wolfvswhale/prose-cadence-stats

Prose cadence statistics

Measurements of 38 stylometric features across 5,402 documents, split by authorship (human or machine) and by register (informal, formal, multi-paragraph).

There is no text in this dataset. Every row is a set of numbers plus a stable reference to the document it was measured from. That is deliberate, and both reasons matter.

The sources carry incompatible licenses, so republishing a merged text corpus would be a mess. Measurements are facts about text rather than the text itself, which sidesteps the problem.

More usefully, the measurements are the artifact people actually want. If you want to set a detection threshold at the 2nd percentile instead of the 5th, or check whether a published threshold is defensible, or fit your own classifier on interpretable features, you can do it from this file without downloading eight gigabytes first.

What's in it

features is one row per document: source, source_ref, register, author, label, then 38 numeric columns. Thirty-two come from prose-eval, covering rule-hit rates and cadence statistics. Six are the gate statistics used by bluepencil, prefixed gate_.

distributions is the summary you would actually consult to set a threshold: mean, standard deviation, and the 1st through 99th percentiles for every feature, broken out by register and authorship.

register	human	machine
informal (forum answers)	1,200	1,200
formal (scientific abstracts)	1,200	1,200
structured (multi-paragraph articles)	602	—

Documents under 120 words were excluded, since cadence statistics on very short text are noise.

The finding that motivated publishing it

A single global threshold is wrong for most registers. Here is the 5th percentile of sentence-length variation in human writing, which is where a "this prose is too flat" gate would fire:

register	5th percentile
formal	0.207
informal	0.283
structured	0.357

That is a 72% spread across three registers of ordinary human prose. A tool calibrated on a pooled corpus, which is what bluepencil currently ships, will be too permissive on academic writing and too aggressive on encyclopedia articles. The right fix is register-conditional thresholds, and this dataset is what makes that possible to build and to check.

A second, smaller result: gate_uniform_paragraphs is measurable on only 632 of the 5,402 documents, because most public human/machine corpora store single blocks of text with no paragraph breaks at all. Any paragraph-shape statistic derived from them is close to meaningless, which is worth knowing before you trust one.

Using it

from datasets import load_dataset

feats = load_dataset("wolfvswhale/prose-cadence-stats", "features")["train"]
dist  = load_dataset("wolfvswhale/prose-cadence-stats", "distributions")["train"]

Recomputing a threshold at your own target false-positive rate:

import pandas as pd

df = feats.to_pandas()
human_formal = df[(df.author == "human") & (df.register == "formal")]
threshold = human_formal["cv_sentence_len"].quantile(0.02)   # 2% FPR, not 5%

Comparing a feature across classes:

d = dist.to_pandas()
d[(d.feature == "rule_editorializing") & (d.register == "informal")][["author", "p50", "p95"]]

Provenance and caveats

Source documents come from HC3 (CC-BY-SA-4.0), AI-and-Human-Generated-Text (MIT), and wikitext-103 (CC-BY-SA-3.0 and GFDL). source_ref identifies the originating row so any measurement can be traced and recomputed. This dataset is CC-BY-SA-4.0 to honor the strongest share-alike term among its inputs.

The machine-written documents were generated in 2022 and 2023. This is the central limitation and it applies to every public paired corpus I could find, including RAID, whose newest generators are GPT-4 and Mistral. Statistics computed here describe how models wrote three or four years ago. Em dashes and negative parallelism, the two constructions a reader in 2026 notices first, barely appear on the machine side, which says more about the corpus than about the rules.

HC3's human text also carries a detokenization artifact, spaces before punctuation, which separates the classes on roughly 73% of documents. It was normalized before measurement, but residual traces likely survive in character-level statistics. See prose-eval for the measurement of that leak.

Registers are labeled by source rather than by inspection. "Formal" means scientific abstracts, not formality assessed per document.

Citation

@misc{lyell2026prosecadence,
  title  = {Prose Cadence Statistics},
  author = {Lyell, J. Alderman},
  year   = {2026},
  url    = {https://huggingface.co/datasets/wolfvswhale/prose-cadence-stats}
}

Built with prose-eval. Applied by bluepencil.