OIKONOMIA-DB: An Auditable Economic Database of Greco-Roman Egypt

195,906 monetary facts, 350,206 gendered person mentions and 21,895 transaction principals, extracted from 61,249 documentary papyri spanning 300 BCE – 700 CE. Every row traces back to a character span in a specific document at a pinned corpus revision.

Documentary papyri — tax receipts, leases, loans, wage accounts, census returns, private letters — are the only large body of evidence for the everyday economy of the ancient world. Historians have read them one at a time for a century. OIKONOMIA-DB is the first attempt to turn the whole Duke Databank into a structured database that can be queried.

• 📄 Source corpus: Duke Databank of Documentary Papyri (DDbDP) + HGV metadata, CC BY 3.0
• 🔒 Pinned revision: d7a34f302d1e44e271256092c2b780733187b478
• 🤖 Extraction models: OIKONOMIA-Grammateus (entities) · OIKONOMIA-Homologia (relations)
• 💻 Code: github.com/abderahmane-ai/oikonomia

---

Dataset Details

Dataset Description

Eight Parquet tables. Seven hold facts at the grain of a single extracted statement; one (autonomy) is a published aggregate, shipped so a headline result is inspectable without recomputation.

Table	File	Rows	One row is	Key
documents	export/documents.parquet	61,249	one text document (the spine)	stem
monetary	monetary.parquet	195,906	one monetary amount	tm_id + span
prices	prices.parquet	98	one clean commodity unit-price	tm_id + span
taxes	taxes.parquet	592	one clean tax payment	tm_id + span
persons	persons.parquet	350,206	one PERSON mention, gendered	stem + span
principals	principals.parquet	21,895	one party to a transaction	stem + span
persons_distinct	export/persons_distinct.parquet	17,362	one distinct person (coref-lite)	person_id
autonomy	autonomy.parquet	32	one century/region bucket (derived)	dimension+bucket

Two extraction regimes produce them, and they have different error profiles that must not be blended:

• monetary, prices, taxes come from a deterministic lexicon + rules over EpiDoc-decoded numerals. Closed-class vocabulary (drachma, artaba, wheat) is matched at gazetteer ceiling. High precision, systematic misses.
• persons, principals, persons_distinct, autonomy come from the trained models, whose measured end-to-end accuracy is entity F1 0.737 (strict) and PARTY_OF F1 0.623 on predicted entities.

Dataset Sources

• Repository: https://github.com/abderahmane-ai/oikonomia
• Corpus: papyri/idp.data — DDbDP texts and HGV metadata, CC BY 3.0
• Authority files: dates from HGV, places from Pleiades, document ids from Trismegistos

---

Uses

Direct Use

Built for quantitative ancient economic history and digital papyrology: price and wage series, fiscal history, gender and legal capacity, prosopography, regional comparison. Also usable as a span-linked evaluation set for Greek information extraction, since every row carries its source offsets.

Out-of-Scope Use

• Not a training corpus for language models. It holds extracted facts, not text.
• Not a source of absolute economic aggregates. Survival bias is uncorrectable (see Limitations); shares within a bucket are interpretable, raw counts across buckets are not.
• Not a payment-direction dataset. Who paid whom is deliberately absent — the relation model scores direction at F1 0.145, too low to ship.
• Not a substitute for reading the papyrus. Every row gives you the span to go and check; for a claim about an individual document, check it.

---

Getting the data

from datasets import load_dataset

prices = load_dataset("ainouche-abderahmane/oikonomia-db", "prices", split="train")

Every table name in the inventory above is a valid config: monetary, prices, taxes, persons, principals, autonomy, documents, persons_distinct.

The whole database (14 MB) as local Parquet:

hf download ainouche-abderahmane/oikonomia-db --repo-type dataset --local-dir oikonomia-db

from huggingface_hub import snapshot_download
path = snapshot_download("ainouche-abderahmane/oikonomia-db", repo_type="dataset")

Quickstart (DuckDB)

-- Women's share of transaction principals, by type of deal.
SELECT deal_type,
       count(*) AS n_gendered,
       round(100.0 * sum(gender = 'female') / count(*), 1) AS pct_women
FROM 'principals.parquet'
WHERE gender <> 'unknown' AND deal_type <> '?'
GROUP BY 1 HAVING count(*) >= 40
ORDER BY pct_women DESC;

┌─────────────────┬────────────┬───────────┐
│    deal_type    │ n_gendered │ pct_women │
├─────────────────┼────────────┼───────────┤
│ sale            │        191 │      30.4 │
│ loan            │        144 │      28.5 │
│ contract        │       3708 │      23.0 │
│ receipt         │       2200 │      10.2 │
│ delivery        │         59 │       5.1 │
└─────────────────┴────────────┴───────────┘
  (abridged: the full result is 15 rows)

-- The monetization transition: gold's share of dated money facts, by century.
SELECT century, count(*) AS n,
       round(sum(system = 'gold') * 1.0 / count(*), 3) AS gold_share
FROM 'monetary.parquet'
WHERE century IS NOT NULL AND system IN ('gold','silver','bronze')
GROUP BY 1 ORDER BY 1;

---

Dataset Structure

How the tables join

Two provenance keys, both resolving to the source corpus:

• stem — the unique per-document key (DDbDP file stem). Used by the person-derived tables (documents, persons, principals).
• tm_id — the Trismegistos document id, used by the money-derived tables (monetary, prices, taxes). tm_id is not unique: 61,249 documents carry 60,862 distinct TM ids, so a tm_id join fans out (Pitfall 2).

documents ──stem──< persons ──stem+span──> principals ──folded──> persons_distinct
documents ──tm_id──< monetary ──filtered subset──> prices, taxes
persons   ──aggregated by century/region──> autonomy

prices and taxes are not separate extractions: they are monetary with precision filters applied, same grain, same columns plus two.

Referential integrity is exact: 0 of 195,906 monetary rows have a tm_id absent from documents; 0 of 21,895 principals rows have a stem absent from documents.

Character-span provenance

Every fact table carries a start/end offset pair into the canonical text view (edited_text) of its document: amount_start/amount_end in the money tables, person_start/person_end in the person tables. Offsets are 0-based and end-exclusive (Python slicing). DuckDB's substr is 1-based, so the slice is substr(edited_text, start + 1, end - start).

Column reference

"cov" is the percentage of non-null rows, measured on this build — it tells you what a filter on that column costs.

documents — the spine (61,249 rows)

One row per text document: metadata plus per-document counts folded in from every other table. Every document survives the fold, so 0 means "nothing extracted", not "no data".

column	type	cov	meaning
stem	VARCHAR	100%	unique document key (DDbDP file stem)
tm_id	VARCHAR	100%	Trismegistos id — not unique, 60,862 distinct
century	DOUBLE	97.1%	signed century from the HGV date (+2 = 2nd c. AD, −2 = 2nd c. BC)
place_pleiades	DOUBLE	74.1%	Pleiades id of the document's place
deal_type	VARCHAR	100%	primary genre; '?' when the corpus gives none (17,932 docs)
n_persons	BIGINT	100%	PERSON mentions in the document
n_women_mentions, n_men_mentions	BIGINT	100%	of those, by attributed gender
n_principals, n_women_principals	BIGINT	100%	principals, by gender
has_guardian_woman	BOOLEAN	100%	a woman with a μετὰ-/χωρὶς-κυρίου formula is present
n_money_facts	BIGINT	100%	monetary facts — folded by tm_id, shared by TM-siblings
has_price, has_tax	BOOLEAN	100%	a clean price / tax row exists — also tm_id-shared

monetary — the fact table (195,906 rows)

One row per monetary amount carrying a currency, with its normalized value and whatever the relation graph attaches: the commodity it prices, that commodity's quantity and unit, the tax it discharges.

column	type	cov	meaning
tm_id	VARCHAR	100%	document key (not unique)
amount_start, amount_end	BIGINT	100%	char span of the amount (0-based, end-exclusive)
amount_text	VARCHAR	100%	the Greek surface string of that span
value_num	DOUBLE	100%	the decoded numeral inside the span
currency_id	VARCHAR	100%	canonical denomination id (see vocabularies)
system	VARCHAR	100%	silver | gold | unknown — never aggregate across systems
value_base	DOUBLE	98.7%	value in drachmas (silver) or nomismata (gold)
commodity_id	VARCHAR	3.9%	the priced commodity, when a HAS_PRICE link exists
quantity	DOUBLE	3.4%	how much of it
unit_id	VARCHAR	0.3%	the measure (artaba, metretes, …)
unit_price_base	DOUBLE	3.3%	value_base / quantity — raw, over-divides (Pitfall 4)
tax_id	VARCHAR	3.4%	the tax discharged, when a CHARGED_UNDER link exists
confidence	DOUBLE	100%	labeler confidence — constant 0.82, carries no information
date_lo	DOUBLE	93.4%	earliest year of the HGV date range (signed; negative = BC)
date_hi	DOUBLE	78.3%	latest year of the range
date_mid	DOUBLE	94.3%	range midpoint — what century is derived from
century	DOUBLE	94.3%	signed century, range −4 … +10
bin50	DOUBLE	94.3%	start year of the 50-year bin, floored toward −∞ (−124 → −150)
place_pleiades	DOUBLE	80.7%	Pleiades id
genres	VARCHAR	100%	JSON array as a string, e.g. ["list","account"] (Pitfall 7)

Low commodity_id/tax_id coverage is expected: most amounts in a papyrus are bare sums (a receipt total, a wage, a rent) with nothing attached. The 3.9% that do carry a commodity are what the price series is made of.

prices — clean price observations (98 rows)

monetary restricted to genuine, comparable unit prices. The filters drop the value_num == quantity double-link artifact (48% of naive candidates), bronze chalkous, units that are not the commodity's own dry/liquid measure, and implausible quantity/price combinations. All monetary columns, plus:

column	type	cov	meaning
unit_price	DOUBLE	100%	drachmas per unit — the series value
commodity	VARCHAR	100%	wheat (70) | wine (14) | barley (11) | oil (3)

taxes — clean tax payments (592 rows)

monetary restricted to rows carrying a named tax. These are payments (often installments), not tax rates. Cleaner than prices because no per-unit division is involved. All monetary columns, plus:

column	type	cov	meaning
payment	DOUBLE	100%	the payment in drachmas
tax	VARCHAR	100%	laographia (539, the poll tax) | demosia (53, the land tax)

persons — gendered person mentions (350,206 rows)

Gender and guardian status for every PERSON span the NER model found. One row per mention, not per person.

column	type	cov	meaning
stem, tm_id	VARCHAR	100%	document keys
person_start, person_end	BIGINT	100%	char span of the mention
person_text	VARCHAR	100%	the full mention as written ("name son-of-father …")
head_text	VARCHAR	100%	the person's own name, split out of the blob
father_text	VARCHAR	36.8%	the patronymic, when present
gender	VARCHAR	100%	male (82,817) | female (22,901) | unknown (244,488)
gender_basis	VARCHAR	100%	which rule decided it (see vocabularies)
gender_confidence	DOUBLE	100%	0.0 (unknown) … 0.97 (guardian formula)
guardian	VARCHAR	100%	with (1,628) | without (143) | none (348,435)
date_mid, century, bin50	DOUBLE	95.9%	document date
place_pleiades	DOUBLE	81.7%	Pleiades id
genres	VARCHAR	100%	JSON array string

Only 30% of mentions are gender-attributable, by design: the rules fire only when a name or formula is decisive, and record which rule fired rather than guessing. Nothing is imputed.

principals — principals, gendered and deal-typed (21,895 rows)

The people a deal turns on: a PERSON the relation model links as PARTY_OF a transaction, or PAID_BY/PAID_TO an amount, joined to its gender/guardian/ patronymic from persons. Spans 11,002 documents. All the persons columns above (with father_text at 53.9% — principals are named formally in contracts), plus:

column	type	cov	meaning
roles	VARCHAR	100%	sorted, |-joined subset of party, payer, payee
transaction_term	VARCHAR	68.9%	the Greek verb/noun naming the deal (ὁμολογῶ, ἐμίσθωσεν, …)
deal_type	VARCHAR	100%	the document's primary genre; '?' for 3,514 rows
confidence	DOUBLE	100%	relation-model score, 0.34 … 1.00 (median 0.85) — informative, filter on it

persons_distinct — coreference-lite people (17,362 rows)

Principal mentions folded into distinct people by a conservative surface key (normalized name, normalized patronymic, place). Answers "how many distinct women", not "how many mentions". It under-merges — a person named without their father, or appearing in two nomes, splits into two rows — so the count is an upper bound, the safe direction for a "not fewer than" claim. This is not full prosopographical coreference.

column	type	cov	meaning
person_id	VARCHAR	100%	stable 16-hex hash of the identity key; unique, identical across rebuilds
head_text	VARCHAR	100%	representative name
father_text	VARCHAR	57.7%	representative patronymic
place_pleiades	DOUBLE	85.0%	representative place
gender	VARCHAR	100%	folded across mentions (attributed beats unknown, majority wins) — female 1,414 | male 5,608 | unknown 10,340
guardian	VARCHAR	100%	folded with without winning: one unambiguous χωρὶς-κυρίου attestation establishes she acted alone
n_mentions	BIGINT	100%	mentions folded into this person (max 32)
deal_types	VARCHAR	100%	|-joined set of deal types the person appears in
first_century	DOUBLE	98.5%	earliest century of attestation

autonomy — the published curve (32 rows)

A derived summary, not a fact table: the χωρὶς-κυρίου share of guardian-formula women, by century and by region. Shipped so the headline finding is inspectable. To re-slice by nome, deal type or a different time granularity, group principals yourself — that is the underlying table.

column	type	cov	meaning
dimension	VARCHAR	100%	century (8 rows) or region (24 rows)
bucket	VARCHAR	100%	the signed century as a string (e.g. "3.0") or the place name
n_with, n_without	BIGINT	100%	women attested with / without a guardian
n	BIGINT	100%	n_with + n_without
autonomous_share	DOUBLE	100%	n_without / n

Controlled vocabularies

Ids are canonical lexicon ids, not surface forms — the labeler resolves inflected Greek to these before anything is written. Counts are from monetary unless noted.

system · silver 140,040 (Ptolemaic–Roman drachma system, base unit the drachma) · gold 54,771 (Byzantine solidus system, base unit the nomisma) · unknown 1,095 (a money word with no fixed denomination).

currency_id · Silver ladder (1 talent = 6,000 drachmas; 1 drachma = 6 obols; 1 obol = 8 chalkoi): drachma 65,672 · obol 23,253 · talent 12,513 · diobol 7,317 · triobol 7,099 · chalkous 6,745 · hemiobelion 6,680 · tetrobol 6,346 · pentobol 3,422 · argyrion 993 (generic "silver money", no denomination → value_base is null). Gold (24 keratia = 1 nomisma): nomisma 36,168 · keration 18,047 · chrysion 556 (generic gold, including gold as metal, not only coin).

commodity_id · grain 2,098 · garden 1,066 · wheat 1,057 · wine 758 · oil 673 · barley 429 · donkey 395 · hay 212 · vegetables 161 · land 149 · house 148 · camel 139 · sheep 139 (+ a long tail).

unit_id · artaba 399 (dry measure) · aroura 59 (land) · metretes 34 and keramion 28 (liquid) · xestes 26 · kotyle 19 · myriad 18 · choinix 10 · litra 10 · naubion 7 · pechys 6.

tax_id · prosdiagraphomena 2,788 (Roman surcharge) · demosia 1,722 (land tax) · phoros 700 · laographia 574 (poll tax) · merismos 349 · phylakitikon 254 (Ptolemaic police tax) · telesma 110 · stephanikon 40 · genema 36 · syntaxis 17. A small tail of ids in this slot are not taxes — drachma 20, obol 3, hemiobelion 2, chalkous, chrysion, aroura, year, time — 8 ids / 33 dated facts (0.5%), a measured contamination rate. Filter to the named taxes above rather than taking every non-null tax_id.

deal_type (in documents) · ? 17,932 · receipt 15,194 · contract 5,117 · list 4,193 · letter_private 3,603 · account 3,342 · mummy_label 2,212 · order 2,146 · petition 1,908 · letter 1,724 · letter_official 1,228 · declaration 734 · register 643 · delivery 417 · sale 318 (+ loan, lease, a tail). ? means "the corpus records no genre", not "other" — exclude it, don't bucket it.

gender_basis (in persons), in precision order · guardian 1,770 (a κύριος formula, conf 0.97) · nomen 21,864 (Αὐρήλιος / Αὐρηλία, 0.9) · kin 5,781 (θυγάτηρ / υἱός, 0.9) · gazetteer 31,153 (attested name list, 0.8) · egypt_prefix 45,122 (Egyptian Τα- female / Πα- male, 0.72) · ethnic 28 · none 244,488 (no rule fired → unknown).

guardian · with — a μετὰ κυρίου formula (she transacts under a guardian) · without — χωρὶς κυρίου (she transacts alone) · none — no formula in the window.

roles (in principals) · party 13,738 · payee 3,605 · payer 3,176 · party|payee 1,111 · party|payer 227 · payee|payer 32 · party|payee|payer 6. Multi-valued: one person can be both a party to the contract and the payer of its price. Match with LIKE '%payer%' or list_contains(str_split(roles,'|'),'payer'), never with =.

---

Dataset Creation

Curation Rationale

Economic historians extract these facts by hand, one papyrus at a time. That limits every quantitative claim about the ancient economy to the sample one scholar can read. The goal here was a database where each fact is (a) derived mechanically from a named revision of a public corpus and (b) traceable to the exact characters it came from, so a reader can disagree with any individual row.

Source Data

DDbDP EpiDoc XML at a pinned git revision, parsed into a dual-view text representation (a diplomatic view and an edited view) at a parse rate of 1.000 over all 67,980 documents; 61,249 carry real text. HGV supplies dates (95–98% coverage) and places (74–76%, linked to Pleiades authority ids); EpiDoc <num> elements supply decoded numeral values. So dates, places and numerals are given by the corpus, not re-extracted — the database normalizes and assembles them.

Annotations

The papyri carry no entity or relation markup upstream (0% over a 200-document audit). All supervision was built for this project:

1. A deterministic lexicon + rules labeler (mined lexicons: 132 entries / 545 attested surface forms, 0 unattested) produced silver labels over ~49k training documents.
2. 115 documents form the reference set — 2,995 entities, 710 relations — drafted by one language model and re-checked by a second. Mechanically validated (offsets, byte-identity, numeral coverage, schema legality) but not adjudicated by a papyrologist; the maintainers do not read Ancient Greek. Model scores are agreement with this reference, not expert accuracy.
3. Models were silver-pretrained and gold fine-tuned, then run over the whole corpus (1,368,079 entities; 228,945 relations).
4. The money layer runs on the rules, not the models, because closed-class vocabulary is matched at ceiling by a gazetteer.

Personal and Sensitive Information

The named individuals are inhabitants of Greco-Roman Egypt who died between roughly 1,300 and 2,300 years ago, named in documents published openly by papyrologists for over a century. There is no living-person privacy concern.

Gender attribution is inferred from onomastic and formulaic evidence, and is recorded as an inference with its basis (gender_basis) and confidence, never as ground truth about an individual. It is fit for aggregate history, not for a claim about any one person.

---

What has been found with it

Five results, strongest first. Full write-up in the project repository; every number below was recomputed from these shipped tables.

1. The monetization transition, recovered unsupervised. Gold's share of dated money facts sits at ~0.00 for eleven centuries, then 4th c. AD 0.155 → 5th c. 0.931 → 8th c. 1.000 (n = 195,906). That is the textbook silver-to-solidus transition, reproduced without being told about it. Restricted to coined gold (nomisma/keration), pre-4th-century attestations number 22 — the earlier residue is chrysion, gold as metal.

2. The fiscal-regime map periodizes itself. 6,441 dated tax facts across 18 tax ids: laographia (the Roman poll tax) is 560 of 569 in the 1st–3rd c. AD, zero after, zero in the Ptolemaic centuries; prosdiagraphomena is 99.9% Roman; phylakitikon is 73% Ptolemaic; demosia is 1,596 of 1,674 in the 6th–8th c. Poll-tax medians (~4 dr) read as installments, with a p90 of 16–39 dr matching the known annual rate.

3. Women's legal autonomy rises sharply (novel). The χωρὶς-κυρίου ("without a guardian") share of guardian-formula women: 0% up to the 1st c. AD (n=646, zero autonomous) → 1.2% (2nd) → 38.8% (3rd) → 80% (4th, n=35). Validated against gold: the over-count is on the μετὰ side, so the rise is conservative. Quote it in tiers, not decimals — the 4th-century cell is directional.

4. Women as principals, by type of deal (novel). 21,895 principals; women are 18.0% of gendered mentions and 20.1% of distinct people. The finding is the gradient: sale 30.4% · loan 28.5% · contract 23.0% vs receipt 10.2% · delivery 5.1%. Dropping the female-only guardian channel gives a 13.0% floor and the ordering survives (Spearman ρ 0.856).

5. Prices (weakest — flagged as such). 98 clean observations. Only 2nd c. AD wheat, 13.33 dr/artaba [IQR 6.0–27.5], n=37 is defensible; the literature says ~7–12, which the IQR brackets. See Limitations.

---

Bias, Risks, and Limitations

Limitations

• prices is 98 observations, not a millennium-long series. Only 2nd c. AD wheat (n=37) has enough observations to defend; the 3rd century has 9 and the rest fewer. Two known defects remain: unit_price = value / quantity over-divides where the recorded amount is already per-unit, and the non-wheat commodities are too sparse to use (some wine rows are unit errors, not prices). Treat this table as a validated method on a thin sample, not as a price history.
• Two different error regimes, never to be blended. Rules (high precision, systematic misses) for money; models (entity F1 0.737 strict, PARTY_OF 0.623 end-to-end) for people. Do not put one error bar across both.
• Payment direction is absent by choice. There are no paid_by/paid_to columns. The relation model scores PAID_BY at F1 0.145, so direction is deliberately missing rather than present and wrong. principals records that someone is party to a deal, not which side of the payment they stand on.
• 58% of principals have no attributable gender, excluded from gendered shares and never imputed. They are not damaged text — 100% carry a parsed head name; the exclusion reflects closed-vocabulary coverage of the gender rules. One rule (the guardian formula) can only ever return female, so women's share is 18.0% including it and 13.0% without it — use 13.0% as the conservative floor. The deal-type ordering is stable either way.
• Mentions are not people. persons and principals are mention-grain; persons_distinct is the only head-count table, and it under-merges.
• Survival bias is not correctable. Everything counts surviving, published, digitized papyri, skewed toward the Arsinoite nome and dry sites. Shares within a bucket are interpretable; raw counts across buckets are not.
• Dates are HGV's, assigned to a century by range midpoint, which smears sharp transitions across a boundary.
• Regional labels are findspot-based, and a papyrus can be written in one nome and found in another.

Pitfalls — read before publishing a number

1. Never aggregate value_base across system. Silver drachmas and Byzantine gold nomismata are different metals six centuries apart and are not convertible. Always WHERE system = 'silver', or GROUP BY system.
2. tm_id is not unique; a tm_id join fans out. 61,249 documents share 60,862 TM ids. In documents, n_money_facts/has_price/has_tax are TM-level facts shared across siblings, not document-level ones.
3. Mentions ≠ people. Women are 18.0% of principal mentions and 20.1% of distinct gendered principals — different denominators, both correct. Say which.
4. monetary.unit_price_base over-divides and must not be used for a published series. Use prices.unit_price, which has the filters applied.
5. confidence means two different things. In monetary/prices/taxes it is the rule labeler's constant 0.82 and carries no information — filtering on it does nothing. In principals it is the relation model's score (0.34–1.00, median 0.85) and is a usable precision knob.
6. Trust the contrasts, not the absolute totals. Extraction error is real, so the robust claims are relative ones — across deal types, centuries, regions — where error is roughly common-mode.
7. genres is a JSON string, not a list: query it with list_contains(cast(json(genres) AS VARCHAR[]), 'receipt'), or just use the already-resolved deal_type.
8. century skips year zero, and '?' is not a category. +1 covers 1–100 AD, −1 covers 1–100 BC, so abs() merges a BC and an AD century. bin50 floors toward −∞ and is safe to sort across the BC/AD line.

Recommendations

Report the denominator you used (mentions vs people, gendered vs all). Prefer within-bucket shares to cross-bucket counts. When a claim rests on a small cell, give n — several buckets here are under 50 rows. And when a single row matters, follow its span back to the papyrus and read it.

---

Reproducibility

Re-running the pipeline at the same corpus_rev reproduces every table bit-for-bit: the assembly stages are deterministic and the two model outputs are frozen artifacts. manifest.json in the repo records the revision, schema version, licence and per-table inventory, so the export is self-describing.

oik db build --sample 0   # monetary facts  → monetary.parquet
oik db prices             # price series    → prices.parquet
oik db taxes              # tax payments    → taxes.parquet
oik db persons            # gendered people → persons.parquet
oik db autonomy           # the curve       → autonomy.parquet
oik db principals         # principals      → principals.parquet
oik db export             # spine + people  → export/

---

Citation

Source texts are derived from the Duke Databank of Documentary Papyri (DDbDP) and the Heidelberger Gesamtverzeichnis (HGV) under CC BY 3.0, which requires attribution. Please cite both the corpus and this dataset.

@dataset{oikonomia_db_2026,
  author    = {Ainouche, Abderahmane},
  title     = {{OIKONOMIA-DB}: An Auditable Economic Database of Greco-Roman Egypt},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/ainouche-abderahmane/oikonomia-db}
}

Duke Databank of Documentary Papyri (DDbDP), Duke Collaboratory for Classics Computing (DC3) and papyri.info, CC BY 3.0.

Dataset Card Contact

Issues and corrections: https://github.com/abderahmane-ai/oikonomia/issues