README.md

metadata

language:
  - en
license: other
license_name: pd-seeds-deterministic-rejected
license_link: https://www.gutenberg.org/policy/permission.html
task_categories:
  - text-generation
tags:
  - dpo
  - orpo
  - preference
  - prose
  - writing-style
  - single-axis
size_categories:
  - n<1K

Kadenz-DPO

German "Kadenz": cadence.

596 pairs teaching one lesson — vary your sentence length.

chosen is unedited human prose from US-public-domain novels. rejected is the same passage with its rhythm regularised: long periods split at coordinators and semicolons, short adjacent sentences fused, every sentence pushed toward the mean length. Nothing else changes. No model touched either side.

field	content
prompt	a continuation instruction plus the first 60 words of the passage
chosen	the passage's actual continuation, human-written, untouched
rejected	the same continuation, rhythm-flattened
delta	how much sent_len_cv dropped from chosen to rejected
edits	number of split/fuse operations applied

Why this axis

Measured against a human reference distribution built from schneewolflabs/fundgrube (1,483 passages, 8 style features), every instruct model tested writes with markedly less sentence-length variation than human novelists — one of the two largest deviations from human prose, alongside repetitive sentence openers:

	sent_len_cv	z vs human
human (fundgrube)	0.74	—
Qwen3.6-27B (instruct)	0.55	−1.1
Qwen3.6-27B-TIES	0.46	−1.7
Fable-Fusion-711-27B	0.51	−1.4

Humans mix four-word sentences with forty-word ones. These models write metronomically. The degradation applied here (sent_len_cv −0.209 mean) is deliberately sized to span that gap, so the preference signal covers the distance between human prose and real model output rather than a subtler distinction the model would never encounter.

Construction

Deterministic, no LLM in the loop — the same method as schneewolflabs/hecke-dpo, which injects hedges into the same corpus.

1. 320-word passages drawn from fundgrube chapters.
2. Passages with >25% dialogue sentences are skipped — quoted speech has its own rhythm and re-bounding across quotation marks produces ungrammatical splices.
3. Long sentences (>1.6× the passage mean) are split at ; and coordinating conjunctions; adjacent short sentences (<0.5× mean) are fused with ", and".
4. Content preservation is enforced, not assumed. Every pair is checked so that rejected contains the same words as chosen modulo connectives (≤2% word-multiset difference). An earlier version dropped short fragments during splitting, which truncated clauses — that would have taught "prefer the complete text" rather than "prefer varied rhythm". Pairs failing the check, or containing punctuation artifacts, are discarded.
5. Pairs where the transform failed to degrade sent_len_cv by >0.02 are discarded.

Verification stats over the released set:

	value
pairs	596
sent_len_cv degradation	mean 0.209, median 0.182, max 1.108
split/fuse edits per passage	4.5
words per side	260 (identical between chosen and rejected by construction)

Example

chosen

…sought for her there had been wakened that wonderful vision to which alone are wonderful things revealed; the mere shapes and patterns of things becoming, as it were, refined, and gaining a kind of symbolical value…

rejected

…for her there had been wakened that wonderful vision to which alone are wonderful things revealed. The mere shapes and patterns of things becoming, as it were, refined. …

Limitations

• Archaic chosen side. The source is pre-1930 public-domain fiction, so training on it nudges diction toward period style as a side effect of teaching rhythm. Stacking this with other fundgrube-derived sets (e.g. hecke-dpo) compounds that drift. Mixing in modern permissively-licensed prose on the chosen side would isolate the axis more cleanly.
• Some rejected sentences are fragments. Splitting at a coordinator occasionally yields a clause without a finite verb. That is a second, unintended defect riding along with the intended one; a POS-aware splitter would avoid it.
• Narration only. Dialogue-heavy passages are excluded, so the axis is untrained for dialogue rhythm.
• Not decontaminated against any benchmark.

Related

• schneewolflabs/fundgrube — the human source corpus
• schneewolflabs/hecke-dpo — same method, hedging axis
• Auftakt-DPO (planned) — the companion axis: sentence-opener monotony