serialhex commited on
Commit
d7b48f7
·
verified ·
1 Parent(s): e365f26

Add dataset card

Browse files
Files changed (1) hide show
  1. README.md +148 -0
README.md ADDED
@@ -0,0 +1,148 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: cc0-1.0
3
+ task_categories:
4
+ - text-generation
5
+ - question-answering
6
+ language:
7
+ - en
8
+ tags:
9
+ - theology
10
+ - christianity
11
+ - bible
12
+ - pastoral-counseling
13
+ - reasoning
14
+ - synthetic
15
+ pretty_name: Open Biblical Dataset
16
+ size_categories:
17
+ - 10K<n<100K
18
+ configs:
19
+ - config_name: default
20
+ data_files:
21
+ - split: train
22
+ path: master.jsonl
23
+ ---
24
+
25
+ # Open Biblical Dataset
26
+
27
+ **11,105 examples.** Synthetic instruction-tuning data for a Christian
28
+ theology / pastoral-counseling fine-tune: a real situated question, a
29
+ reasoning trace, and a grounded pastoral answer — every answer required to
30
+ quote and stay inside the bounds of a real source text, never an isolated
31
+ verse or an appeal to unnamed authority ("many theologians have held...").
32
+
33
+ Built with [`data-forge`](https://github.com/) — a custom generation
34
+ pipeline (question-gen → answer-gen → offline lint pass), not a
35
+ template-filled or scraped dataset. Every example traces back to a real,
36
+ public-domain source passage, which is carried alongside the example as
37
+ `context` so the grounding is checkable.
38
+
39
+ ## What's in the box
40
+
41
+ | Tier | Source | Examples |
42
+ |---|---|---|
43
+ | 0 — Norming norm | WEB Bible (single-pericope) | 2,952 |
44
+ | 0 — Norming norm | WEB Bible (paired-passage) | 5,000 |
45
+ | 1 — Normed norms | Nicene/Apostles' Creeds + Didache | 216 |
46
+ | 1 — Normed norms | Patristic witnesses (Justin Martyr, Ignatius, Clement of Rome, Polycarp, Barnabas, Clement of Alexandria, Shepherd of Hermas, et al.) | 999 |
47
+ | 2 — Wisdom/illumination | Logic & philosophy (Mill, Lewis Carroll, Boethius, Whately, Watts, Port-Royal Logic, Euclid) | 1,938 |
48
+
49
+ Each tier carries a different citation discipline, matching how each source
50
+ actually functions:
51
+
52
+ - **Tier 0 (scripture)** is read at the level of the full passage, never an
53
+ isolated verse, with any genuinely contested question (grace/election,
54
+ eschatology) presented as a real, live disagreement rather than resolved
55
+ on the model's own authority.
56
+ - **Tier 1 creeds/Didache** are cited as what they are — fixed, named
57
+ communal confessions ("Nicaea, 325, art. 3") — illuminating and binding
58
+ within their own scope, never scripture's equal.
59
+ - **Tier 1 patristic witnesses** are cited as named INDIVIDUAL thinkers
60
+ ("Justin Martyr, First Apology") with their own argument — worth taking
61
+ seriously, but measured against scripture, not assumed correct merely for
62
+ being early.
63
+ - **Tier 2 wisdom/logic** sources are explicitly subordinate tools —
64
+ analogical or logical aids that sharpen reasoning but settle nothing
65
+ theological on their own authority.
66
+
67
+ Across every tier: no appeal to an unnamed authority ("the church has long
68
+ held...") is allowed — a source is either named specifically, or the point
69
+ is made in the model's own voice with no historical framing at all.
70
+
71
+ ## Format
72
+
73
+ One JSON object per line. `conversations` follows a ShareGPT-style
74
+ `system` / `human` / `gpt` turn structure; the `gpt` turn carries the
75
+ answer in `value` and its reasoning trace in `think` (not rendered inline —
76
+ pulled out as its own field so a fine-tune can map it to whatever
77
+ thinking-tag convention the target chat template uses).
78
+
79
+ ```json
80
+ {
81
+ "conversations": [
82
+ {"from": "system", "value": "<the standing commitments every answer follows>"},
83
+ {"from": "human", "value": "<a realistic, situated question>"},
84
+ {"from": "gpt", "value": "<the grounded pastoral answer>",
85
+ "think": "<the reasoning trace behind it>"}
86
+ ],
87
+ "context": "<the real source passage this example is grounded in>",
88
+ "pericope_ref": "<a human-readable reference, e.g. \"Romans 8:28-30\" or \"Justin Martyr, First Apology, Chapter IV\">",
89
+ "source": "<which source corpus this came from, e.g. \"web_bible\", \"creeds_didache\">",
90
+ "source_tier": 0,
91
+ "variation": "<which register/angle this example was generated under, if any>",
92
+ "lint_flags": {"...": "advisory-only structural/phrase flags from the offline lint pass - not yet used to filter"}
93
+ }
94
+ ```
95
+
96
+ ## How it was made
97
+
98
+ 1. **Chunk** each source into pericopes — a complete argumentative or
99
+ narrative unit (a creed clause, a Didache chapter, a Boethius Book
100
+ section, a Bible pericope), never a fixed-size window.
101
+ 2. **Question-gen**: a realistic, situated question a real person would
102
+ actually bring to the material — grief, doubt, a skeptic's challenge, a
103
+ personal failure — never "explain this passage."
104
+ 3. **Answer-gen**: single call, producing both the reasoning trace and the
105
+ final answer together, under a standing set of commitments (reading
106
+ method, divine-action framing, contested-questions handling, citation
107
+ discipline) that stays byte-identical across every tier's prompt.
108
+ 4. **Offline lint pass** (advisory in this release — see Limitations):
109
+ phrase-based and structural checks for filler, unnamed-authority
110
+ appeals, and essay-shaped reasoning traces, recorded per-example in
111
+ `lint_flags`.
112
+
113
+ Generated across a mix of model backends (Gemini/Gemma, Mistral, NVIDIA-
114
+ hosted models) routed through a graceful-degradation router — no single
115
+ model's fingerprint dominates the set.
116
+
117
+ ## Licensing
118
+
119
+ - **Source texts**: every passage behind every example is public domain —
120
+ the World English Bible, the Nicene/Apostles' Creeds, the Didache and
121
+ patristic writings (Ante-Nicene Fathers, Roberts-Donaldson translation),
122
+ and the Tier 2 logic/philosophy texts (Project Gutenberg / pre-1923
123
+ scans: Mill, Lewis Carroll, Boethius, Whately, Watts, the Port-Royal
124
+ Logic, Euclid).
125
+ - **Generated text** (questions, reasoning traces, answers): produced by
126
+ third-party model APIs (Google/Gemini, Mistral, NVIDIA-hosted models).
127
+ This dataset's own curation is released CC0; if your use case cares about
128
+ upstream model-output terms, check the relevant provider's usage policy
129
+ before redistributing at scale.
130
+
131
+ ## Limitations — stated honestly
132
+
133
+ - **Reasoning traces are essay-shaped, not terse.** The offline lint pass
134
+ that flags this (`think_slop_structural` in `lint_flags`) is advisory
135
+ only in this release — it was not used to filter or rewrite anything
136
+ here. A companion dataset applies an actual compression pass (see the
137
+ `grug-think`-inspired rewrite) to fix this; this release is the
138
+ unmodified base.
139
+ - **Some Tier 2 OCR sources carry period-typesetting noise.** Whately,
140
+ Watts, and the Port-Royal Logic are archive.org scans of 18th/19th-
141
+ century editions; mechanical cleanup was applied (boilerplate/header
142
+ stripping, hyphenation rejoining) but individual OCR character errors
143
+ (a misread "s" as "f," etc.) were not hand-corrected.
144
+ - **Register-rotated child/teen slices and a safety-escalation slice exist
145
+ separately** and are not included in this release — they're held back
146
+ pending a dedicated human review pass.
147
+ - **No held-out eval split.** This is training-shaped data as generated;
148
+ slice your own split before fine-tuning.