sinew / README.md
LucasGalhardoLima's picture
Update dataset card
8929686 verified
|
Raw
History Blame Contribute Delete
6.61 kB
---
pretty_name: "Sinew — sourced Bible connection-graph"
license: cc-by-4.0
language:
- en
tags:
- bible
- cross-references
- knowledge-graph
- versification
- openbible
size_categories:
- 100K<n<1M
configs:
- config_name: connections
data_files: parquet/connections.parquet
default: true
- config_name: verses
data_files: parquet/verses.parquet
- config_name: texts
data_files: parquet/texts.parquet
- config_name: books
data_files: parquet/books.parquet
- config_name: versification_map
data_files: parquet/versification_map.parquet
- config_name: dataset_meta
data_files: parquet/dataset_meta.parquet
---
# Sinew — a *sourced* Bible connection-graph
**An open database of the Bible and its internal connections** — text + canonical IDs +
versification reconciliation + cross-references in one queryable artifact, **every connection
carrying its provenance**.
> *"…the sinews and the flesh came up upon them… and they lived." — Ezekiel 37:8,10*
> Sinew is the connective tissue of the canon — the layer that links verse↔verse *with provenance*,
> so others build on reconnected Scripture instead of feeding a PDF to an AI.
🧑‍💻 **Code, build & MCP server:** <https://github.com/LucasGalhardoLima/sinew>
🔭 **Live explorer (navigate by meaning, or search a theme):** <https://huggingface.co/spaces/LucasGalhardoLima/sinew-explorer>
[![Sinew explorer demo](https://raw.githubusercontent.com/LucasGalhardoLima/sinew/main/docs/demo.gif)](https://huggingface.co/spaces/LucasGalhardoLima/sinew-explorer)
## Why
Anyone building on the Bible (study apps, sermon tools, AI agents, research) redoes the same
foundational work — parsing text, canonical IDs, versification reconciliation, cross-reference
lookup — and worse, **derives connections from text similarity, which is itself an interpretive
assumption** (similarity ≠ connection). Sinew separates **fact from interpretation**: every
connection is `type + source + weight`, the dataset *attributes, never asserts*, and nothing is
silently dropped. v1 is entirely **public-domain text + CC-BY data** → frictionlessly redistributable.
## The three tiers (never blurred)
- **Tier 1 — Facts:** verse text, canonical IDs, book metadata, versification map.
- **Tier 2 — Sourced connections:** every edge has `type + source + weight (+ review_status)`.
- **Tier 3 — Derived (optional):** computed embeddings/similarity, *"computed, not authoritative"*
powers the live explorer's meaning-terrain **and its in-browser theme search**; built opt-in, never
joined to the fact tables (and not part of this published dataset, which is Tier-1/2 only).
## What's in v1
| | |
|---|---|
| **Text** | World English Bible (WEB), public domain — 31,095 verses / 1,189 chapters / 66 books |
| **Canonical IDs** | `Book.Chapter.Verse` (e.g. `John.3.16`), **eng/KJV** numbering base |
| **Versification** | `eng` (identity) + `org` (Hebrew) per verse, via the Copenhagen Alliance spec |
| **Connections** | OpenBible cross-references → **613,998** Tier-2 edges (`type=cross_reference`, `weight=votes`) |
| **Formats** | `sinew.sqlite` (all tables + joins) **and** per-table Parquet (this dataset's viewer) |
## Files
| Path | What |
|---|---|
| `sinew.sqlite` | The whole dataset in one SQLite file — all tables, ready for joins |
| `parquet/*.parquet` | One file per table (powers the dataset viewer above) |
| `sources.lock.json` | Exact upstream retrieval URLs + sha256 pins (reproducible build) |
## Load it
**Parquet, via 🤗 `datasets`** (each table is a named config):
```python
from datasets import load_dataset
edges = load_dataset("LucasGalhardoLima/sinew", "connections", split="train")
verses = load_dataset("LucasGalhardoLima/sinew", "verses", split="train")
texts = load_dataset("LucasGalhardoLima/sinew", "texts", split="train")
```
**SQLite, for joins** (one file, all tables):
```python
from huggingface_hub import hf_hub_download
import sqlite3
db = hf_hub_download("LucasGalhardoLima/sinew", "sinew.sqlite", repo_type="dataset")
con = sqlite3.connect(db)
# John 3:16 and its strongest sourced cross-references (with provenance)
con.execute("SELECT text FROM texts WHERE verse_id='John.3.16'").fetchone()
con.execute("""SELECT target_verse_id, source, weight FROM connections
WHERE source_verse_id='John.3.16' AND review_status='ok'
ORDER BY weight DESC LIMIT 5""").fetchall()
# Hebrew (org) numbering for an eng verse
con.execute("SELECT scheme_ref FROM versification_map WHERE verse_id='Joel.2.32' AND scheme='org'").fetchone()
# -> ('Joel.3.5',)
```
## Schema (tables)
- **`verses`**`verse_id (PK)`, `book`, `chapter`, `verse`, `canonical_order`, `tier=1` *(stable, translation-independent address)*
- **`books`**`book (PK)`, `name`, `testament`, `book_number`, `chapter_count`
- **`texts`**`verse_id`, `translation`, `text`, `tier=1` *(PK `verse_id+translation`; v1: WEB only)*
- **`versification_map`** — `verse_id`, `scheme ∈ {eng,org}`, `scheme_ref`, `status ∈ {present,merged,split}`, `tier=1`
- **`connections`** — `source_verse_id`, `target_verse_id`, `type`, `source`, `weight`, `confidence`, `review_status`, `turpie_class (NULL in v1)`, `tier=2`
- **`dataset_meta`**`key`, `value`
**Rules:** no `connections` row without `source`; endpoints reference the canonical **address**
(`verses.verse_id`), never a `texts` row; an edge whose endpoint doesn't resolve (≈311, versification
gaps) is **flagged in `review_status`, never dropped** — the "no silent failure" guarantee.
## Sources & license
| Source | Provides | License |
|---|---|---|
| [World English Bible](https://ebible.org/web/) (via getbible.net v2) | PD English text | Public Domain |
| [Copenhagen Alliance versification](https://github.com/Copenhagen-Alliance/versification-specification) | eng/org mappings | CC-BY / open |
| [OpenBible.info cross-references](https://www.openbible.info/labs/cross-references/) | ~340k weighted edges | **CC-BY** |
**This compilation is licensed CC-BY-4.0** and **requires attribution to OpenBible.info**; the
underlying Bible text is public domain. Exact retrieval URLs + sha256 pins are in `sources.lock.json`
and the `dataset_meta` table.
**Roadmap (P1):** typed NT→OT quotations (Turpie 1868, classified A–E); deeper Tier-3 semantic search
beyond the explorer (a query/MCP-layer semantic endpoint — the explorer's in-browser theme search
already ships); original-language (Macula) lemma/morphology; more PD translations + multilingual alignment.