Ched-ai/voynich-transcription-mismatch

Voynich Transcription Mismatch Index

Dataset Summary

This dataset provides a line-by-line comparison across five different transcription sources of the Voynich Manuscript. It tracks agreements and disagreements between transcribers, enabling research on transcription uncertainty and consensus.

The primary comparison is between EVA-based transcriptions (ZL and IT), with additional tracking of Currier, FSG, and v101 transcription systems.

Key Statistics

Metric	Value
Total lines	4,072
Lines with both EVA sources	4,069
EVA exact matches	901 (22.1%)
EVA normalized matches	293 (7.2%)
EVA high similarity (≥95%, but not identical)	2,220 (54.6%)
EVA substantive disagreements (<95%)	655 (16.1%)

The headline finding: the two major EVA transcriptions are fully identical on only 29.3% of lines (exact + normalized matches), and differ substantively on 16.1%. The remaining 54.6% are close (≥95% similar) but not identical. Any analysis built on a single transcription inherits this uncertainty — this dataset exists to make it quantifiable.

Transcription Sources

ID	Name	Alphabet	Lines	Description
zl	Zandbergen-Landini	EVA	4,072	Primary reference (most complete)
it	Takahashi	EVA	4,069	Secondary EVA transcription
cd	Currier/D'Imperio	Currier	2,154	Historical Currier alphabet
fg	Friedman Study Group	FSG	3,980	NSA research group transcription
gc	Glen Claston	v101	4,070	High-granularity v101 alphabet

Dataset Structure

Data Instance

{
  "page_id": "f1r",
  "line_number": 1,
  "line_id": "f1r:1",
  "zl_text": "fachys.ykal.ar.ataiin.shol.shory.cthres.y.kor.sholdy",
  "it_text": "fachys.ykal.ar.ataiin.shol.shory.cthres.y.kor.sholdy",
  "cd_text": "VAS92.9FAE.AR.APAM.ZOE.ZOR9.QOR92.9.FOR.ZOE89",
  "fg_text": "8ACIS.ICAL.AB.APAM.SOL.SOB2.QPBES.2.COB.SOLD2",
  "gc_text": "fZC*9S,9kAL,Ar,AtAiiN,Sol,Sor9,C!reS,9,kor,SolD9",
  "status": "exact_match",
  "eva_agreement": true,
  "similarity_score": 1.0,
  "sources_present": ["zl", "it", "cd", "fg", "gc"],
  "sources_missing": [],
  "notes": []
}

Data Fields

Field	Type	Description
page_id	string	Page identifier
line_number	int	Line number within page
line_id	string	Unique identifier: {page_id}:{line_number}
zl_text	string/null	ZL (EVA) transcription
it_text	string/null	IT (EVA) transcription
cd_text	string/null	Currier transcription
fg_text	string/null	FSG transcription
gc_text	string/null	v101 transcription
status	string	Comparison status (see below)
eva_agreement	bool/null	EVA transcriptions agree
similarity_score	float/null	Text similarity (0-1)
sources_present	list[string]	Sources with this line
sources_missing	list[string]	Sources missing this line
notes	list[string]	Additional notes

Status Values

Status	Description	Count
exact_match	ZL and IT are identical	901
normalized_match	Match after removing uncertainty markers	293
high_similarity	≥95% similar	2,220
content_mismatch	<95% similar	655
zl_missing	Line missing from ZL	0
it_missing	Line missing from IT	3

Use Cases

Finding Uncertain Lines

from datasets import load_dataset

ds = load_dataset("Ched-ai/voynich-transcription-mismatch")

# Find lines where transcribers disagree
mismatches = ds['train'].filter(lambda x: x['status'] == 'content_mismatch')
print(f"Disagreements: {len(mismatches)}")

# Find low-similarity mismatches
low_sim = ds['train'].filter(
    lambda x: x['similarity_score'] is not None and x['similarity_score'] < 0.5
)
print(f"Low similarity: {len(low_sim)}")

Robustness Testing

Use this dataset to test whether your analysis results are robust across transcription variants:

# Check if a pattern appears in both ZL and IT
def pattern_in_both(record, pattern):
    zl_has = pattern in (record['zl_text'] or '')
    it_has = pattern in (record['it_text'] or '')
    return zl_has and it_has

Cross-Reference with Metadata

Join with voynich-manuscript-metadata to analyze disagreements by section:

# Which sections have most disagreements?
meta = load_dataset("Ched-ai/voynich-manuscript-metadata", "pages")
# Create page_id -> section mapping
# Then filter mismatches by section

Dataset Creation

Processing

1. Parse all five transcription files (IVTFF format)
2. Align lines by line_id (page:line_number)
3. Compare EVA transcriptions with normalization
4. Compute similarity scores for mismatches
5. Record presence/absence for all sources

Methodology Notes

• EVA comparison: Only ZL and IT are directly compared (same alphabet)
• Normalization: Removes uncertainty markers (?, !, *) and editorial brackets
• Similarity: SequenceMatcher ratio (0-1)
• High similarity threshold: ≥0.95

Considerations

Limitations

• Cross-alphabet comparison not attempted: CD, FG, GC use different alphabets; direct text comparison not meaningful
• Line alignment by ID: May miss reflow differences where transcribers split lines differently
• Normalization choices: May hide meaningful distinctions

Why Transcriptions Disagree

• Glyph identification: Scribal hands are difficult to read
• Character boundaries: Where one glyph ends and another begins
• Damaged areas: Faded or damaged sections
• Word boundaries: Interpreting spaces between words
• Transcription conventions: Different systems make different choices

Related Datasets

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

• voynich-eva: Complete EVA transcription
• voynich-manuscript-metadata: Page, folio, quire metadata

Quick Start

from datasets import load_dataset

ds = load_dataset("Ched-ai/voynich-transcription-mismatch")

# Summary statistics
total = len(ds['train'])
exact = len(ds['train'].filter(lambda x: x['status'] == 'exact_match'))
normalized = len(ds['train'].filter(lambda x: x['status'] == 'normalized_match'))
high_sim = len(ds['train'].filter(lambda x: x['status'] == 'high_similarity'))
mismatch = len(ds['train'].filter(lambda x: x['status'] == 'content_mismatch'))

print(f"Fully identical:          {(exact + normalized) / total:.1%}")
print(f"Close but not identical:  {high_sim / total:.1%}")
print(f"Substantive disagreement: {mismatch / total:.1%}")

Licensing

Processing code and dataset structure: MIT License

Underlying transcription data: The five source transcriptions (ZL, IT, CD, FG, GC) are scholarly works whose authors have not published explicit license statements. They are publicly distributed for research via voynich.nu. This dataset is released as a research resource: cite the original transcribers, and verify rights independently for commercial applications. See SOURCES_LICENSE.md for full provenance details.

Citation

@misc{voynich-transcription-mismatch,
  author = {VCAT Contributors},
  title = {Voynich Transcription Mismatch Index},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/Ched-ai/voynich-transcription-mismatch}
}

References

• Zandbergen, R. & Landini, G. ZL Transcription. https://www.voynich.nu/
• Takahashi, T. IT Transcription.
• Currier, P. (1976). Papers on the Voynich Manuscript. NSA.
• Friedman First Study Group. FSG Transcription.
• Claston, G. v101 Transcription.

Contact

• Repository: GitHub
• Issues: GitHub Issues