Datasets:
You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This dataset is distributed under CC BY 4.0. Access is gated for provenance and attribution stewardship; please review the attribution note in the dataset card.
Log in or Sign Up to review the conditions and access this dataset content.
NuBerea/discourse-graph
One discourse graph over the canonical Bible. Each row of the two edge tables is an edge between
two atomic claims, classifying their rhetorical relation under an 8-class schema (supports /
qualifies / refutes / precondition / elaborates / sequence / contradicts / none)
with directionality and confidence, plus a fine relation_subtype, the verse-opening
connective, a per-edge model-tier trail, an additive ambiguity layer, and a textual-variant
sidecar. The third table groups claims that stand in coordinate parallelism (ordered narrative
runs and unordered lists) instead of representing them as pairwise edges.
Claims are the nodes (they live in the estate's claim-extraction layer); this repo ships the
edges and groups only. Every edge row keeps its grain (match_grain) and its OSIS verse anchors
on both sides (source_ref_a / source_ref_b, pericope_id_a / pericope_id_b), so the graph
can be re-cut at verse or pericope level without loss.
Configs
Canonical names follow the {corpus}_{pole}_{grain} convention; the legacy names remain
dual-listed and keep loading. Both names of a pair read the same data folder.
| Config (canonical) | Legacy alias | Rows | Columns | Description |
|---|---|---|---|---|
nt_contention_binding |
nt_edges |
78,000 | 47 | NT claim-to-claim discourse edges |
ot_contention_binding |
ot_edges |
155,150 | 47 | OT claim-to-claim discourse edges |
bible_contention_groups |
coordinate_groups |
8,261 | 9 | Coordinate-parallelism groups: span (ordered narrative run) / set (unordered list); members are claim nodes; edges reference a group via attaches_to_group |
All configs are licensed cc-by-4.0; the config list above matches the live file listing on the
repo exactly.
Edge table columns (47)
- Identity:
edge_id,safe_id_a/b,claim_index_a/b,source_ref_a/b,pericope_id_a/b - Relation spine:
relation,directionality(a_to_b|b_to_a|symmetric),confidence,reasoning,relation_subtype,subtype_source,connective - Edge meta:
cross_pericope,via_intertext_target,match_grain,model_id,build_date,concept_relation - Arbitration and tier trail:
arbitration_status,arbitration_reason,arbitration_burn,qwen_relation,haiku_relation,sonnet_relation,opus_relation - Ambiguity layer:
relation_alt,ambiguity_kind,candidate_relations - Variant sidecar:
variant_anchor,variant_relation,variant_directionality,variant_confidence,variant_relation_alt,variant_ambiguity_kind,variant_reason,variant_provenance - Coordinate-group and bridge attachment:
attaches_to_group,attach_grain,bridge_id_a/b,bridge_profile_a/b,bridge_default_rank
Labels and their review status
The relation labels are frozen model output that went through a staged quality-correction ladder
(triage, adjudication, arbitration; per-edge tier trail in the columns above). The negative
classes (refutes, contradicts) received a dedicated same-referent review: every surviving
contradicts edge and roughly three quarters of the surviving refutes edges carry an
adjudicated or arbitrated read above the primary labeller; the remainder rest on the primary
label plus a secondary-model check. Treat refutes rows whose arbitration_status is
confirmed or contradicts_demoted as the least-reviewed cell. Coarse label error on the
worst-measured cells was brought under 10% against an arbitrated reference; within-pericope
qualifies remains the hardest relation.
Attribution
| Source | License |
|---|---|
| Westminster Leningrad Codex (Hebrew OT alignment substrate, via MACULA Hebrew) | Public Domain |
| SBLGNT (Greek NT alignment substrate, via MACULA Greek) | CC BY 4.0 |
| NuBerea claim extraction and relation labelling (frozen model output) | CC BY 4.0 |
NuBerea project. Licensed CC BY 4.0.
What this is not
This is a machine-labelled relation layer, not a scholarly commentary and not ground truth. The validation figures are relative to a model-arbitrated reference, not to an independent gold standard. No source text edition is redistributed here; the tables carry claim identifiers, verse anchors, labels, scores, provenance codes, and short model reasoning strings only.
- Downloads last month
- 119