Style models (the "fingerprint" models)

Authorship style embeddings (Amendment 24): the time-locked model fine-tuned with a contrastive objective so that passages by the same author land close together. 297m/style.pt and 590m/style.pt are the encoders on the 297M and 590M backbones (style_embed.train.StyleEncoder in the GitHub repository; load_encoder(path, cfg, "base") or "mid"), with their training logs. Benchmark: 297M 0.839 right author, 0.500 false matches (benchmark_297m.json); 590M 0.839, 0.628 (benchmark_590m.json).

The second model (v2/): the one behind the library search

v2/style.pt is the second style model, on the 297M backbone. It is the model that searched all 23,872 authors of the period for the writing closest to the Book of Mormon. The first model memorised its training authors; this one was trained on eight times as many passages per book, with half its steps forcing it to match an author across different kinds of writing (style_embed/v2.py). Load it the same way: load_encoder("v2/style.pt", cfg). A book is scored against an author by its closest episode (sweep/episodes.py). v2/log.jsonl and v2/run.json are its training record. SHA-256 of v2/style.pt: 8a4e09b0ba9cef5de5beb73c58c7412a1041d60473ae52746094bf0dbff9fd29.

How it did, on books whose author is known (results files in the GitHub repository):

  • Finding the author in a library of 18,288: on 727 known books, the true author comes first 34% of the time and in the top ten 59%. For a book unlike its author's other writing: 17% and 38% (data/sweep/coverage_episodes_v2.json).
  • Confirming or rejecting one named author (200 books, 2,200 pairs, data/sweep/nobody.json): separation 0.955, against 0.895 for LambdaG (2024) and 0.861 for LUAR (2021) on the same pairs.
  • Its limits: set to confirm 95% of true authors, it still accepts 33% of wrong ones. It was trained and tested on English of 1780-1830 only.

The study and what it found

This is one of the models from a finished, open study of the English text of the Book of Mormon: who writes like it, what it borrows, and when its language dates from.

In short: the book copies almost nothing from earlier print. Among the eight people usually proposed as its author, the closest writing is Joseph Smith's revelations, and a search of 23,872 authors of the period puts those revelations first or second. No tool we found, ours included, can tell a person from the voice they are writing in, or reliably say "it is none of these people". The study takes no position on inspiration.

About

Part of an open, pre-registered study of the authorship, borrowing and dating of the earliest Book of Mormon text (https://github.com/lifehug/book-of-mormon-authorship; pre-registration https://osf.io/6ugxc/). The project does not argue for or against divine inspiration. Everything is published so it can be checked and redone; the corpus texts we agreed to keep private are not, and the repository's docs/replication.md gives the steps to obtain the same access.

Licence: CC BY-NC 4.0. Training data: English printed 1780-1830 (Internet Archive, Harvard Institutional Books 1.0, HathiTrust public-domain volumes, Text Creation Partnership), with the Book of Mormon and everything after 1830 removed; see the base model cards (etherfuse/timelocked-1830-125m). Contains derivatives of Harvard Institutional Books 1.0, Institutional Data Initiative, Harvard Law School Library; used for noncommercial research. Checksums: every file's SHA-256 is in the GitHub repository, data/manifests/hf_<repo>.jsonl.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support