Edition Fingerprint v0.1: edition-variant loci as a retrospective provenance probe for LLM training data

Edition Fingerprint v0.1: edition-variant loci as a retrospective provenance probe for LLM training data. Correction notice (4 November 2026): Subsequent review identified extra spaces in insertion/deletion comparisons and errors in the reporting of correlation subsets in the pilot analysis. The archived v0.1 analysis requires correction and should not be used to estimate edition mixtures or establish that a particular source file was used in training. Byungwoong Yoo, Independent Researcher. ORCID: 0009-0002-1797-3100. This dataset contains edition-variant loci, corpus n-gram counts, per-locus model scores and analysis code for Chinese canonical texts, Frankenstein and On the Origin of Species. The pilot compares model preferences at variant loci with reading counts in public pretraining corpora. The archive includes four English text pairs and exploratory Chinese comparisons. Correlation analyses include loci with a nonzero count for at least one reading, including cases where only one reading occurs. Associations with reading counts do not by themselves identify document-level edition proportions. The supplied zenodo_v0.1.zip is preserved byte-for-byte. Kanripo-derived text excerpts retain CC BY-SA 4.0; the dataset is distributed under CC BY-SA 4.0, and the included code is identified as MIT. Source attribution: Kanseki Repository (Kanripo). The license displayed on this Zenodo record clarifies the blanket CC BY 4.0 wording in the bundled README and metadata.