tadad/twkm-maya-glyphs

TWKM Maya Glyphs

An ML-ready, relational snapshot of the University of Bonn's Text Database and Dictionary of Classic Mayan (TWKM) digital sign catalogue. It packages the catalogue as seven Parquet configurations and embeds the explicitly CC BY 4.0 standardized graph drawings in the graphs configuration.

This is an independent preservation and interoperability package, not an official TWKM publication. The source of authority for sign classification and interpretation remains TWKM.

Why this dataset exists

This dataset retains TWKM identifiers while making the data directly loadable by Hugging Face Datasets, DuckDB, Polars, pandas, and other Parquet readers.

Snapshot

Retrieved from the live English catalogue on 2026-09-04.

Configuration	Rows	Unit
graphs	1,606	Concrete graph/allograph records; 1,583 rows have embedded drawings
signs	1,098	Abstract sign records
decipherments	735	Proposed readings, types, confidence levels, and evidence criteria
artefacts	639	Referenced inscription-bearing objects
concordances	7,705	Mappings from TWKM graph numbers to eleven other catalogues
bibliography	7,686	Bibliographic authorities referenced by the catalogue
entities	1,304	Catalogues, iconographic concepts, materials, people, and places

The source snapshot is identified by SHA-256:

d414f9aed953b5975a777bcd37198898828a51dd69e0ecec5b5ba47824682812

See source/manifest.json, source/drawing-manifest.jsonl, and SHA256SUMS for file- and object-level provenance.

Usage

from datasets import load_dataset

graphs = load_dataset("tadad/twkm-maya-glyphs", "graphs", split="train")
signs = load_dataset("tadad/twkm-maya-glyphs", "signs", split="train")

print(graphs[0]["graph_no"])
graphs[0]["image"].show()

The split is named train only for compatibility with the Hub ecosystem. It is the complete snapshot, not a recommended experimental split. Researchers should create artifact-, site-, period-, and allograph-aware partitions to avoid leakage.

Data model

TWKM distinguishes an abstract sign from a concrete graph or allograph. Image-recognition work should generally use graph_no as the observed visual label and join through sign_no when the task requires linguistic interpretation.

Relationships are represented by stable source identifiers:

graphs.sign_no                 -> signs.sign_no
graphs.artefact_ids[]          -> artefacts.artefact_id
graphs.concordance_ids[]       -> concordances.concordance_id
concordances.external_catalog_id -> entities.entity_id where entity_type == "catalog"
decipherments.sign_no          -> signs.sign_no (when sign_reference_resolved is true)

Nested source fields that vary structurally are retained losslessly as compact JSON strings. These columns end in _json. The unmodified catalogue response is preserved at source/twkm-sign-catalog.en.json.

graphs

Important fields include:

• graph_no, sign_no, and variant: TWKM identifiers.
• image: embedded standardized drawing, or null where TWKM publishes no image identifier.
• source_image_id, source_image_url, and image_sha256: image provenance and integrity.
• translation and occurrence_count: source catalogue values, not independently verified labels.
• artefact_ids, iconography_ids, and concordance_ids: foreign keys.
• drawing_creator and drawing_license: per-row attribution.

There are 1,581 unique downloadable drawing files. Two images are reused by more than one graph record, producing 1,583 image-bearing rows. Twenty-three graph records have no source image identifier.

Other configurations

• signs records abstract signs and translations.
• decipherments stores logographic/phonographic values and TWKM's eight-level confidence assessment. The source snapshot has one unresolved relationship: decipherment 285 names sign 716, which is absent from signs. The row is retained losslessly and marked sign_reference_resolved = false; the build manifest records the exception.
• artefacts stores object labels, types, dates, Maya Long Count values, and links to controlled entities.
• concordances connects TWKM graphs to historical and contemporary catalogue identifiers. Third-party catalogue images are explicitly excluded.
• bibliography stores citation text and source Zotero URIs.
• entities combines small controlled vocabularies with an entity_type discriminator.

Feature schemas are versioned under schema/.

Provenance and preparation

The build process:

1. Retrieves the catalogue's public “Complete catalog data (JSON)” representation.
2. Records retrieval time, response validators, byte count, and SHA-256.
3. Retrieves only TWKM's standardized graph drawings from the catalogue's graph-image collection.
4. Validates every retrieved file as a decodable JPEG and records dimensions and SHA-256.
5. Normalizes the relational entities into Parquet without changing source identifiers or readings.
6. Embeds licensed drawings in the graphs Parquet file.
7. Validates primary-key uniqueness, cross-table references, row counts, media presence, and the third-party-image exclusion boundary.

Rebuild and validate locally with:

uv run --python 3.12 --with 'datasets>=4.0,<5' --with 'pillow>=11,<13' scripts/build_dataset.py
uv run --python 3.12 --with 'datasets>=4.0,<5' --with 'pillow>=11,<13' scripts/validate_dataset.py

Validation results are committed in reports/validation-report.json.

Licensing and rights

This repository is marked license: other because the components do not share one confirmed licence.

• Standardized graph drawings: Christian Prager, Text Database and Dictionary of Classic Mayan, University of Bonn, licensed CC BY 4.0 by TWKM. Attribution is repeated in every image-bearing row.
• Complete catalogue data and normalized metadata: TWKM describes the catalogue as open access and available for scholarly and public use, but its legal notice does not assign the entire database an explicit Creative Commons licence. No additional licence is asserted here.
• Third-party catalogue images: not redistributed. Their identifiers remain in concordances.external_image_ids solely to preserve the source relationship.
• This repository's build and validation scripts: released under Apache-2.0; see LICENSE-CODE.

Read RIGHTS.md and the TWKM legal notice before reuse. Users planning redistribution or commercial use of the catalogue metadata should obtain clarification from TWKM. This dataset does not provide legal advice.

Coverage and known omissions

This is a sign-catalogue snapshot, not a complete corpus of Maya inscriptions.

TWKM's project description estimates more than 10,000 surviving text carriers. The current public catalogue exposes 639 referenced artifacts, a rough public-coverage proxy of at most about 6.4% by carrier. This is not a percentage of glyph blocks, words, or fully transcribed texts: carrier sizes vary, records may be partial, and unpublished internal TextGrid data may exist.

Not included:

• Full occurrence-level TEI/XML inscription transcriptions.
• Glyph-block reading-order trees or complete linguistic annotations from ALMAH.
• Bounding boxes or segmentation masks on original inscriptions.
• Artifact photographs, scans, or drawings with unconfirmed rights.
• Third-party catalogue illustrations.

The 2026 IDIOM+/Text+ project is developing an FCS-compatible interface connecting TWKM's RDF metadata, TEI/XML inscriptions, and JSON linguistic data. A future version should ingest that corpus only after a documented endpoint and redistribution terms become available.

Intended uses

Appropriate uses include:

• Sign- and allograph-classification research.
• Visual retrieval and representation learning.
• Catalogue concordance and authority reconciliation.
• Digital epigraphy interfaces and teaching tools.
• Controlled experiments in structured glyph recognition.

This dataset alone is not sufficient for reliable OCR, transliteration, translation, decipherment, or historical claims. translation and decipherments reflect evolving expert scholarship and include explicit uncertainty.

Bias, leakage, and responsible use

• The catalogue is shaped by preservation, excavation, publication, collection, and scholarly-selection biases.
• Standardized drawings remove much of the damage, texture, lighting, material, and scribal variation found in real inscriptions.
• Multiple graphs can derive from the same sign or artifact; random row splits can severely overestimate model performance.
• Frequently occurring, well-preserved, and deciphered signs are better represented than rare or uncertain ones.
• Maya hieroglyphic writing and its study are part of living Maya cultural heritage. Applications should involve relevant Maya scholars and communities, preserve attribution and uncertainty, and avoid presenting automated readings as cultural or scholarly authority.

Maintenance

TWKM is a living catalogue updated frequently. Snapshot updates should preserve the previous source checksum, run the committed build and validation scripts, document additions/removals/reclassifications in CHANGELOG.md, and create a Hub tag for materially different versions.

Issues should identify the configuration, source snapshot SHA-256, relevant TWKM identifier, and evidence. Corrections to the authoritative catalogue should also be reported to TWKM.

Citation

Please cite both this packaging dataset and the underlying TWKM catalogue. TWKM's requested catalogue citation is reproduced in CITATION.cff; the authoritative wording is available in its 2025 working paper.

Acknowledgements

All scholarly catalogue work and graph drawings are credited to the TWKM project and its contributors. This package contributes format conversion, preservation metadata, validation, and Hugging Face interoperability only.