Ched-ai/voynich-eva

Voynich Manuscript EVA Transcription

Dataset Summary

One row per locus (line, label, circle or radius text) of the Zandbergen-Landini (ZL) transliteration of the Voynich Manuscript (Beinecke MS 408, Yale University), in EVA (the Extensible Voynich Alphabet), version 3b of 13/05/2025. Each row keeps the source text exactly as written (text_raw) beside VCAT's readings of it, so nothing the transliterator recorded is lost: every locus of the source file is a row, and the build fails if one is not.

This dataset is part of the Voynich Computational Analysis Toolkit (VCAT).

Every layer as a configuration

The default configuration is the lines, as published since version 0.2.3; every other layer of VCAT is a configuration of this repository too (roadmap item 2.9), so one load_dataset("Ched-ai/voynich-eva", "<config>") reaches all of it. The repositories voynich-manuscript-metadata and voynich-transcription-mismatch keep their own configurations, so every name and path that loads today still loads.

Config	File	One row is
default	eva_lines.parquet	one locus of ZL (the lines)
words	eva_words.parquet	one word of ZL's lines
pages	pages.parquet	one page
folios	folios.parquet	one folio
quires	quires.parquet	one quire
bifolia	bifolia.parquet	one sheet (conjugate leaves)
page_orders	page_orders.parquet	one page in one named order
page_canvases	page_canvases.parquet	one link from a page to Yale's images
layers	layers.parquet	one interpretive layer and its attribution
mismatch	mismatch_index.parquet	one locus of ZL or IT, compared
transliteration_zl	zl.parquet	one reading of ZL
transliteration_it	it.parquet	one reading of IT
transliteration_it2a	it2a.parquet	one reading of IT2a
transliteration_cd	cd.parquet	one reading of CD
transliteration_fg	fg.parquet	one reading of FG
transliteration_gc	gc.parquet	one reading of GC
transliteration_rf1b	rf1b.parquet	one reading of RF1b
transliteration_vt	vt.parquet	one reading of VT
transliteration_lsi	lsi.parquet	one reading of LSI
locus_variants	locus_variants.parquet	one locus, every EVA reading side by side (not a consensus)
word_alignments	word_alignments.parquet	one reader's words of a locus aligned to the frame

What each configuration holds is described on its own dataset card: the metadata tables on voynich-manuscript-metadata, the mismatch index on voynich-transcription-mismatch, the per-source readings and the variant tables (not a consensus) on voynich-transliterations.

Dataset Structure

Data Instance

The first row of the build, as committed:

{
  "alternates": [
    [
      "cth",
      "oto"
    ]
  ],
  "char_count": 43,
  "comments": [
    "@254;"
  ],
  "currier_language": "A",
  "drawing_breaks": [],
  "hand": "1",
  "has_alternatives": true,
  "has_high_ascii": false,
  "has_illegible": false,
  "has_uncertain": false,
  "high_ascii": [],
  "illegible": [],
  "illustration_type": "T",
  "ligatures": [],
  "line_id": "f1r:1",
  "line_index": 1,
  "line_number": 1,
  "line_type": "paragraph",
  "locus_type": "@P0",
  "markers": [
    "<%>"
  ],
  "page_id": "f1r",
  "position": "@",
  "quire": "A",
  "section": "text_only",
  "source": "zandbergen_landini",
  "source_line": 24,
  "source_version": "bf5b6d4ac1e3",
  "text": "<%>fachys.ykal.ar.ataiin.shol.shory.[cth:oto]res.y.kor.sholdy<!@254;>",
  "text_clean": "fachys.ykal.ar.ataiin.shol.shory.cthres.y.kor.sholdy",
  "text_raw": "<%>fachys.ykal.ar.ataiin.shol.shory.[cth:oto]res.y.kor.sholdy<!@254;>",
  "uncertain_spaces": 0,
  "word_count": 10
}

Data Fields

The fields of schemas/transcription_lines.schema.json, in order (this table is rendered from the schema):

Field	Type	Description
page_id	string	IVTFF page name (e.g., f1r, f86v3, fRos)
line_number	integer	Source locus ID within page (may have gaps - preserves original numbering)
line_index	integer	Sequential index within page (always 1..n, no gaps)
line_id	string	Unique line identifier: {page_id}:{line_number}
text	string	Transcription text for this line with IVTFF markup, stripped, continuation lines joined (text_raw is the exact source text)
text_raw	string	The transliterated text exactly as in the source: after the locus identifier and the whitespace that follows it, to the end of the line (a wrapped locus keeps its continuation lines and line terminators)
locus_type	string	The IVTFF locus code as written, locator and locus type (IVTFF 2.0 Tables 8 and 9), e.g. @P0, &Lz
source_line	integer	1-based line number of the locus identifier in the source file
alternates	array of array of string	The readings of each uncertain reading [a:b] or [a:b:c], as written, most likely first
ligatures	array of string	The characters of each ligature {...}, as written
uncertain_spaces	integer	Number of uncertain word spaces (,)
illegible	array of string	Each run of unreadable characters (?, or ??? for an unknown number), as written
high_ascii	array of integer	Each high-ASCII code @NNN; in the transliterated text (not in comments)
comments	array of string	The text of each in-line comment <!...>
drawing_breaks	array of string	Each interruption by a drawing, as written (a word space)
markers	array of string	Each paragraph start <%>, paragraph end <$> and text tag <@X=y>, as written
text_clean	string	The normalization profile analysis-v1 of text_raw (vcat.profiles; the build report's text_clean_policy): ligature glyphs without braces, the first reading of each alternate, <-> and <~> as '.', '?' and the legacy EVA markers '!' '*' dropped, ',' kept, comments, paragraph marks, text tags, whitespace and wraps removed; glyphs as written
line_type	string	Type of text
position	string	IVTFF locator, the locus position relative to the previous item (IVTFF 2.0 Table 8)
quire	string or null	Quire identifier (A-T)
section	string or null	VCAT's one name for the page header's $I illustration type: a renaming of what the source states, the source's layer section_name (data_sources/layers.yaml); the metadata's section by folio ranges is VCAT's judgement
currier_language	string or null	Currier's language classification
hand	string or null	Scribe hand identifier
illustration_type	string or null	Type of illustration on page
source	string	Source identifier (e.g., zandbergen_landini)
source_version	string	SHA256 hash prefix of source file
char_count	integer	Count of EVA characters in clean text
word_count	integer	Count of words (separated by . or ,)
has_uncertain	boolean	text holds a '?' (an unreadable character) outside comments, counting only the first reading of an alternate
has_illegible	boolean	text holds a legacy EVA illegibility marker '!' or '*' outside comments; never true for an IVTFF 2.0 EVA file
has_alternatives	boolean	Contains alternative readings ([a:b])
has_high_ascii	boolean	Contains high-ASCII codes (@NNN;)

text_clean is the normalization profile analysis-v1; the other profiles (raw-v1, diplomatic-v1, no-uncertain-spaces-v1) are computed from text_raw by vcat.profiles. The rules are in docs/data_model.md.

Data Splits

One split, train, holding every row.

The word level (words config)

One row per word of the lines above (eva_words.parquet, from eva_words.jsonl), built by python -m builders.build_eva_words from the committed lines. The tokenizer is the one EVA tokenizer of the repository (vcat/words.py, which the lines' word_count also uses): a word is a run of a line's text_clean between word spaces; . is a certain space, , an uncertain one (the transliterator was not sure; it ends a word here, and space_after says uncertain, so a reader can join across it); a high-ASCII code @NNN; is part of its word. The tokenizer decides nothing about glyphs and never splits or joins what text_clean does not.

Positions: word_index, line_initial, line_final within the line; paragraph_id, line_in_paragraph, paragraph_initial, paragraph_final within the source's paragraph, which runs from the P line carrying the start mark <%> to the one carrying the end mark <$> (IVTFF 2.0 section 6.9: where a paragraph begins and ends is the transcriber's decision). Labels, circles, radii and the few free-floating lines outside a marked pair are in no paragraph. Each word carries its line's hand, currier_language, section and quire. The first row of output/eva_words.jsonl, as committed:

{
  "currier_language": "A",
  "hand": "1",
  "line_final": false,
  "line_id": "f1r:1",
  "line_in_paragraph": 1,
  "line_index": 1,
  "line_initial": true,
  "line_type": "paragraph",
  "locus_type": "@P0",
  "page_id": "f1r",
  "paragraph_final": false,
  "paragraph_id": "f1r.p1",
  "paragraph_initial": true,
  "quire": "A",
  "section": "text_only",
  "space_after": "certain",
  "word": "fachys",
  "word_id": "f1r:1.1",
  "word_index": 1
}

Field	Type	Description
word_id	string	The word's id: its line's line_id, a period and its position in the line (f1r:1.3)
line_id	string	The line (locus) it is in, as in the lines config
page_id	string	The page
line_index	integer	The line's position on its page, as in the lines config
word_index	integer	The word's position in its line, from 1
word	string	The word: a run of the line's text_clean between word spaces (vcat/words.py)
space_after	string or null	The space after it: certain (.), uncertain (,), or null for the last word of the line
line_initial	boolean	Whether it is the first word of its line
line_final	boolean	Whether it is the last word of its line
paragraph_id	string or null	The source's paragraph it is in (<page>.p<n>, the n-th on the page), or null outside every paragraph
line_in_paragraph	integer or null	Its line's position in the paragraph, from 1; null outside one
paragraph_initial	boolean	Whether it is the first word of its paragraph
paragraph_final	boolean	Whether it is the last word of its paragraph
line_type	string	The line's type, as in the lines config (paragraph, label, circle, radius)
locus_type	string	The line's IVTFF locus code, as in the lines config (@P0)
hand	string or null	The line's hand ($H), as in the lines config
currier_language	string or null	The line's Currier language ($L), as in the lines config
section	string or null	The line's section, as in the lines config
quire	string or null	The line's quire ($Q), as in the lines config

Dataset Creation

Source Data

• Source: the ZL transliteration, file ZL3b-n.txt, version 3b of 13/05/2025
• URL: https://www.voynich.nu/data/ZL3b-n.txt
• SHA-256: recorded in data_sources/sources.yaml; the release manifest says whether the build read the pinned bytes (source.pinned, checked by both validators), and python -m data_sources.verify_sources checks the cached file
• Format: IVTFF 2.0
• Transliterators: René Zandbergen and Gabriel Landini

Processing

python -m builders.build_eva_lines parses the file (parsers.IVTFFParser), emits one row per locus, accounts for every source locus and page (a locus that is neither a row nor reported fails the build), and writes JSONL, Parquet, a build report, a SHA-256 file and a release manifest. Two validators (validators/validate_phase1_outputs.py, validators/verify_invariants.py) check the rows against the source; the suite checks that the committed files are a fresh build. Repository: voynich-data.

Statistics

From the committed build report (output/eva_lines_build_report.json; this table is rendered from it):

Metric	Value
Pages	227
Lines (loci)	5,385
Source loci	5,385
Words (word_count, summed)	39,018
EVA characters (char_count, summed)	194,437
Lines holding a high-ASCII code	132
circle lines	84
label lines	1,029
paragraph lines	4,130
radius lines	142
Pages of Currier language A	114
Pages of Currier language B	83
Pages of Currier language unknown	30

The word level, from its committed build report (output/eva_words_report.json; this table is rendered from it):

Metric	Value
Words	39,018
Lines holding a word	5,385
Paragraphs (the source's <%> to <$>)	740
Words in a paragraph	35,056
Words followed by an uncertain space	2,748

Known Limitations

• One transliteration. The rows are ZL's readings. Other transliterations read many loci differently: ZL and Takahashi (IT) give the same reading on about half of the loci they share (see voynich-transcription-mismatch). VCAT does not choose between readings.
• Uncertain readings stay uncertain. Alternates ([a:b]), unreadable characters (?) and uncertain spaces (,) are kept in text_raw and listed in their fields; text_clean takes the first reading of each alternate, so it is one reading among those the source allows.
• section is VCAT's one name for the page header's $I illustration type: a renaming of what the source states (the source's layer section_name, data_sources/layers.yaml), not a field of its own in the source; the metadata's section by folio ranges is VCAT's judgement.
• Counts are of the profile's text. char_count and word_count count text_clean; another profile counts differently.
• Stable schema. From 1.0.0, field names and meanings change only as STABILITY.md allows (a major version); pin a version for reproducible work.
• The text is undeciphered. The rows record glyphs, not meaning.

Related Datasets

This dataset is part of the Voynich Computational Analysis Toolkit (VCAT):

• voynich-eva (this dataset): one row per locus of the ZL transliteration
• voynich-manuscript-metadata: page, folio and quire metadata
• voynich-transcription-mismatch: ZL against IT, locus by locus

Datasets can be joined on page_id.

Quick Start

from datasets import load_dataset

ds = load_dataset("Ched-ai/voynich-eva")
print(ds["train"][0])

# Rows of Currier language A
lang_a = ds["train"].filter(lambda x: x["currier_language"] == "A")

# The analysis text of every row
texts = [row["text_clean"] for row in ds["train"]]

# Another layer: its configuration name
words = load_dataset("Ched-ai/voynich-eva", "words")

From a checkout of the repository, vcat.load(source=, unit=, profile=) reads the same build without the network:

import vcat

lines = vcat.load()                                     # ZL's lines under analysis-v1
it_words = vcat.load(source="IT", unit="word")          # IT's words, cut by the EVA tokenizer
raw = vcat.load(source="LSI", unit="line", profile="raw-v1")

Versioning

• Current version: v1.0.0
• Schema status: stable from 1.0.0, under STABILITY.md (identifiers never reused or silently changed)
• Source version: ZL 3b of 13/05/2025

Licensing

Data: CC-BY-4.0 for the repository, which holds VCAT's annotations in some configurations (the metadata's section judgements and missing folios, the links to Yale's images, the mismatch verdicts and the variant tables: the VCAT layers of data_sources/layers.yaml). The default lines, the words and every transliteration_* configuration are the sources' text and functions of it, which voynich.nu makes available under CC0 (the statement is recorded in data_sources/sources.yaml, read at its source and quoted in docs/SOURCES_LICENSE.md): those fields stay CC0-1.0 wherever they are reused.

Code: MIT (the repository's LICENSE).

Please cite the transliterators (below) in published work. Per-source details: SOURCES_LICENSE.md.

Citation

@misc{voynich-eva,
  author = {Noah Chelednik},
  title = {Voynich Manuscript EVA Transcription},
  year = {2026},
  publisher = {Hugging Face},
  version = {1.0.0},
  doi = {10.5281/zenodo.23685586},
  url = {https://huggingface.co/datasets/Ched-ai/voynich-eva}
}

Source Citation

Every transliteration this repository holds, with its transliterators (from data_sources/sources.yaml); cite the ones you use, and the dataset above:

Source	Transliteration	Transliterators	Upstream version
ZL	Zandbergen-Landini transliteration (ZL), EVA	René Zandbergen, Gabriel Landini	Version 3b of 13/05/2025
IT	Takahashi transliteration (IT), EVA, extracted from the Landini-Stolfi interlinear file	Takeshi Takahashi	Version 1a of 27/12/2022
IT2a	Takahashi transliteration (IT), EVA, version 2a, extracted from the Landini-Stolfi interlinear file	Takeshi Takahashi	Version 2a of 02/02/2023 modified 25/06/2025
CD	Currier transliteration (CD), Currier alphabet	Prescott Currier, Mary D'Imperio	IVTFF version 2a of 02/02/2023
FG	First Study Group transliteration (FG), FSG alphabet	First Study Group (William Friedman)	IVTFF version 2a of 02/02/2023
GC	Glen Claston transliteration (GC), v101 alphabet	Glen Claston (Tim Rayhel)	IVTFF version 2a of 05/02/2023 modified 25/06/2025
RF1b	Reference transliteration (RF1b), EVA, full form	René Zandbergen	v. 1b
VT	voynichese.com transliteration (VT), EVA, mainly by Takeshi Takahashi	Takeshi Takahashi	v. 0e
LSI	Landini-Stolfi interlinear file (LSI), EVA, several transliterations per locus	Gabriel Landini, Jorge Stolfi, and the transcribers it collects	Release 1.6e6 - ?? Dec 1998

@misc{zandbergen-landini,
  author = {Zandbergen, René and Landini, Gabriel},
  title = {Voynich Manuscript Transliteration},
  url = {https://www.voynich.nu/transcr.html}
}

Contact

• Repository: voynich-data
• Issues: GitHub Issues

---

This project does not claim to solve the Voynich Manuscript. It provides infrastructure for rigorous computational study.