Ched-ai/voynich-eva

Voynich Manuscript EVA Transcription

Dataset Summary

This dataset provides a line-level transcription of the Voynich Manuscript in EVA (Extensible Voynich Alphabet) encoding. The Voynich Manuscript is a 15th-century codex written in an undeciphered script, held at Yale University's Beinecke Rare Book & Manuscript Library (MS 408).

This transcription is derived from the Zandbergen-Landini (ZL) transcription, the most complete and accurate EVA transcription available.

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

Dataset Structure

Data Instance

{
  "page_id": "f1r",
  "line_number": 1,
  "line_index": 1,
  "line_id": "f1r:1",
  "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",
  "line_type": "paragraph",
  "position": "@",
  "quire": "A",
  "section": "text_only",
  "currier_language": "A",
  "hand": "1",
  "illustration_type": "T",
  "char_count": 43,
  "word_count": 10,
  "has_uncertain": false,
  "has_illegible": false,
  "has_alternatives": true,
  "has_high_ascii": true,
  "source": "zandbergen_landini",
  "source_version": "bf5b6d4ac1e3"
}

Data Fields

Field	Type	Description
page_id	string	Folio identifier (e.g., "f1r" for folio 1 recto)
line_number	int	Source locus ID within page (may have gaps - preserves original numbering)
line_index	int	Sequential index within page (always 1..n, no gaps)
line_id	string	Unique identifier: {page_id}:{line_number}
text	string	Raw transcription with IVTFF markup
text_clean	string	Cleaned text (markup removed, @NNN; codes preserved - lossless)
line_type	string	Type: "paragraph", "label", "circle", "radius"
position	string	Position indicator: "@" (new), "+" (continue), "=" (end), "*" (new para)
quire	string	Quire identifier (A-T)
section	string	Manuscript section (herbal, astronomical, biological, etc.)
currier_language	string	Currier's language classification: "A" or "B"
hand	string	Scribe hand identifier
illustration_type	string	Type of illustration on page (H=herbal, A=astro, etc.)
char_count	int	Count of EVA characters
word_count	int	Count of words (separated by . or ,)
has_uncertain	bool	Contains uncertain reading markers (?)
has_illegible	bool	Contains illegible markers
has_alternatives	bool	Contains alternative readings ([a:b])
has_high_ascii	bool	Contains high-ASCII codes (@NNN;)
source	string	Source transcription identifier
source_version	string	SHA256 hash prefix of source file

Data Splits

Split	Records	Description
train	4,072	Complete transcription (no splits)

Dataset Creation

Source Data

• Source: Zandbergen-Landini (ZL) Transcription v3b
• URL: https://www.voynich.nu/data/ZL3b-n.txt
• Format: IVTFF 2.0 (Intermediate Voynich Transliteration File Format)
• Attribution: René Zandbergen and Gabriel Landini

The ZL transcription is based on the European Voynich Manuscript Transcription (EVMT) project and represents the most accurate complete transcription of the Voynich Manuscript.

Processing

Built by VCAT processing scripts:

1. Parse IVTFF format
2. Extract page-level metadata from headers
3. Clean text (remove markup, keep first alternatives)
4. Validate against EVA character set
5. Compute line-level statistics
6. Export to Parquet

Pipeline: fetch → parse → validate → export

Repository: voynich-data

Validation

• Schema validation: ✓ Passed
• EVA character validation: ✓ 170,564 characters validated
• All smoke tests: ✓ Passed (10/10)

Statistics

Metric	Value
Total pages	226
Total lines	4,072
Total EVA characters	170,564
Total words	33,711
Paragraph lines	3,957
Label lines	115
Language A pages	114
Language B pages	82

Top Characters by Frequency

Character	Count	Percentage
o	22,022	12.9%
e	17,823	10.4%
h	16,133	9.5%
y	15,819	9.3%
a	12,628	7.4%
c	11,917	7.0%
d	11,673	6.8%
i	10,745	6.3%
k	9,854	5.8%
l	9,380	5.5%

Considerations

Limitations

• Transcription uncertainty: The Voynich script is ambiguous; some readings are uncertain (flagged with has_uncertain and has_alternatives)
• No semantic meaning: This is a character-level transcription; the text remains undeciphered
• Section classification: Section assignments are approximate and may differ from other sources
• Single transcription: This dataset uses ZL only. The two major EVA transcriptions (ZL and Takahashi) are fully identical on only ~29% of lines and differ substantively on 16% — see voynich-transcription-mismatch to quantify how this affects your analysis

Schema Status

⚠️ Pre-1.0: Schema may change. Pin to specific versions for reproducibility.

Related Datasets

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

• voynich-eva (this dataset) - Line-level EVA transcription
• voynich-manuscript-metadata - Page, folio, quire metadata
• voynich-transcription-mismatch - Cross-transcription comparison

Datasets can be joined on page_id.

Quick Start

from datasets import load_dataset

# Load the dataset
ds = load_dataset("Ched-ai/voynich-eva")

# View first record
print(ds["train"][0])

# Filter by Currier language
lang_a = ds["train"].filter(lambda x: x["currier_language"] == "A")
lang_b = ds["train"].filter(lambda x: x["currier_language"] == "B")

# Get herbal section
herbal = ds["train"].filter(lambda x: x["section"] == "herbal")

# Get clean text for analysis
texts = [row["text_clean"] for row in ds["train"]]

Versioning

• Current version: v0.2.3
• Schema status: Pre-1.0 (may change)
• Source version: ZL3b (May 2025)

Licensing

Processing code and dataset structure: MIT License

Underlying transcription data: The ZL transcription is provided by voynich.nu for research purposes. The transcription authors (Zandbergen & Landini) have not published an explicit license statement. Based on the public availability and academic nature of the work, we believe research use is permitted. Users should:

• Cite the original transcribers (see Source Citation below)
• Verify rights independently for commercial applications
• Contact the transcription authors if in doubt about specific uses

This dataset is released as a research resource. If you are aware of more specific licensing terms, please open an issue. See SOURCES_LICENSE.md for full provenance details for all transcription sources.

Citation

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

Source Citation

@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.