Lynote/stealthhumanizer-corpus-style-model

StealthHumanizer Corpus Style Model (Lynote)

A lightweight, fully open writing-style statistics model trained by Lynote from a corpus of 10,000 Q1 open-access papers (OpenAlex, ranked by cited_by_count). It describes the statistical fingerprint of high-quality academic human writing and is used by StealthHumanizer as the target-style reference when rewriting AI-flavored prose.

This is a JSON statistics model (no neural weights): small (~14 KB), deterministic, and trivially auditable.

Contents

• global — corpus-level style statistics: average sentence length, burstiness (sentence-rhythm variance), vocabulary diversity, contraction frequency, passive-voice ratio, first-person pronoun rate, hedging frequency, transition-word frequency, AI-phrase frequency, sentence starters, punctuation profile.
• byDomain — the same profile split across 11 disciplines (Physical Sciences, Medicine, Computer Science, Social Sciences, Education, Mathematics, Life Sciences, Engineering, Economics & Business, Environmental Science, Multidisciplinary).
• sentenceLengthDistribution / burstinessProfile / vocabularyDiversityRange — fine-grained distributions used by the humanizer to shape output rhythm.
• humanWritingExamples — 50 excerpted examples of human academic prose.

Training / lineage

• Corpus: Q1 OA papers via OpenAlex (cited_by_count-ranked, with DOI).
• Pipeline: scripts/papers/batch-download-and-train.mjs in StealthHumanizer, run daily by scripts/ops/daily-train.sh with a quality gate (accuracy/f1/auroc ≥ 0.55/0.55/0.6 on the held-out benchmark).
• Lineage: data/models/lineage/experiments.jsonl in the same repo.

License

MIT. Part of the StealthHumanizer project (fendouai/StealthHumanizer, MIT).