OIKONOMIA-Grammateus — Entity Recognition for Greek Documentary Papyri

γραμματεύς grammateús, "the scribe" — the village clerk of Greco-Roman Egypt, whose job was to write down who, where, and how much.

A token-classification model that reads the everyday economic language of the documentary papyri of Greco-Roman Egypt — tax receipts, leases, loans, wages, sales, census returns — and tags the entities an economic historian extracts by hand: people, places, money, commodities, units, taxes, dates, ages, occupations.

It is the entity arm of OIKONOMIA, a project turning the ~68,000 Duke Databank (DDbDP) papyri into a structured, auditable database of ancient economic life. Run over the whole corpus it produced 1,368,079 entities across 61,249 documents, which became the published OIKONOMIA-DB.

Model Details

Model Description

• Developed by: Abderahmane Ainouche (OIKONOMIA project)
• Model type: RoBERTa-base encoder + token-classification head (BIO tagging)
• Language: Ancient Greek (grc) — documentary, not literary
• Licence: apache-2.0, inherited from the backbone
• Base model: bowphs/GreBerta + domain-adaptive pretraining (DAPT) on the papyri corpus
• Parameters: 125.4M · Context: 512 tokens · Vocabulary: 52k, case-preserving
• Labels: 15 entity types → 31 BIO tags

Case preservation matters here and drove the backbone choice: capitalisation is the strongest single cue for PERSON and PLACE in this corpus, and keeping it costs only ~0.6% more tokens.

Model Sources

• Repository: https://github.com/abderahmane-ai/oikonomia
• Sibling model: OIKONOMIA-Homologia (relations)
• Dataset produced with it: OIKONOMIA-DB

Uses

Direct Use

Information extraction over documentary Ancient Greek: tagging the entities in a papyrus so they can be normalized, linked and counted. Suited to papyrology, digital humanities, and ancient economic history.

Downstream Use

Designed to feed OIKONOMIA-Homologia, which links the tagged spans into transactions. Run Grammateus first, pass its spans to Homologia. Also a reasonable starting point for fine-tuning on other documentary Greek annotation schemes.

Out-of-Scope Use

• Literary Greek, epigraphy, or Modern Greek. The domain adaptation is specifically to documentary papyri; performance elsewhere is unmeasured.
• As a normalizer. It finds spans. Mapping surface forms to canonical currency/commodity/unit ids is a separate lexicon step.
• As a person disambiguator. A PERSON span is the name as written, often name + patronymic together; it does not resolve who that individual is.

How to Get Started

from transformers import AutoTokenizer, AutoModelForTokenClassification
import torch

repo = "ainouche-abderahmane/grammateus"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForTokenClassification.from_pretrained(repo).eval()

text = "πυροῦ ἀρτάβας δύο δραχμῶν ἑκατόν"  # "two artabas of wheat, a hundred drachmas"
enc = tok(text, return_tensors="pt", return_offsets_mapping=True)
offsets = enc.pop("offset_mapping")[0]
with torch.no_grad():
    pred = model(**enc).logits[0].argmax(-1)

for (s, e), p in zip(offsets.tolist(), pred.tolist()):
    if e > s:
        print(text[s:e], model.config.id2label[p])

Documents longer than 512 tokens must be windowed; the corpus run used a strided window with span-offset remapping (oikonomia.ner.inference).

Entity labels (15)

Training Details

Training Data

• Corpus: the Duke Databank of Documentary Papyri via papyri/idp.data, CC BY 3.0, pinned to revision d7a34f302d1e44e271256092c2b780733187b478. Not available as a Hub dataset; it is EpiDoc XML in a git repository.
• The papyri carry no entity markup upstream (0% over a 200-document audit). All supervision was built for this project:
  • Silver: a deterministic lexicon + rules labeler over 48,891 training documents (measured against gold at micro F1 0.667).
  • Reference set: 115 documents — 2,995 entities. Model-drafted and model-re-checked, not expert-validated; see the provenance note below.
• Splits are leak-free: near-duplicate clusters (2.89% of the corpus) and TM-sibling documents are grouped before assignment, so no document's near-twin sits across the split boundary.

Training Procedure

Three stages: DAPT, then silver pretraining, then gold fine-tuning.

Stage 0 — domain-adaptive pretraining. Full fine-tune of GreBerta on the papyri corpus (8.25M train / 1.10M dev tokens), masked-LM objective. Full-FT beat head-only and adapter variants (dev perplexity 4.54). This checkpoint is the backbone for everything below.

Stage 1 — silver pretraining. Token classification over the 48,891 silver-labelled documents.

Stage 2 — gold fine-tuning. The same model fine-tuned on the 115 gold documents.

Training Hyperparameters

A generalized-cross-entropy loss (GCE, q=0.7) was tried on the silver stage to absorb label noise and rejected — it cost 5.7 strict F1.

Compute

Single NVIDIA A10 (24 GB) on Modal. Emissions were not tracked; a run of this size is on the order of GPU-hours, not GPU-days.

Evaluation

Testing Data, Factors & Metrics

Testing data. 5-fold cross-validation over the 115-document reference set (2,995 entities). No separate held-out test set exists — the set is small enough that spending part of it on a test split would make every number noisier than cross-validation does.

Note on provenance. This reference set was drafted by one language model and re-checked span by span by a second, different one. It is mechanically constrained (offsets computed not typed, text byte-identical to the corpus, every numeral either labelled or explicitly skipped with a reason, every relation schema-legal) but no papyrologist or other domain expert has adjudicated it, and the maintainers do not read Ancient Greek. Scores below are therefore agreement with this reference, not accuracy against expert ground truth. Independent expert annotation is the top outstanding item.

Metrics. Micro-averaged span F1, in two regimes: strict requires an exact span and label match; relaxed credits an overlapping span of the right label. Relaxed is the fairer read for a downstream normalizer that only needs to find the right token region.

Results

Per-label strict F1:

Three labels are missing, and that is a gap in the record, not a suppression. The cross-validation run logged per-label F1 for 12 of the 15 labels; COMMODITY, PERSON_ROLE and TAX_TERM were not captured, and re-deriving them honestly requires re-running the 5-fold CV rather than scoring the shipped checkpoint (which was trained on all 115 gold documents, so scoring it on them would be train-on-test). Expect them to sit at the low end: PERSON_ROLE and TAX_TERM are the two labels named below as consistency-bound.

Precision and recall were not recorded separately for the entity model; only micro F1 at both strictness levels was logged. The sibling relation model reports full P/R.

Domain-adaptive pretraining is what pays

Against an identical fine-tune on the raw backbone with no papyri DAPT (a paired run on the same folds):

(This comparison was run at an earlier stage on a 65-document gold set, so the absolute values are lower than the headline 0.737; the paired contrast is the result.) Gains concentrate exactly where predicted — the open-class onomastic labels — and no label regressed. Lexicon-reachable labels barely move (CURRENCY 0.78 → 0.79), which is the expected shape: DAPT buys you the words a gazetteer cannot enumerate.

How much to trust these numbers

115 documents over 5 folds is roughly 23 documents per fold, so every figure above is a point estimate on a small sample. The micro totals rest on 2,995 entities and are the sturdiest. Per-label figures for rare labels rest on far fewer instances and should be read as indicative of rank, not as precise values — ten more annotated documents could move them. Per-fold variance was not recorded and is a known gap. The headline comparisons (DAPT vs no-DAPT, silver-only vs silver→gold) are paired runs on identical folds, which is what makes their direction reliable at this n.

Bias, Risks, and Limitations

• Label ceilings are annotation-consistency-bound, not data-bound. TRANSACTION, PERSON_ROLE and TAX_TERM are open, formulaic classes where the limit is how consistently a human can draw the span, not how much data the model sees. More annotation of the same kind will not lift them much.
• Fragmentary text degrades accuracy. Papyri carry lacunae, editorial brackets and supplied readings; heavily broken passages are harder for the model in the same way they are for a reader.
• Onomastic bias. PERSON recognition leans on Greek and Egyptian name shapes attested in the corpus. Rare, foreign or heavily abbreviated names are recognised less reliably, which means the entities this model finds are not a uniform sample of the people in the documents.
• Downstream gender inference is not the model's output. OIKONOMIA-DB infers gender from names and formulae after this model tags a span. Those inferences are aggregate-level and rule-based, not a claim by this model about an individual — see the dataset card.
• Collapsed persons. A PERSON span is the name as written, frequently name + patronymic in one span; splitting individuals for kinship is a downstream step.
• Corpus bias is inherited. The training corpus is surviving, published, digitized papyri — skewed toward the Arsinoite nome and dry sites.

Recommendations

Report relaxed F1 alongside strict when your downstream step only needs the right token region. For quantitative history, prefer relative contrasts (across periods, regions, document types) to absolute totals, since extraction error is roughly common-mode across buckets but does not cancel in a total.

Licence & lineage

Released under apache-2.0, inherited from bowphs/GreBerta (apache-2.0). The training corpus (DDbDP) is CC BY 3.0 and must be attributed. No ancestor carries a NonCommercial term — deliberately: the release is gated in code by a licence firewall that refuses any artifact of NonCommercial or unverified lineage (oikonomia.models.licensing; audit trail in MODEL_LICENSES.md).

Citation

@misc{oikonomia_grammateus_2026,
  title  = {{OIKONOMIA-Grammateus}: Entity Recognition for Greek Documentary Papyri},
  author = {Ainouche, Abderahmane},
  year   = {2026},
  url    = {https://huggingface.co/ainouche-abderahmane/grammateus},
  note   = {Base model bowphs/GreBerta; trained on the Duke Databank (DDbDP, CC BY 3.0)}
}

@inproceedings{riemenschneider-frank-2023-exploring,
  title     = {Exploring Large Language Models for Classical Philology},
  author    = {Riemenschneider, Frederick and Frank, Anette},
  booktitle = {Proceedings of ACL 2023},
  year      = {2023},
  url       = {https://arxiv.org/abs/2305.13698}
}

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

Model Card Contact

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