serialhex's picture
Upload README.md with huggingface_hub
401b754 verified
|
Raw History Blame Contribute Delete
2.84 kB
metadata
license: cc0-1.0
language:
  - en
tags:
  - theology
  - christianity
  - reasoning
  - instruction-tuning
pretty_name: Open Biblical Dataset - GT (Grug-Think)

Open Biblical Dataset - GT (Grug-Think)

GT = Grug-Think. (Also, yes, like the Nissan/Ford trim and Dragon Ball GT — no apology.)

This is a rewritten-reasoning variant of serialhex/Open-Biblical-Dataset. Same 11,105 examples, same questions, same final answers, same sources — the only thing that changed is the think field.

Why this exists

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.

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).

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.

Process

Think traces were rewritten by a mix of agents against the target/ceiling/filler-ban rules above. Every one of the 11,105 records passes validation: 10-200 words, no banned filler phrases, same reasoning moves and conclusion as the original.

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.

Structure, sources, and licensing

Identical to the base 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.

Known limitations

  • Reasoning-trace rewriting was done by smaller/faster models without independent verification of semantic fidelity to the original trace — see Process above.
  • Records were validated for form (length, filler phrases) but reviewed only by spot-check, not exhaustively, for content drift.