Persepolis

A multimodal corpus of the Achaemenid ceremonial capital at Persepolis (Pārsa, Takht-e Jamshid) and the adjacent royal necropolis at Naqsh-e Rustam.

The corpus combines photographs of the standing reliefs and architecture, an archival finding-aid index covering the 1923–1934 excavations of Ernst Herzfeld, and the cuneiform tablet catalog from the same provenience. 13,480 records in all: 5,650 images bundled in this repo (4.6 GB), 1,624 archival finding-aid entries, and 6,089 cuneiform catalog entries with publication and language metadata.

config	rows	size	modality
wikimedia_images	5,767	4.6 GB images	photographs and drawings of the standing site
wikimedia_hires	37	3.7 GB TIFFs	high-resolution originals for delegation reliefs and BNF Persepolis plates
wikimedia_deferred	14	~25 MB PNGs	rasterized royal-tomb inscription PDFs (Tomb ARTXA2, DAR1A2-2, DAR2A2, X1A2, WELLA4) and 4 SVGs (Naqsh-e Rustam, Persepolis, Apadana-stair schemas)
smithsonian_finding_aid	1,624	metadata only	Herzfeld FSA.A.06 catalog, Persepolis-tagged
cdli_catalog	6,089	metadata only	Pārśa-provenience cuneiform artifacts
cdli_inscriptions	5,663	ATF text	C-ATF transliterations for the same artifacts
delegation_labels	24	metadata	one row per named Apadana delegation, with Wikipedia summary and file count
wikimedia_delegation_assignments	253	metadata	per-file delegation assignment from Wikimedia subcategory; flags files appearing in multiple delegations
wikimedia_image_clusters	5,701	metadata	dHash + pHash per image, with cluster id and cluster role flagging the 470 near-duplicates across 126 multi-image clusters
tasks	358	task definitions	benchmark splits for apadana_delegation_classification (192 records, 19 classes, cluster-disjoint 70/15/15 split, seed 42) and relief_motif_classification (156 records, 7 classes, same split policy)
default	13,480	full manifest	union of the catalog configs (inscriptions are joinable on cdli_inscriptions.artifact_id == cdli_catalog.source_id; delegation labels are joinable on wikimedia_delegation_assignments.wikimedia_title == wikimedia_images.title)

Loading

from datasets import load_dataset

ds  = load_dataset("phanerozoic/persepolis", split="all")
img = load_dataset("phanerozoic/persepolis", "wikimedia_images", split="all")
fa  = load_dataset("phanerozoic/persepolis", "smithsonian_finding_aid", split="all")
cd  = load_dataset("phanerozoic/persepolis", "cdli_catalog", split="all")
ins = load_dataset("phanerozoic/persepolis", "cdli_inscriptions", split="all")

Image rows carry a file_path column pointing at the bundled file under images/wikimedia/<sha[:2]>/<sha[2:4]>/<sha>.jpg.

Coverage

The Wikimedia stream walks the Persepolis World Heritage category tree with bounded recursion. It captures:

• the Apadana stairs (north and east), with the 23 named ethnic delegation panels each in its own subcategory: Arabian, Arachosian, Arian, Armenian, Assyrian, Babylonian, Bactrian, Cappadocian, Carian, Drangianian, Egyptian, Elamite, Gandarian, Indian, Ionian, Libyan, Lydian, Median, Nubian, Sagartian, Scythian, Sogdian, Susian, Thracian
• the Tachara, Hadish, Tripylon, Treasury, and the Hundred-Column Hall
• the Gate of All Nations
• recurring relief motifs: lion-attacking-bull, royal hero stabbing monsters (lion, griffin, bull, composite figure), audience scene of Darius / Xerxes, parasol-bearer, Faravahar, lotus, cypress, Achaemenid warriors
• the royal necropolis at Naqsh-e Rustam: tombs of Darius I, Xerxes I, Artaxerxes I, Darius II, plus the Naqsh-e Rajab Sasanian-period reliefs
• Old Persian inscription panels including DNa

Schema

column	type	meaning
record_id	string	stable id source:source_id
source	string	wikimedia, smithsonian, cdli
source_id	string	source-native identifier
source_url	string	landing page on the source
modality	string	image, metadata
title	string	short title
description	string	longer description or archival path
date_original	string	original date when known
license_canonical	string	canonical license id
license_raw	string	source-declared license string
license_url	string	canonical license URL
attribution	string	required attribution text
share_alike	bool	derivative requires share-alike
attribution_required	bool	use requires attribution
file_url	string	original file URL on the source
file_url_thumb	string	resized thumbnail URL when available
file_path	string	bundled file relative path (image rows)
sha256	string	sha256 of the bundled file
mime, width, height, byte_size	various	image metadata
categories	list	source taxonomy tags
notes	string	provenance notes
raw_metadata	string	JSON-encoded source extras
date_iso	string	parsed ISO-8601 date when extractable; "" otherwise

Benchmark tasks

The tasks config defines two named classification tasks with deterministic cluster-disjoint train/val/test splits. "Cluster-disjoint" means every near-duplicate cluster (from wikimedia_image_clusters) is placed in a single split, so train/test leakage from repeat photographs of the same panel cannot occur.

apadana_delegation_classification — 19-class image classification on the unambiguously delegation-tagged Apadana panels. 192 records, stratified 70/15/15 split, seed 42. Five of the 24 named delegations (Carian, Drangianian, Egyptian, Elamite, Susian) are excluded from this task because every Wikimedia file in those subcategories is also tagged as belonging to another delegation (label-conflict, see wikimedia_delegation_assignments).

relief_motif_classification — 7-class image classification on the recurring relief motifs that have dedicated Wikimedia subcategories: audience-scene (112), faravahar (26), parasol-bearer (7), royal-hero-killing-{bull, composite-figure, griffin, lion} (3, 4, 1, 3). Long-tailed by design.

The catalog configs support additional usage. The Smithsonian and CDLI rows cross-reference the same site at different layers (1934 archival photo records; 6,089 cuneiform tablets from the fortification archive) and support retrieval against the bundled photographs. Old Persian, Elamite, and Akkadian inscriptions are tagged in cdli_catalog.languages with companion photographs in wikimedia_images.

Sources

• Wikimedia Commons. Photographs and drawings under per-file CC0, public domain, CC BY 1.0–4.0, or CC BY-SA 1.0–4.0. Rights metadata is queried per-file via the MediaWiki imageinfo API at harvest time and preserved in the license_* columns. NC/ND/ARR files are excluded.
• Smithsonian Open Access — Ernst Herzfeld Papers (FSA.A.06). Catalog metadata is CC0 and is what this dataset ships. The digital reproductions of the photographs carry a collection-level rights statement that requires permission from the National Museum of Asian Art Archives, so the images stay at the source. source_url resolves to the canonical SOVA record for image-reproduction requests.
• Cuneiform Digital Library Initiative. CDLI catalog text is freely reusable with attribution. Includes catalog metadata for 6,089 Pārśa-provenience artifacts: designation, museum number, language, period, genre, materials, publication history. CDLI tablet photographs are non-commercial under CDLI Terms of Use and are not bundled.

Caveats

• The 23-delegation labels for Apadana panels come from the Wikimedia subcategory in which a file appeared. They have not been re-verified against Walser 1966 Die Völkerschaften auf den Reliefs von Persepolis or Roaf 1983 Sculptures and Sculptors at Persepolis.
• Site names appear in multiple transliteration conventions across the source records (Persepolis, Pārsa, Pārśa, Takht-e Jamshid; Naqsh-e Rustam, Naqsh-i-Rustam, Naqsh-e Rostam). Cross-reference by record_id.
• The royal-tomb inscription PDFs and four cartographic SVGs are rasterized and ship under wikimedia_deferred. 95 PDFs and 2 SVGs of 1971 Pahlavi-era newspaper coverage of the 2500-year celebration are dropped as out-of-scope for an Achaemenid iconographic dataset. WebM walkthroughs are not included.
• The delegation_labels and wikimedia_delegation_assignments configs document the bootstrap delegation tag for each Apadana panel and flag the 27 cases where a single file is tagged as belonging to two or more delegations on Wikimedia. These conflicts are unresolved and represent the upper bound on per-image label noise in the bootstrap labels.
• 470 of 5,701 hashed images are flagged as near-duplicates in wikimedia_image_clusters (126 multi-image clusters). The Met Museum Luigi Pesce 1840s–60s albumen prints account for many of these: consecutive DP\d+ accession numbers are paired scans of the same object from different angles. Cluster-disjoint splits in tasks prevent leakage across train/val/test from these duplicates.

License

The compilation is CC BY-SA 4.0. Per-record licenses are preserved in license_canonical and license_url and must be respected when reusing individual items. Admitted licenses are CC0, public domain, CC BY 1.0–4.0, CC BY-SA 1.0–4.0, and the CDLI attribution-style catalog license. NC, ND, ARR, and unstated-license records are excluded.

Citation

@dataset{persepolis_corpus,
  title   = {Persepolis: a multimodal corpus of the Achaemenid ceremonial capital},
  author  = {Norton, Charles Cornell},
  year    = {2026},
  url     = {https://huggingface.co/datasets/phanerozoic/persepolis},
  license = {CC BY-SA 4.0}
}

Per-record attribution from attribution and license_url is also required: Wikimedia Commons attribution lines, the Ernst Herzfeld Papers collection citation for finding-aid records, and the Cuneiform Digital Library Initiative for catalog records.

Acknowledgments

Wikimedia Commons contributors, the Smithsonian Institution National Museum of Asian Art Archives, and the Cuneiform Digital Library Initiative.