Sphragis

Sphragis (σφραγίς, "sigil") is a benchmark for Ancient Greek (grc) prose and verse authorship attribution (AA). Its input is the complete curated union of the human-annotated CoNLL-U trees in the supported treebank projects, including lemma, part of speech, morphology, and dependency relations. It defines 1-, 5-, 10-, and 50-sentence attribution tasks. The complementary scanned-line benchmark is published separately as Urdatorn/sphragis-metre.

For the archaic Greek poets, "sphragis" (or "sphregis" in some dialects) was a term of art roughly equivalent to the modern "authorial watermark", referring to text composition that intentionally aims at making it harder for adversaries to falsely pass off the work as their own. The phenomenon of "sphragizing" is one of the earliest chapters in the history of AA. (See Theognidea, 19-30)

Sphragis does not require a sentence to occur in any second annotation stream. Prose and verse enter on the same basis: a retained sentence from any supported treebank is eligible. Exact duplicates are consolidated with all treebank provenance preserved.

Tasks

There are four task sizes on each of two tracks. Prose and verse sentences share one closed-set authorship problem; the genre column preserves their source representation.

The main track (sentence_*) splits rows randomly within author and work. The source-held-out track (source_held_out_*) splits them by annotation project: each author's validation and test sentences come from a treebank that supplies none of that author's training sentences. See Source-held-out track for why this exists and Model-facing representation for what was done to the main track in consequence.

Configurations

Configuration	Authors	Row unit	Train rows	Validation rows	Test rows
sentence_1	28	sentence	50,500	5,650	5,650
sentence_5	28	5-sentence chunk	10,100	1,130	1,130
sentence_10	28	10-sentence chunk	5,050	565	565
sentence_50	28	50-sentence chunk	1,010	113	113
source_held_out_1	12	sentence	28,500	4,550	4,550
source_held_out_5	12	5-sentence chunk	5,700	910	910
source_held_out_10	12	10-sentence chunk	2,850	455	455
source_held_out_50	12	50-sentence chunk	570	91	91

The source_held_out_* configurations use the same rows, the same chunk sizes and the same 100-row bottleneck, over the eleven authors that more than one project annotates. Two of them, Hesiod and Lucian, do not reach the main track's threshold under random splitting but do under this arrangement.

The suffix gives the chunk size in every split. _50, the largest, defines one shared source corpus for all four tasks: an author is retained only when that author independently has at least 50 source rows in both validation and test, and every split is reduced to a multiple of 50 source rows per author. Every smaller task size divides 50, which is what lets the same retained rows fill all four. _1 publishes the shared rows atomically, while _5, _10 and _50 regroup those exact same train, validation, and test rows into fixed benchmark chunks. Thus author inventory and total source text are identical across task sizes.

Source-row selection uses seed 776. For each retained author and split, the unavoidable n mod 50 source rows are selected for exclusion once by a stable hash. The shared retained rows are put into their best available textual order within each work for each chunking task. Complete single-work chunks are emitted first; only residual work tails are combined, and such rows are marked by chunk_is_mixed_work. Chunk IDs, boundaries and exclusions are therefore reproducible. The complete manifest is in metadata/dataset_variants.json.

Every row retains its treebank genre and provenance.

Sentences per author label

These are the exact retained atomic sentences in sentence_1. The same sentences are represented in the 5-, 10-, and 50-sentence configurations.

Author label	Train	Validation	Test	Total
Aeschines	500	50	50	600 (1.0%)
Aeschylus	2,600	300	300	3,200 (5.2%)
Antiphon	400	50	50	500 (0.8%)
Appian	750	50	50	850 (1.4%)
Aristophanes	1,450	150	150	1,750 (2.8%)
Aristotle	650	50	50	750 (1.2%)
Athenaeus	1,950	200	200	2,350 (3.8%)
Demosthenes	2,400	300	300	3,000 (4.9%)
Diodorus Siculus	1,000	100	100	1,200 (1.9%)
Dionysius of Halicarnassus	850	100	100	1,050 (1.7%)
Euripides	600	50	50	700 (1.1%)
Herodotus	5,900	700	700	7,300 (11.8%)
Hesiod	700	50	50	800 (1.3%)
Homeric-Iliad	6,250	750	750	7,750 (12.5%)
Homeric-Odyssey	4,850	600	600	6,050 (9.8%)
Josephus	800	100	100	1,000 (1.6%)
Lucian	550	50	50	650 (1.1%)
Lysias	750	50	50	850 (1.4%)
Menander	700	50	50	800 (1.3%)
Paul	700	50	50	800 (1.3%)
Plato	900	100	100	1,100 (1.8%)
Plutarch	1,300	150	150	1,600 (2.6%)
Polybius	3,000	350	350	3,700 (6.0%)
Procopius	950	100	100	1,150 (1.9%)
Sextus Empiricus	500	50	50	600 (1.0%)
Sophocles	3,100	350	350	3,800 (6.1%)
Thucydides	950	100	100	1,150 (1.9%)
Xenophon	5,450	650	650	6,750 (10.9%)
All labels	50,500	5,650	5,650	61,800 (100.0%)

from datasets import load_dataset

sentences = load_dataset("Urdatorn/sphragis", "sentence_10")

Splits

Splits use random seed 776 and begin with independent stratification within each author/work group: approximately 80% train, 10% validation and 10% test by row. Groups with 3–9 rows receive one validation and one test row; groups with fewer than three stay in training. The 50-row eligibility and remainder bottleneck is then applied to all four task sizes. The train, validation, and test source sets are all exact multiples of 50 per author. This makes the published source-row ratio slightly different from the initial 80/10/10 split. A stricter work-held-out evaluation can be added as a later track.

Source-held-out track

Sphragis merges seven annotation projects, and those projects do not overlap in authors. UD PROIEL contributes Herodotus and nobody else; Pedalion contributes Aristophanes and Procopius; the Odyssey is entirely AGDT while most of the Iliad is UD Perseus. Any convention that distinguishes the projects therefore also distinguishes the authors, and a model can score well on the main track by recognizing the annotator instead of the writer.

The source_held_out_* configurations remove that possibility by construction. For each author, the second-largest source project is held out entirely: its sentences become that author's validation and test data, split evenly, and every other project supplies training. An author is retained only when the held-out project yields at least 50 validation and 50 test rows and at least 50 training rows remain. A model that has learned to recognize annotators is actively penalized here, because every author arrives at evaluation in a scheme it was never trained in.

Which project is held out per author, and the row counts that follow, are recorded under source_held_out_track in metadata/build_report.json. The held_out_source column repeats it on every row.

This track is the benchmark's answer to the confound. The normalization described below removes the parts of the leak that are indefensible on their own terms, but it cannot remove all of it: projects genuinely disagree about tree shape, and no relabelling can normalize that away. Scores on the two tracks are not comparable, and the gap between them is the interesting number.

Principal columns

All configurations contain:

• id, author, work, work_id, genre, text;
• conllu, cts_urn, passage, treebank_source;
• native_syntax: one list per unit of audit records — sentence, token, source_token, relation, head, head_repair — for every token whose source project used a non-UD scheme;
• source_records: every contributing upstream record, with revision, URL, license, annotation provenance and syntax scheme;
• licenses, dedup_key, and split.

The source_held_out_* configurations add held_out_source.

Column structure

Every structured column is a real nested Arrow type rather than JSON packed into a string, so load_dataset hands back Python lists and dictionaries with nothing left to parse.

text is a list of strings, one per sentence. Atomic _1 rows hold one; _10, _50, and _100 hold exactly 10, 50, or 100. No whitespace separator encodes unit boundaries.

conllu has the same number of units as text, and each unit is a list of sentences, each sentence a list of token objects:

row["conllu"][0][0][2]
{"id": 3, "form": "βῆ", "lemma": "βαίνω", "upos": "VERB",
 "feats": "Aspect=Perf|Mood=Ind|Number=Sing|Person=3|Tense=Past|VerbForm=Fin|Voice=Act",
 "head": 0, "deprel": "root"}

The field names are the CoNLL-U column names. feats keeps the canonical Case=Nom|Gender=Masc spelling and is null when empty. deps, misc and xpos are not published: the first two were empty on every token, and the nine-position tag is folded into feats — see One morphological analysis.

The extra sentence level exists because one Urdatorn/sphragis-metre line can span a sentence boundary. Every Sphragis sentence unit holds exactly one sentence, so row["conllu"][i][0] is the token list for unit i.

To get canonical CoNLL-U text back:

from scripts.conllu_units import render_conllu
print(render_conllu(row["conllu"][0]))

The build validates every published unit by rendering it and parsing the result with the official CoNLL 2018 shared-task loader, so the structure is provably still CoNLL-U.

The _10, _50, and _100 configurations additionally contain chunk_size, chunk_target_size, ordered constituent_ids, constituent_provenance, chunk_work_ids, chunk_works, and chunk_is_mixed_work. Every row has chunk_size equal to the configuration suffix. Aggregated text and conllu follow constituent order, while constituent provenance retains the original passage-level locations.

All sentence rows share one schema. There are no metrical, scanned-line, or external-alignment columns. genre has exactly two values: prose and verse.

Model-facing representation

Every published conllu document and every text unit is reduced to one scheme-neutral representation, so that no category boundary the source projects disagree about can be used as an authorship feature.

No comments. CoNLL-U documents contain token rows only. The # text comment was removed in favour of the dedicated text column, which already carried the same string; comments carrying sent_id, source, cts, passage or newdoc id had already been stripped.

No punctuation. Editorial punctuation tokens are removed from the trees and re-indexed, and any dependent hanging from one is re-attached to its nearest surviving ancestor. Punctuation is likewise removed from text. UD PROIEL does not tokenize punctuation at all, so its absence alone identified its single author.

Lowercase throughout. Word forms, lemmas and text are lowercased with grc_utils.lower_grc, which handles the precomposed polytonic letters correctly. Capitalization marked sentence openings and proper nouns inconsistently between projects and editions.

Empty MISC. The MISC column is _ on every token. NativeRel, NativeHead and HeadRepair sat on every token of the natively converted projects and on none of the UD ones, and SpaceAfter occurred in UD Perseus alone. They now live in the native_syntax column, which records the source relation, source head and any head repair per token, keyed by both the published token index (i) and the source document's own numbering (src).

Shared inventories. UPOS is derived from the normalized nine-position tag, so it expresses one scheme rather than each project's own, and the categories the projects disagree about are merged:

Column	Published inventory	Merged because
UPOS	NOUN VERB ADJ PRON ADV ADP NUM INTJ X	PROPN is PROIEL-only; AUX is UD-only; δέ is PART in AGDT, ADV in Gorman and CCONJ in UD Perseus; the article is DET in UD and a pronoun in the AGDT lineage
DEPREL	root nsubj obj ccomp obl nmod advmod case conj dep	cc, cop and aux are never produced by the native conversions; iobj is UD Perseus only; subtypes such as nsubj:pass and obl:arg are PROIEL only; the two conversions of the same AGDT trees split attribution between amod, nmod and det. appos, discourse, dislocated and vocative were merged away after measurement — see What annotator variation remains
FEATS	Aspect Case Degree Gender Mood Number Person Tense VerbForm Voice	PronType, Definite, Polarity, Poss and Reflex occur in one project only. Aspect and VerbForm are published but never inherited: the build derives them from the normalized tag, so they are uniform across projects

One morphological analysis

The nine-position AGDT/Perseus tag is no longer published beside feats. Keeping both invited them to drift, and they already had: each carried distinctions the other lost, and they disagreed with each other on roughly 136,000 tokens.

UD's Tense cannot express the Greek tense system. Aorist, imperfect, perfect and pluperfect all read as Tense=Past, so the positional tag was the only record of which one a verb was. It was likewise the only record that a verb was non-finite: 20,798 infinitives and participles carried no VerbForm at all, while others carried a non-standard Mood=Inf. And it alone separated the medio-passive from the true middle, which feats flattened onto Voice=Mid. Meanwhile feats was the finer of the two on ambiguous genders such as Masc,Neut, which the tag could not represent.

The build now folds the tag into feats and drops it. All seven Greek tenses stay distinct, in standard UD values:

Greek tense	published feats
present	Tense=Pres|Aspect=Imp
imperfect	Tense=Past|Aspect=Imp
aorist	Tense=Past|Aspect=Perf
perfect	Tense=Pres|Aspect=Perf
pluperfect	Tense=Pqp|Aspect=Perf
future	Tense=Fut
future perfect	Tense=Fut|Aspect=Perf

VerbForm is Fin, Inf, Part or Ger, and Mood is restricted to the four moods a finite verb can carry. Voice distinguishes Mid,Pass from Mid. Gender keeps its feats value, which is the finer of the two.

Aspect and VerbForm were previously excluded from the shared inventory because only the UD sources supplied them, so their presence identified the project. Deriving them from the normalized tag removes that objection: every token now gets them on the same terms.

Native AGDT/Arethusa trees are serialized as CoNLL-U while preserving heads. Their dependency relations are conservatively mapped to Universal Dependencies v2 before the shared inventory is applied; opaque source-specific categories fall back to dep. Repairs required after removing artificial ellipsis nodes—or to break an upstream beta self-loop/cycle—are recorded as repair in native_syntax.

Every distinct published conllu document is loaded during testing with the official CoNLL 2018 shared-task load_conllu parser, and separately checked against every invariant above. This checks sentence termination, ten-column structure, token IDs, heads, roots and dependency cycles across all configurations, and that the CoNLL-U forms reconstruct the text column exactly.

What annotator variation remains

Twelve authors are annotated by more than one project. Holding the author fixed and changing only the annotator isolates convention from style, because the Greek is the same; changing the author inside one project gives the stylistic baseline to compare against. Categories are published only when they survive that test.

Four were removed because they failed it:

removed	annotator variation	authorial variation
vocative	0.36	0.00
Gender=Com	0.53	0.02
appos	0.43	0.04
discourse, dislocated	6.66	4.70

vocative is the clearest case: it did not vary between sixteen authors of one project at all, so it carried no authorial signal, and Case=Voc identifies vocatives exactly. discourse is the largest: UD Perseus marks it on 0.19% of tokens where the AGDT lineage marks 8–10%, but its advmod rate is higher by almost exactly as much. Merging the two drops their combined annotator variation from 6.66 and 6.17 to 1.68.

Two categories fail the test and are published anyway, because no relabelling reaches them:

• case is 6.6–8.9% of tokens in the natively converted projects and 13.9–17.2% in the UD ones. The UD conversions create copula, auxiliary and coordinator tokens that the AGDT lineage never produces at all, so the residual is which tokens exist rather than what they are called. Merging case with ccomp moves its variation only from 5.14 to 4.15, and adding root makes it worse.
• conj is 13.4–19.3% against 4.8–4.9%. Coordination is attached differently, which is a difference in tree shape.

Both are annotator-dependent by construction, and are the reason the source_held_out_* track exists.

Residual leakage

The normalization above removes the deterministic tells. It does not remove the confound, and it was never going to: the projects disagree about tree shape itself, not only about labels. AGDT-lineage trees mark conj on roughly 14% of tokens where UD Perseus marks 4% of the same Iliad, and no relabelling normalizes a genuine analytic disagreement.

scripts/audit_leakage.py measures what remains by fitting classifiers that see annotation markup and no Greek at all, and writes metadata/leakage_report.json. Those numbers are the floor a real model must clear: a system that does not beat them has demonstrated source identification rather than authorship attribution. They are published on the leaderboard for that reason, and the source_held_out_* track exists because they cannot be driven to chance.

Deduplication

Raw treebank input comprised 110,849 prose and verse sentences. Authorship curation retained 92,141 source sentences before deduplication. Exact normalization-based deduplication then removed 23,282 repeated rows, and 1,291 rows in 253 exact-text groups were omitted because their retained author labels conflict, leaving 67,568 eligible sentences before the shared 100-row task bottleneck. Normalization lowercases, maps final sigma, removes diacritics and retains letters only. The chosen parse follows a deterministic source priority, while all exact-source provenance is merged into source_records.

This is aggressive exact deduplication, not passage-level fuzzy deduplication. Segmentation or edition differences can therefore leave near duplicates for future review.

Sources and licensing

The corpus draws on AGDT 2.1, UD Perseus, UD PROIEL, UD PTNK, Gorman's Greek Dependency Trees, Pedalion, and Harrington Trees.

This is a mixed-license dataset, including NonCommercial and ShareAlike data. Consult LICENSES.md and the licenses recorded on every row. Source commits for this release are frozen in metadata/source_revisions.json.

Conservative authorship policy

The benchmark favors reliable training labels over coverage. It excludes anonymous and unknown authors, Pseudo-* and traditional attributions, fragment-labelled works, corporate or mediated authorship, and works or passages whose attribution is substantially disputed. This includes disputed material transmitted inside the Demosthenic, Antiphontic, Lysian, Platonic, Aeschylean, Isocratean and Xenophontic corpora, as well as the pseudo-Hesiodic Shield of Heracles. The core Theogony and Works and Days are retained under Hesiod. Romans is retained under Paul, while disputed and anonymous New Testament works are excluded.

The Iliad and Odyssey are retained as distinct conventional corpora under the author labels Homeric-Iliad and Homeric-Odyssey; the labels do not claim a shared biographical author. Every work labelled Fragments is excluded.

Where an upstream corpus files two works under one name and only one of them is that author's, the works are named separately. The Cynegetica is transmitted under Oppian's name but is not by the poet of the Halieutica, so Oppian is the Halieutica alone and the Cynegetica is Pseudo-Oppian.

Pseudo-Oppian is retained rather than excluded, which is an exception to the rule above. A Pseudo- label is normally dropped because it names nobody: the work could be by the author it is transmitted under, by another author already in the benchmark, or by several hands. The Cynegetica is instead the work of one Syrian poet, demonstrably neither the Cilician poet of the Halieutica nor anyone else published here. Its attribution is secure in everything except the poet's name, so it is a usable label. The exceptions are listed in RETAINED_PSEUDONYMOUS_AUTHORS. Machine-readable reasons and pre-deduplication counts are recorded in metadata/build_report.json.