(1) Overview
Repository location
The dataset is openly available on Zenodo: https://doi.org/10.5281/zenodo.20754025.
Context
Fulvio Orsini (1529–1600) was librarian to the Farnese family in Rome and one of the founders of systematic antiquarian scholarship (de Nolhac, 1887; Cellini, 2004; Matteini, 2013). His Familiae Romanae quae reperiuntur in antiquis numismatibus (Orsini, 1577) is the first systematic catalogue of Roman coinage organised by issuing family (gens). The work matches physical coins from the Farnese cabinet — one of the foremost numismatic collections of the Renaissance — to literary attestations of the Roman families that produced them, drawing on Cicero, Livy, Appian, and other ancient sources. The book contains 223 copperplate plates illustrating coins from 166 gentes, accompanied by a Latin commentary. It constitutes the origin point of numismatic prosopography as a scholarly method.
Despite its foundational status, no complete concordance of Orsini’s engravings against modern reference systems has been published in the four and a half centuries since its appearance. The book itself has been read and cited continuously — the Fontes Inediti Numismaticae Antiquae database (FINA ID 3015) records its appearance in some fifty pieces of learned correspondence alone — but that reception has never included a systematic mapping of its engravings to modern type-references. Babelon’s Description historique et chronologique des monnaies de la République romaine (Babelon, 1885–1886), organised by gens in conscious imitation of Orsini, provides partial cross-references but does not systematically map every engraving to a Crawford number; Cohen (1857) and Sydenham (1952) likewise precede the modern standard. Crawford’s Roman Republican Coinage (Crawford, 1974) — now that standard for Republican types — contains no Orsini concordance. The result is that the 767 individual coin images in the Familiae Romanae have never been systematically identified in modern terms. This dataset fills that gap.
The concordance was built as an independent research project (March–June 2026) from a single copy of the 1577 first edition in the author’s collection (contemporary limp vellum, 36 × 24.5 cm, 107 folia complete per the printer’s regestum). The copy preserves unreported cancel plates on pp. 101 and 104 correcting a Furia/Fufia plate swap, a cancel on the Aquillia plate (p. 29), and eight French manuscript annotations datable c. 1848. The principal numismatic findings are reported separately.
Figure 1
The self-contained HTML concordance viewer, showing the Iulia family (pp. 113–123). Each entry displays the SAM-extracted engraving, museum specimen photographs (obverse and reverse), type descriptions, Crawford/RIC identification, CRRO reference link, and Orsini’s commentary. Specimen photographs © American Numismatic Society (numismatics.org), reproduced under CC BY-NC 4.0; not covered by the CC-BY licence of this article.
(2) Method
The pipeline from physical book to structured concordance comprises four stages (Figure 2).
Figure 2
From book page to concordance entry. Left: the full page of Orsini’s Familiae Romanae (1577, p. 7). Centre: the SAM-extracted engraving of Aemilia_2, obverse (laureate head of Roma) and reverse (equestrian statue on arches). Right: the matching museum specimen (Crawford 291/1, Mn. Aemilius Lepidus, c. 114 BC; ANS). Specimen photograph © American Numismatic Society (numismatics.org), reproduced under CC BY-NC 4.0; not covered by the CC-BY licence of this article.
Steps
Stage 1: Image segmentation
High-resolution photographs of all pages were taken under controlled lighting (iPhone 17, using the vFlat scanning application, under diffuse natural lighting, ~48 MP per frame). The 223 plates follow a consistent visual grammar: a cartouche banner bearing the gens name, followed by rows of engraved coin pairs (obverse and reverse side by side within circular frames), with small metal-designation cartouches (AR, AE, AV) between the two sides. Blank roundels indicate types the Farnese cabinet lacked.
Individual coin pairs were extracted using Meta’s Segment Anything Model (SAM ViT-B; Kirillov et al., 2023), guided by a deterministic grid algorithm. The grid algorithm exploits the regularity of Orsini’s plate layout: it detects the plate rectangle using edge detection and Hough line transforms, identifies the cartouche banner via horizontal intensity profiling, divides the remaining area into rows, and computes bounding boxes for each coin pair. Per-family layout specifications were manually compiled for all 166 gentes, encoding the number of coins, row layout (e.g., “2,2,1” for five coins in three rows), and any anomalies.
SAM received point prompts (grid box centres) and bounding-box prompts (expanded by 15%). Quality checks rejected outputs where the detected radius differed from the grid estimate by more than 40%, the centre was displaced by more than 50% of the estimated radius, or the mask area fell below 30% of the expected area. SAM succeeded on approximately 95% of coin pairs; the remainder used grid fallback. The process produced 767 PNG images organised in directories by gens.
Stage 2: Automated identification
For each coin image, obverse and reverse descriptions were generated using a multimodal language model (Claude Opus 4.5, model identifier claude-opus-4-5), prompted to describe only what was visible on the engraving — head type, direction, legend text, reverse scene, metal cartouche — without reference to catalogue knowledge, to avoid confirmation bias. A strict vision-first protocol governed all identifications: the engraving was described before consulting any database, to prevent the model from projecting catalogue descriptions onto the images — the single largest source of errors in early sessions. Legend text was then matched against the Coinage of the Roman Republic Online (CRRO; Meadows & Gruber, 2014) using fuzzy string matching (Levenshtein distance ≤ 2 after normalization). Legends were stripped of interpuncts, spaces, and ligatures and upper-cased before comparison — for example, the engraved legend C·ABVRI·GEM was normalised to CABVRIGEM). Where the legend yielded a unique match, the identification was accepted provisionally. Where multiple Crawford types shared the same legend, obverse and reverse type descriptions were compared to disambiguate. Head direction proved unreliable as a diagnostic feature: Orsini’s engravings commonly mirror coin images as a consequence of the intaglio printing technique.
Non-Crawford types were routed to secondary databases: Online Coins of the Roman Empire (OCRE) for Imperial types (numismatics.org/ocre/), Roman Provincial Coinage Online (RPC) for provincial issues, and IRIS (greekcoinage.org) for Greek provincial. This stage resolved approximately 90% of identifications (c. 680 of 758 coins); the remaining ~10% (c. 78 coins) were carried forward to Stage 4 (Human validation, below). The 91 Roman Imperial Coinage (RIC) I identifications are predominantly Augustan moneyers (19–9 BC) whom Orsini filed under their Republican family names.
Stage 3: Vision-based verification
Specimen photographs were retrieved programmatically from the American Numismatic Society (ANS; 667), the BnF/Gallica (38), the British Museum (16), the Berlin Münzkabinett (15), the Ashmolean (11), CoinArchives (2), and the Kunsthistorisches Museum Vienna (1). A vision sweep — performed with the multimodal model (Claude Sonnet 4, model identifier claude-sonnet-4-20250514) — then presented each Orsini engraving alongside the candidate specimen (obverse and reverse), checking: (a) head direction (left vs right); (b) principal obverse type; (c) reverse scene; (d) legend compatibility. 47 coins were flagged for review. Of these, 12 proved to be genuine identification errors (subsequently corrected), eight were head-direction discrepancies where Orsini’s engraver had reversed the orientation, and 27 were false alarms caused by worn specimens or atypical engraving styles.
Stage 4: Human validation
The approximately 10% of coins not resolved by automated legend matching (c. 78 coins) required iterative expert refinement across four review rounds. These fell into three categories: coins with illegible or absent legends (c. 30), where identification relied on visual type comparison; coins with ambiguous matches (c. 35), requiring close comparison of secondary features; and coins absent from CRRO (c. 13), including three types identifiable only through Babelon (1885–1886) and Morell (1734). Recent work on object detection in numismatics (Cabral, de Iorio, & Harris, 2025) demonstrates the growing role of computational methods in the field; the present pipeline extends this approach from individual coin recognition to full-corpus concordance building. Each round produced an interactive HTML review file for side-by-side comparison. Status fields tracked each coin through: tentative → identified → confirmed → validated.
Stage 5: Quality control
The concordance is a human-validated dataset produced with the assistance of artificial intelligence (AI), and every entry was verified across five independent dimensions — identification, free-text description, specimen linkage, image integrity, and the fidelity of Orsini’s commentary — rather than against a single error figure, because these dimensions fail independently: a fluent description can mask a wrong-coin specimen photograph or a transposed reverse. The decisive principle was to check each dimension against an external ground truth. Identifications and specimens were verified against the open linked-data collections (CRRO/OCRE and the cabinets of the ANS, the British Museum, Berlin, Vienna, the Ashmolean, and the BnF), which made the dataset both auditable and repairable from the same open corpora; this corrected the silent, mechanical failures that automated matching introduces — wrong-coin specimens, obverse/reverse swaps, and mis-mapped crops. The two AI-drafted free-text layers — descriptions and commentary — were then re-checked against the engravings and against Orsini’s own 1577 text, gens by gens, with an agent proposing only evidence-anchored corrections and the human adjudicating each. The errors this surfaced — some 600 in all — fell overwhelmingly here, in the generated prose, not in the identifications, which were already secure; each was written back to the dataset, and a note was left empty wherever Orsini says nothing. Because the audit layers themselves over-flag, their flags were treated as candidates for expert adjudication, not verdicts, and residual uncertainty is exposed per entry through the status field (all 758 validated, none tentative).
The generalisable and encouraging lesson for AI-assisted catalogues is that, verified per dimension and against the sources, an AI-drafted concordance becomes a reliable scholarly resource — and it is the open linked-data ecosystem that makes such verification possible.
(3) Dataset Description
Repository name – Zenodo (CERN, Geneva).
Object name – Complete Concordance of Orsini’s Familiae Romanae (1577).
Format names and versions – JSON (a single structured concordance file, orsini_concordance.json, built from eleven per-letter working files via a rebuild script, ~1.4 MB); JSON-LD (a nomisma.org-aligned linked-data export using the nomisma ontology and resolvable nomisma.org type URIs, ~0.6 MB); CSV (a flat one-row-per-engraving export with the catalogue, metal, status, and cross-listing fields, ~0.5 MB); HTML (self-contained concordance viewer; portable version with embedded images, ~557 MB); PNG (767 extracted coin engraving images); JPEG (specimen photographs, obverse and reverse, for the identified types); Python 3.9+ scripts.
Creation dates – 2026-03-13 to 2026-04-07 (initial build); verified release (v2.0, all commentary checked against Orsini’s 1577 text) 2026-06-18.
Dataset creators – Christophe Jurczak – sole creator. ORCID: 0009-0001-8204-7333.
Language – English (descriptions, commentary). Latin (Orsini’s text, coin legends). Greek (legends on provincial types).
License – CC-BY 4.0 International.
Publication date – 2026 (v2.0 verified release).
Data model and structure
The data is organised as eleven per-letter working files, one per alphabetic group of gentes. Each file contains an array of family objects; each family contains an array of coin entries. Per-coin fields include: obverse and reverse descriptions (from the engraving), metal, modern catalogue identification, moneyer, date, verification status, reference URL, specimen URLs, Orsini’s commentary, and flagged errors. These working files are merged via a rebuild script into the single authoritative concordance file (orsini_concordance.json) that is deposited. The self-contained HTML concordance (Figure 1) presents each entry with the extracted engraving, specimen photographs, catalogue identification, and commentary; it requires no external dependencies.
What the data shows
The structured data permits a quantitative portrait of the Farnese cabinet that has not previously been available (Figure 3). Crawford types dominate (649, 85.7%), but the 91 RIC I identifications (12.0%) reveal a deliberate structural choice: Orsini extended his prosopographic method into the Augustan Principate, filing imperial moneyers under their Republican family names (RPC I provincial issues account for 13, specialist references for 4). The metal distribution — 654 silver, 89 bronze, 15 gold — reflects both the Farnese collection’s bias toward legible denarii and the practical reality that silver coins bear the moneyer legends on which prosopographic attribution depends. The gold coins are concentrated in the civil wars of 49–40 BC; the earliest gold (Crawford 28/1, c. 218 BC) is the first gold coinage of the Roman Republic. Temporally, the corpus peaks sharply in 100–1 BC: the political crisis of the late Republic — Sulla, the Social War, Caesar, the Triumvirs — generated both the most prolific coinage and the most politically prominent families.
Figure 3
Concordance statistics. Left: temporal distribution by decade, showing the concentration in the late Republic (100–1 BC). Right: composition by catalogue, metal, and period.
Two individual entries illustrate the kind of result the concordance makes possible. The first is a small bronze that Orsini, reading the legend CABE, placed in “Cabae, a city of the province of Africa” under M. Aemilius Lepidus’s triumviral command; matching the engraving to the modern corpus instead identifies it as the Cabellio issue of Gallia Narbonensis (RPC I 528) — a precise geographic re-attribution. The second is the corpus’s one entry with no genuine ancient prototype: a denarius reading BRVTVS·IMP whose reverse (globe, caduceus, ship’s rudder) is in fact the Caesarian type of L. Mussidius Longus (Crawford 494/39), the globe being Caesar’s own calendar-reform emblem and absent from all coinage of Brutus. With that type’s cornucopia dropped, a generic Roma head substituted, and — unlike every genuine BRVTVS·IMP issue, which names the lieutenant who struck it — no naming magistrate, it is a Renaissance fantasy or pasticcio that Orsini faithfully engraved from a sixteenth-century cabinet. Systematic matching to the modern corpus is exactly what exposes it, leaving it the single type with no counterpart.
Several coin types appear under multiple gentes — a structural feature of Orsini’s method. When a coin bears the names of magistrates from different families, Orsini engraved it on both plates: Crawford 475/1a appears under both Munatia and Iulia, and Crawford 505/1 under both Cassia and Servilia. These cross-references are not errors but prosopographic indexes: every family mentioned on a coin receives its own entry. The concordance also documents nine non-coin entries: eight manuscript annotations by a mid-nineteenth-century French reader traceable to Mionnet (1847), and one printing anomaly (a ghost coin on an unreported cancel plate).
(4) Reuse Potential
The pipeline — SAM segmentation, linked-data legend matching, vision-based verification — is generalisable to other illustrated numismatic catalogues of the sixteenth through nineteenth centuries (Morell, Vaillant, Eckhel, Cohen, Babelon). As a pilot transfer test — with the vision-language reading performed by hand rather than the full automated pipeline (SAM segmentation and automated specimen retrieval were not run) — the method was applied to one plate of Le Blanc’s Traité historique des monnoyes de France (1690): with Duplessy (1988) substituted for CRRO as the reference corpus, it confirmed the silver and billon types as genuine issues of Louis IX and reassigned the gold types Le Blanc had attributed to the king to the fourteenth century — reproducing against a modern reference the doubts Le Blanc had himself recorded.
The method’s reach is bounded by that reference infrastructure, however: antiquity is served by nomisma.org’s linked-open-data type corpora and aggregated specimen images, whereas post-antique national series (here Duplessy in print, with specimens scattered across dealer and museum holdings) have no equivalent, so legend-to-type matching transfers more readily than the automated retrieval of vetted comparanda. Applied across several such catalogues, the method would allow concrete research questions to be addressed at corpus scale: recovering lost or unillustrated types attested only in early modern plates; quantifying iconographic change across the Republican series; detecting engravers’ systematic distortions (such as the mirroring noted above); and comparing the holdings of different Renaissance cabinets. The dataset has already enabled substantive numismatic results in its own right — the recovery of a coin type known only from sixteenth-century engravings and the identification of a pre-Roman issue present in the Farnese cabinet — demonstrating its scholarly reuse value.
The data architecture (per-family JSON files, derived concordance, specimen links) provides a reusable template. To make the concordance directly reusable within the international numismatic Linked Open Data networks, the deposit includes a nomisma.org-aligned JSON-LD export: each identified engraving is expressed in the nomisma ontology (nmo:) and carries a resolvable nomisma.org type URI (741 of the 767 entries; the remainder are RPC or specialist types, the nine non-coin entries, and the one flagged engraving, for which the source reference is retained where applicable). This allows the data to be ingested and queried alongside the CRRO/OCRE typology portals rather than matched record by record. The 758 coin images, paired with structured identifications and specimen photographs, also constitute a labelled dataset for vision-based machine learning: iconographic similarity graphs, automatic type classification, and engraving-style analysis are all enabled by the combination of images and metadata. Preliminary experiments with SigLIP 2 and DINOv2 embeddings showed that off-the-shelf vision models capture engraving style rather than iconographic content, suggesting that domain-specific fine-tuning would be a productive direction for future work.
This work forms part of a broader turn toward AI methods in numismatics, spanning object detection (Cabral, de Iorio, & Harris, 2025) and generative approaches to coin identification and reconstruction (Altaweel & Khelifi, 2024); the present contribution is a concordance-building application within that wider development. More broadly, the project shows that AI methods can substantially lower the barrier to producing digital concordances of historical illustrated works: a phone camera, segmentation, and a multimodal language model can turn what would once have taken months of manual cataloguing into a usable draft in weeks. That draft is the starting point, not the finished product — but verified per dimension and against the sources, as described above, it becomes a reliable resource, and on that basis the method makes it feasible to revisit the great numismatic reference works of the sixteenth through nineteenth centuries more systematically than before.
Supplementary files
Supplementary File 1
Interactive HTML concordance (orsini_concordance_portable.html; HTML, ~557 MB): self-contained viewer with embedded coin engravings, specimen photographs, and full metadata for all 767 entries; opens in any modern browser without an internet connection. DOI: https://doi.org/10.5334/johd.557.s1
Supplementary File 2
Structured data (orsini_concordance.json; JSON, ~1.4 MB): the authoritative per-entry data. DOI: https://doi.org/10.5334/johd.557.s2
Supplementary File 3
Linked-data export (orsini_concordance.jsonld; JSON-LD, ~0.6 MB): nomisma.org-aligned, using the nmo: ontology and resolvable type URIs. DOI: https://doi.org/10.5334/johd.557.s3
Supplementary File 4
Flat export (orsini_concordance.csv; CSV, ~0.4 MB): one row per engraving with catalogue, metal, status, and cross-listing fields. DOI: https://doi.org/10.5334/johd.557.s4
Supplementary File 5
Pipeline prompts: the verbatim model prompts used for description, identification, and the vision-verification sweep. DOI: https://doi.org/10.5334/johd.557.s5
AI Declaration
Generative AI tools were used at three stages of the data creation pipeline: (1) generating initial obverse and reverse descriptions from the coin engraving images, using Claude Opus 4.5 (model identifier claude-opus-4-5); (2) the vision-based verification sweep, in which the model automatically compared each engraving with candidate specimen photographs, using Claude Sonnet 4 (model identifier claude-sonnet-4-20250514); and (3) a source-grounded verification of the AI-drafted descriptions and commentary, in which AI agents (Claude Sonnet 4, model identifier claude-sonnet-4-20250514) proposed corrections checked against Orsini’s own 1577 text, each adjudicated by the author. All AI-generated outputs were reviewed and validated by the author. The final concordance depends on human-verified identifications. Generative AI was also used as a writing aid in the preparation of this manuscript; the author is responsible for the final text.
Data Accessibility Statement
The complete dataset described in this paper — the structured JSON, a nomisma.org-aligned JSON-LD export, a flat CSV, and a self-contained HTML concordance viewer that embeds all extracted engraving images and specimen photographs — is openly available on Zenodo at https://doi.org/10.5281/zenodo.20754025 under a Creative Commons Attribution 4.0 (CC-BY 4.0) licence. The verbatim pipeline prompts are provided as Supplementary File 5.
Acknowledgements
Specimen photographs remain the copyright of the holding institutions and are reproduced under their individual terms of use (predominantly CC BY-NC), as credited individually in the concordance data; they are not covered by the CC-BY licence of this article. Coin-type data and identifiers are drawn from the nomisma.org linked-open-data ecosystem — Coinage of the Roman Republic Online (CRRO), Online Coins of the Roman Empire (OCRE), and Roman Provincial Coinage Online (rpc.ashmus.ox.ac.uk). This work depends entirely on these open data collections and infrastructures, whose maintainers I gratefully acknowledge.
Author Contributions
Christophe Jurczak: Conceptualization, Data curation, Investigation, Methodology, Validation, Writing – original draft.