serialhex commited on
Commit
e36f063
·
verified ·
1 Parent(s): 6d37af8

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +43 -0
README.md ADDED
@@ -0,0 +1,43 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: cc0-1.0
3
+ language:
4
+ - en
5
+ tags:
6
+ - theology
7
+ - christianity
8
+ - reasoning
9
+ - instruction-tuning
10
+ pretty_name: Open Biblical Dataset - GT (Grug-Think)
11
+ ---
12
+
13
+ # Open Biblical Dataset - GT (Grug-Think)
14
+
15
+ **GT = Grug-Think.** (Also, yes, like the Nissan/Ford trim and Dragon Ball GT — no apology.)
16
+
17
+ This is a rewritten-reasoning variant of [serialhex/Open-Biblical-Dataset](https://huggingface.co/datasets/serialhex/Open-Biblical-Dataset). Same 11,105 examples, same questions, same final answers, same sources — the only thing that changed is the `think` field.
18
+
19
+ ## Why this exists
20
+
21
+ The base dataset's own card calls out a known limitation: reasoning traces are "essay-shaped, not terse." They narrate ("Let me consider...", "First, I'll..."), restate the question, and run long — closer to a reader-facing paragraph than a model's actual internal working notes. That shape itself was failing a structural slop check across most of the corpus.
22
+
23
+ This variant runs every `think` field through an offline rewrite pass (inspired by the `grug-think` style: short, dense, declarative reasoning — real hypothesis-and-check, not narrated prose) targeting ~70 words per think block, hard ceiling 200, with an explicit filler-phrase ban ("let me...", "first, I'll...", etc).
24
+
25
+ **What did NOT change:** the question, the final answer, the source citations, the actual reasoning *content* (same passage/word/grammar points noticed, same logical or modal check made, same conclusion reached). This is a compression/restyling pass on the reasoning trace, not a reinvention of it.
26
+
27
+ ## Process
28
+
29
+ - 11,035 of 11,105 (99.4%) rewritten automatically via the Claude Haiku API against the target/ceiling/filler-ban rules above.
30
+ - The remaining 70 failed automatic validation (over-compressed below a 10-word floor, under-compressed over a 200-word ceiling, or a filler-phrase slip) and kept their original essay-shaped think block on that pass.
31
+ - Those 70 were re-rewritten in a follow-up pass with tightened guidance, reaching 100% validation across the full corpus.
32
+ - Every record passes: 10-200 words, no banned filler phrases, same reasoning moves and conclusion as the original.
33
+
34
+ There is no automated check for reasoning *fidelity* the way there is for e.g. verifying a quote against source text — the rewrite prompt itself (keep the same reasoning moves, don't invent a new line of reasoning, don't change the conclusion) is the only guardrail here. Spot-checking is recommended if you're relying on this for anything beyond style transfer.
35
+
36
+ ## Structure, sources, and licensing
37
+
38
+ Identical to [the base dataset](https://huggingface.co/datasets/serialhex/Open-Biblical-Dataset) — same tiers (WEB passages, creeds/Didache/patristics, logic/philosophy texts), same public-domain sources, same CC0 licensing on generated content. See that dataset's card for full details.
39
+
40
+ ## Known limitations
41
+
42
+ - Reasoning-trace rewriting was done by a smaller model (Haiku) without independent verification of semantic fidelity to the original trace — see Process above.
43
+ - The 70 manually-patched records (and the automatic pass generally) were validated for *form* (length, filler phrases) but reviewed only by spot-check, not exhaustively, for content drift.