-commons/bsb-add-books as a reference for evaluation only: only figures and failure cases in words were exchanged, and the choice of the attachments in tree_attachments.json rests on such figures. No BHSA values, codes, links or files are in the rules, the code or the rows of this dataset. The documentation of some rules in clause_layout.toml cites, in words, figures and examples from the comparison.

Step agreement with BHSA's own clause links, laid out the same way: the share of line pairs where the step from one line to the next (deeper, level, shallower) is the same. Rules 0.37 have not been scored yet: it was released without waiting for bsb-add-books' scoring. The figures below are for v0.35 and for a trial with 0.37's tree kinds (and three openers that 0.37 leaves out); the speech levels and the Deuteronomy frame are not in them. On the delivered test chapters, v0.35, measured by bsb-add-books (the law chapters in text_depth_law):

Set	Chapters	Line pairs	text_depth	text_depth_law	v0.22 text_depth
Tuned	8	1,081	88.6%	88.7%	88.3%
Held-out	5	648	88.4%	88.4%	88.0%
Prophets	9	944	83.6%	83.7%	83.3%
Law	26	3,329	85.6%	87.0%	81.7%
Expressions	56	6,871	86.0%	85.9%	85.2% (39 chapters)
E-chapters	13	1,217	83.2%	82.9%	83.0%
All delivered chapters	110	13,284	85.7%	86.0%	84.2% (93 chapters)

These are test sets: chapters chosen because they are hard (speeches, law, prophecy, poetry), not a sample of the whole Old Testament. A chapter in two sets counts in each set once and in "all" once.

Over the whole Old Testament (all 929 chapters, 88,634 line pairs; measured by bsb-add-books; "tree kinds" is the trial of v0.35 with the tree kinds, before the three openers were left out, which by their estimate adds about 0.1 point):

Genre	Line pairs	v0.22	v0.35	v0.35 + tree kinds
All	88,634	86.1%	86.8%	87.2%
Narrative	47,148	88.3%	88.9%	89.3%
Law	4,939	83.2%	86.4%	86.5%
Prophecy	22,102	83.5%	83.9%	84.3%
Poetry (Psalms, Lamentations, Song)	8,135	85.7%	86.4%	86.8%
Wisdom (Job, Proverbs, Ecclesiastes)	6,310	80.9%	81.5%	82.3%

On the 351 steps where the two layouts of conditionals differ, the default matches BHSA in 65%, the profile indented_protasis in 32% (over all delivered chapters: 85.7% against 85.0%).

A test set of 98 lines from bsb-add-books' study of where a line moves out of a level (verse, opening words, the expected step: to the margin, to the speech's main line, out one level, level, in): 0.37 gets 71 right (v0.35: 67). The test is part of the code's test suite.

Mean depth per chapter against BHSA's (v0.35), over the 379 chapters for which a reference was available (Leviticus and Deuteronomy, Ezekiel and four of the Twelve, and the poetic books with Lamentations): a mean absolute gap of 0.356 levels (text_depth_law 0.302). Per book, over all 39 books: 0.149. (v0.22: 0.364, 0.311, 0.208.)

Where the layout follows a convention, the convention is chosen for readers, documented in clause_layout.toml, and sometimes differs from BHSA on purpose. For example, a chapter starts at the margin. Where the layout is deeper than BHSA and the extra level reflects the text (a real subordination, a speech inside a speech), it is kept.

Known differences (see RELEASE_NOTES.md):

• The prophets are the weakest part: where an oracle or a speech opens and closes (Isa 5, Isa 43, Jer 3), and a narrative-looking verb inside an oracle (Isa 3:16, 10:14). Since 0.36 the speaker layer sets where speeches open and close, so its mistakes show here too; those found are fixed in the speaker layer (v0.11: Lev 17; Deuteronomy, where Moses' two long speeches were in pieces and some of God's words inside them were given to Moses; Jer 1, 2, 24, 25).
• Inside a speech nothing reaches the margin (Isa 43:16 and 44:6 "כֹּה אָמַר", Amos 9:11 "בַּיּוֹם הַהוּא" step out to God's words' level), where BHSA starts a new section of the oracle at the margin. Deuteronomy is not affected: Moses is its frame.
• A main clause with no opening word: neither the rules nor MACULA's tree have a reliable signal for when it steps in or out.
• The poetry is, on average, a little shallower than BHSA; Micah 2 still has a messenger speech that doesn't close.

BHSA: Eep Talstra Centre for Bible and Computer, VU Amsterdam; DOI 10.17026/dans-z6y-skyh.

Sources and licence

Part	Source	Licence
Hebrew/Aramaic text, ketiv/qere	unfoldingWord Hebrew Bible (UHB) v3.0.0, based on the Westminster Leningrad Codex	CC BY-SA 4.0
Clauses, words, clause kinds and rules	MACULA Hebrew (Clear Bible / Biblica), WLC/lowfat, commit 47db250bd55d	CC BY 4.0
English verse numbers	bcv-commons/bibles org-to-eng, from STEPBible TVTMS (rev. 902681f77a4a)	CC BY 4.0
Speech levels	Hebrew Bible Speaker Layer v0.11, bcv-commons (hebrew-speaker-layer)	CC BY 4.0
The text-level hierarchy, this compilation	bcv-commons	CC BY-SA 4.0
The rules and code (clause_layout.toml, bcv-query)	bcv-commons	MIT

Attribution when you use this dataset:

Hebrew Bible Clause Layout (v0.37), bcv-commons, CC BY-SA 4.0. Text: unfoldingWord Hebrew Bible (CC BY-SA 4.0). Syntax: MACULA Hebrew, Clear Bible / Biblica (CC BY 4.0). Versification: STEPBible TVTMS (CC BY 4.0). Speeches: Hebrew Bible Speaker Layer, bcv-commons (CC BY 4.0).

Feedback

Issues and requests: bcv-commons/bcv-query. Please name the chapter and verse, the line, and what you expected.