etherfuse commited on
Commit
7b1bfcf
·
verified ·
1 Parent(s): b193a55

Card brought up to date for the finished study: what was found, where to read it, how to use this

Browse files
Files changed (1) hide show
  1. README.md +37 -1
README.md CHANGED
@@ -4,7 +4,7 @@ language: en
4
  tags: [authorship, embeddings, book-of-mormon-authorship, 1780-1830]
5
  ---
6
 
7
- # Style encoder (BOM-30)
8
 
9
  Authorship style embeddings (Amendment 24): the time-locked model fine-tuned with a contrastive
10
  objective so that passages by the same author land close together. `297m/style.pt` and
@@ -13,6 +13,42 @@ in the GitHub repository; `load_encoder(path, cfg, "base")` or `"mid"`), with th
13
  logs. Benchmark: 297M 0.839 right author, 0.500 false matches (`benchmark_297m.json`); 590M 0.839,
14
  0.628 (`benchmark_590m.json`).
15
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
16
  ## About
17
 
18
  Part of an open, pre-registered study of the authorship, borrowing and dating of the
 
4
  tags: [authorship, embeddings, book-of-mormon-authorship, 1780-1830]
5
  ---
6
 
7
+ # Style models (the "fingerprint" models)
8
 
9
  Authorship style embeddings (Amendment 24): the time-locked model fine-tuned with a contrastive
10
  objective so that passages by the same author land close together. `297m/style.pt` and
 
13
  logs. Benchmark: 297M 0.839 right author, 0.500 false matches (`benchmark_297m.json`); 590M 0.839,
14
  0.628 (`benchmark_590m.json`).
15
 
16
+ ## The second model (`v2/`): the one behind the library search
17
+
18
+ `v2/style.pt` is the second style model, on the 297M backbone. It is the model that searched all
19
+ 23,872 authors of the period for the writing closest to the Book of Mormon. The first model
20
+ memorised its training authors; this one was trained on eight times as many passages per book,
21
+ with half its steps forcing it to match an author across different kinds of writing
22
+ (`style_embed/v2.py`). Load it the same way: `load_encoder("v2/style.pt", cfg)`. A book is scored
23
+ against an author by its closest episode (`sweep/episodes.py`). `v2/log.jsonl` and `v2/run.json`
24
+ are its training record. SHA-256 of `v2/style.pt`:
25
+ `8a4e09b0ba9cef5de5beb73c58c7412a1041d60473ae52746094bf0dbff9fd29`.
26
+
27
+ How it did, on books whose author is known (results files in the GitHub repository):
28
+
29
+ - **Finding the author in a library of 18,288:** on 727 known books, the true author comes first
30
+ 34% of the time and in the top ten 59%. For a book unlike its author's other writing: 17% and
31
+ 38% (`data/sweep/coverage_episodes_v2.json`).
32
+ - **Confirming or rejecting one named author** (200 books, 2,200 pairs,
33
+ `data/sweep/nobody.json`): separation 0.955, against 0.895 for LambdaG (2024) and 0.861 for
34
+ LUAR (2021) on the same pairs.
35
+ - **Its limits:** set to confirm 95% of true authors, it still accepts 33% of wrong ones. It was
36
+ trained and tested on English of 1780-1830 only.
37
+
38
+ ## The study and what it found
39
+
40
+ This is one of the models from a finished, open study of the English text of the Book of Mormon:
41
+ who writes like it, what it borrows, and when its language dates from.
42
+
43
+ - **The findings in plain words, with live charts:** https://lifehug.com/book-of-mormon
44
+ - **The paper, the code and every results file:** https://github.com/lifehug/book-of-mormon-authorship
45
+
46
+ In short: the book copies almost nothing from earlier print. Among the eight people usually
47
+ proposed as its author, the closest writing is Joseph Smith's revelations, and a search of 23,872
48
+ authors of the period puts those revelations first or second. No tool we found, ours included, can
49
+ tell a person from the voice they are writing in, or reliably say "it is none of these people".
50
+ The study takes no position on inspiration.
51
+
52
  ## About
53
 
54
  Part of an open, pre-registered study of the authorship, borrowing and dating of the