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, published in two syntax layers: the merged human annotation (conllu_human) and one uniform machine parse of every sentence (conllu_machine). It defines 1-, 5-, and 10-sentence attribution tasks on six tracks. 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)

This is the third edition of the benchmark. The second (September 2026) was rebuilt after an adversarial review of the first, which found that no track held out a work, that the old source-held-out track mostly held out a second copy of the same text under a sibling annotation project, that exact deduplication missed the same passage when two projects segmented it differently, and that the fifty-sentence tasks rested on one test chunk per author. A second review of that edition found what the first had not reached: the human syntax layer still carried a project fingerprint in its tree shape, because the AGDT lineage attaches prepositions, subordinators, conjunctions and the copula the Prague way and the UD projects the UD way; the work-held-out track was, for most of its authors, a genre-identification task; and the content of a sentence, not its style, closed the larger tasks. This edition converts the Prague-style heads to Universal Dependencies before publication, holds out a second whole work per author for model selection, reports the work-held-out authors in two halves, adds work-held-out versions of two matched tracks and a content-masked view of every text, and renames the hexameter track for what it holds. Splits and chunk boundaries of the tracks that already existed are unchanged from the second edition; the trees, the work-held-out selection splits and the track inventory are not, so scores are not comparable across editions.

Tasks

There are three task sizes, 1, 5 and 10 sentences, on each of six tracks. Prose and verse sentences share one closed-set authorship problem; the genre column preserves their source representation.

Track	What its test split holds out	Purpose
sentence_*	random sentences of every work	the main, in-work track: every author, most data
work_held_out_*	one whole work per author, unseen in training, annotated by a project that also annotates the author's training works	attribution that cannot be passage memorisation or annotator recognition
tragedians_*	as sentence_*, Aeschylus, Sophocles and Euripides only	contemporaries of one genre
attic_orators_*	as sentence_*, Aeschines, Antiphon, Demosthenes and Lysias only	contemporaries of one genre
imperial_historians_*	as sentence_*, Appian, Diodorus Siculus, Dionysius of Halicarnassus, Josephus and Plutarch only	one period and genre
archaic_epic_*	as sentence_*, Hesiod and the two Homeric corpora only	one metre, dialect and period; two of its three labels are corpora, not authors (it was hexameter_* in the second edition)
tragedians_work_held_out_*	one whole play of Aeschylus and of Sophocles, all rows from AGDT	attribution between two contemporaries of one genre, one annotator and unseen works
attic_orators_work_held_out_*	one whole speech of Antiphon, Demosthenes and Lysias, all rows from Gorman	attribution among three contemporaries of one genre, one annotator and unseen works

The matched tracks exist because the main track mixes fourteen centuries and every genre, so that period, dialect and genre do much of the work the word "attribution" implies. Their rows are the main track's rows for those authors, with the main track's splits, so a model trained on the main track can be scored on them without retraining, and a model trained on them alone can be compared with it. Their work-held-out versions go one step further: with the genre, the period and the annotator all held fixed and the test works unseen, what is left to tell the authors apart is style and topic, and the gap between a matched track and its work-held-out version is how much of the matched score was the topic of the works trained on. Only two of the four can be built: Euripides, Appian, Diodorus, Dionysius, Josephus and the two Homeric corpora are single works in the treebanks.

Urdatorn/sphragis-metre carries a verse_hexameter_* track of the same three labels as archaic_epic_* over the scanned lines of those poets, so the two publications can be read against each other, and a wider verse_hexameter_all_* track of every hexameter poet it scans. It has no tragedians_* track: Hypotactic scans no Sophocles and no Euripides.

Configurations

Configurations

Configuration	Authors	Row unit	Train rows	Validation rows	Validation2 rows	Test rows
sentence_1	28	sentence	46,910	6,140	3,030	6,140
sentence_5	28	5-sentence chunk	9,382	1,228	606	1,228
sentence_10	28	10-sentence chunk	4,691	614	303	614
work_held_out_1	12	sentence	17,090	2,760	1,350	3,990
work_held_out_5	12	5-sentence chunk	3,418	552	270	798
work_held_out_10	12	10-sentence chunk	1,709	276	135	399
tragedians_1	3	sentence	5,950	770	390	770
tragedians_5	3	5-sentence chunk	1,190	154	78	154
tragedians_10	3	10-sentence chunk	595	77	39	77
attic_orators_1	4	sentence	3,820	500	240	500
attic_orators_5	4	5-sentence chunk	764	100	48	100
attic_orators_10	4	10-sentence chunk	382	50	24	50
imperial_historians_1	5	sentence	4,110	520	250	520
imperial_historians_5	5	5-sentence chunk	822	104	50	104
imperial_historians_10	5	10-sentence chunk	411	52	25	52
archaic_epic_1	3	sentence	10,960	1,450	720	1,450
archaic_epic_5	3	5-sentence chunk	2,192	290	144	290
archaic_epic_10	3	10-sentence chunk	1,096	145	72	145
tragedians_work_held_out_1	2	sentence	3,350	480	240	580
tragedians_work_held_out_5	2	5-sentence chunk	670	96	48	116
tragedians_work_held_out_10	2	10-sentence chunk	335	48	24	58
attic_orators_work_held_out_1	3	sentence	3,290	330	150	560
attic_orators_work_held_out_5	3	5-sentence chunk	658	66	30	112
attic_orators_work_held_out_10	3	10-sentence chunk	329	33	15	56

Sentences per author label, main track (sentence_1)

Author label	Train	Validation	Validation2	Test	Total
Aeschines	490	60	30	60	640 (1.0%)
Aeschylus	2,460	320	160	320	3,260 (5.2%)
Antiphon	380	50	20	50	500 (0.8%)
Appian	720	90	40	90	940 (1.5%)
Aristophanes	1,380	180	90	180	1,830 (2.9%)
Aristotle	650	80	40	80	850 (1.4%)
Athenaeus	1,830	240	120	240	2,430 (3.9%)
Demosthenes	2,270	300	150	300	3,020 (4.9%)
Diodorus Siculus	730	90	40	90	950 (1.5%)
Dionysius of Halicarnassus	790	100	50	100	1,040 (1.7%)
Euripides	590	70	40	70	770 (1.2%)
Herodotus	5,260	700	350	700	7,010 (11.3%)
Hesiod	680	90	40	90	900 (1.4%)
Homeric-Iliad	5,750	760	380	760	7,650 (12.3%)
Homeric-Odyssey	4,530	600	300	600	6,030 (9.7%)
Josephus	770	100	50	100	1,020 (1.6%)
Lucian	530	70	30	70	700 (1.1%)
Lysias	680	90	40	90	900 (1.4%)
Menander	670	90	40	90	890 (1.4%)
Paul	670	90	40	90	890 (1.4%)
Plato	810	100	50	100	1,060 (1.7%)
Plutarch	1,100	140	70	140	1,450 (2.3%)
Polybius	2,850	380	190	380	3,800 (6.1%)
Procopius	890	110	50	110	1,160 (1.9%)
Sextus Empiricus	500	60	30	60	650 (1.0%)
Sophocles	2,900	380	190	380	3,850 (6.2%)
Thucydides	900	120	60	120	1,200 (1.9%)
Xenophon	5,130	680	340	680	6,830 (11.0%)
All labels	46,910	6,140	3,030	6,140	62,220 (100.0%)

Work-held-out track

Author	Held-out test work	Genre	Group	Validation work	Test project	Training works	Train / validation / validation2 / test rows
Aeschylus	Persians (tlg0085.tlg002)	tragedy	rival	Libation Bearers	agdt	4	2,320 / 320 / 160 / 460
Antiphon	antiphon 6 bu2 (tlg0028.tlg006)	oratory	rival	(in-work sentences)	gorman	1	290 / 50 / 20 / 140
Aristophanes	Acharnians (pedalion:aristophanes:acharnians)	comedy	singleton	(in-work sentences)	pedalion	1	660 / 180 / 90 / 910
Demosthenes	dem 45 tree (tlg0014.tlg045)	oratory	rival	dem 57 tree	gorman	11	2,470 / 170 / 80 / 300
Hesiod	Theogony (tlg0020.tlg001)	epic	singleton	(in-work sentences)	agdt, ud_perseus	1	340 / 90 / 40 / 420
Lucian	Judicium vocalium (pedalion:lucian:judicium-vocalium)	epideictic	rival	Prometheus	pedalion	2	550 / 70 / 30 / 50
Lysias	Lysias 1 bu1 (tlg0540.tlg001)	oratory	rival	lysias 19 bu1	gorman	4	610 / 110 / 50 / 120
Paul	Philippians (proiel-phil)	epistle	singleton	Galatians	ud_proiel	2	610 / 120 / 60 / 90
Plato	Plato_Crito_Travis_Kahl (tlg0059.tlg003)	dialogue	rival	Euthyphro	gorman	1	460 / 240 / 120 / 250
Plutarch	plut fortuna romanorum bu1 (tlg0007.tlg086)	epideictic	rival	plutarch alex fort aut virt bu2	gorman	2	860 / 280 / 140 / 180
Sophocles	Trachiniae (tlg0011.tlg001)	tragedy	rival	Antigone	agdt, ud_perseus	3	2,550 / 440 / 220 / 650
Xenophon	xen symp 3-4 (tlg0032.tlg004)	dialogue	rival	Xen_Anab_book_3	gorman	4	5,370 / 690 / 340 / 420

Period-and-genre-matched tracks

Track	Authors
archaic_epic_*	Hesiod, Homeric-Iliad, Homeric-Odyssey
attic_orators_*	Aeschines, Antiphon, Demosthenes, Lysias
imperial_historians_*	Appian, Diodorus Siculus, Dionysius of Halicarnassus, Josephus, Plutarch
tragedians_*	Aeschylus, Sophocles, Euripides

Work-held-out versions of the matched tracks

archaic_epic_work_held_out_*: not built; 1 of the track's authors have a work of the track's project (agdt) that can be held out; a track needs two.

attic_orators_work_held_out_* (project gorman)

Author	Held-out test work	Genre	Group	Validation work	Test project	Training works	Train / validation / validation2 / test rows
Antiphon	antiphon 6 bu2 (tlg0028.tlg006)	oratory	rival	(in-work sentences)	gorman	1	290 / 50 / 20 / 140
Demosthenes	dem 45 tree (tlg0014.tlg045)	oratory	rival	dem 57 tree	gorman	11	2,470 / 170 / 80 / 300
Lysias	Lysias 1 bu1 (tlg0540.tlg001)	oratory	rival	lysias 19 bu1	gorman	3	530 / 110 / 50 / 120

imperial_historians_work_held_out_*: not built; 1 of the track's authors have a work of the track's project (gorman) that can be held out; a track needs two.

tragedians_work_held_out_* (project agdt)

Author	Held-out test work	Genre	Group	Validation work	Test project	Training works	Train / validation / validation2 / test rows
Aeschylus	Suppliants (tlg0085.tlg001)	tragedy	rival	Persians	agdt	4	2,290 / 300 / 150 / 360
Sophocles	Trachiniae (tlg0011.tlg001)	tragedy	rival	Ajax	agdt	3	1,060 / 180 / 90 / 220

from datasets import load_dataset

sentences = load_dataset("Urdatorn/sphragis", "work_held_out_10")

Every configuration has four splits: train, validation, validation2 and test. validation is for selecting a single model; validation2 is reserved for choosing among combinations of models, so that a leaderboard combination is never selected on the split its constituents were tuned on. Atomic models must not read it.

The suffix gives the chunk size in every split. _10, the largest, defines one shared source corpus for all three tasks: an author is retained only with at least 100 training, 50 validation, 10 validation2 and 50 test source rows on that track, and every split is reduced to a multiple of 10 source rows per author. Fifty test sentences are five 10-sentence test chunks, so no author's score on any task rests on a single chunk. _1 publishes the shared rows atomically, while _5 and _10 regroup those exact same rows into fixed benchmark chunks. Author inventory and total source text are therefore identical across the task sizes of one track.

Source-row selection uses seed 917. For each retained author and split, the unavoidable n mod 10 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 canonical 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.

Splits

Splits use random seed 917 and begin with independent stratification within each author and canonical work: 75% train, 10% validation, 5% validation2 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 10-row eligibility and remainder bottleneck is then applied to all three task sizes.

A canonical work is the work a sentence belongs to independently of which file or project holds it. The projects split one work into many files (xen cyr 1_1-2 bu1, xen cyr 1_3-4 bu1, ...) and name one work in several ways (Pedalion's Crito is Gorman's Plato_Crito_Travis_Kahl); work_id keeps the file-level identifier and canonical_work the TLG work it belongs to, which is what the splits, the chunking and the work-held-out track use.

Work-held-out track

For each author, one whole work is held out as that author's entire test split. The candidate works are those with at least 50 sentences and at most half of the author's sentences, annotated by a project that also annotates another work of the same author; the candidate closest to 10% of the author's rows is taken.

A second whole work is then held out for model selection when the author has one to spare: at least 60 sentences, at most half of the author, closest to 15% of the author's rows, and leaving the training works at least 100 sentences and the test work's annotation project. Its sentences are the author's validation and validation2 splits, so that a model is selected out-of-work as well as tested out-of-work; held_out_validation_work names it on every row. Nine of the twelve authors have such a work. Antiphon, Aristophanes and Hesiod have two works in the treebanks, so their validation and validation2 are random sentences of the training work, as in the second edition, and the column is null.

Every row also carries held_out_genre, the literary genre of the author's test work, and held_out_group. An author is a rival when another author of the track has a work of that genre, a singleton when none has. The second review showed the difference: Hesiod is the only epic poet on the track, Paul the only epistolographer, Aristophanes the only comedian, and a model scores 90–100% on them by recognising the genre. The leaderboard reports the nine rivals (two tragedians, three orators, two dialogue writers, two writers of epideictic prose) as the track's headline and the three singletons as a genre floor. Every score on the track should carry an interval over its twelve works rather than over its sentences: the work is the statistical unit.

The within-project condition is deliberate. The first edition's source-held-out track held out an annotation project, and because UD Perseus is a conversion of the AGDT trees and Vanessa Gorman annotated both her own repository and the AGDT files of several prose authors, it mostly held out a second copy of the same text. Holding out a work inside a project changes the text and keeps the annotator, so a drop from sentence_* to work_held_out_* is a drop in attribution, not in annotation recognition. Authors with a single work, or with no work satisfying the conditions, are absent from the track; the reasons are recorded under work_held_out_track in metadata/build_report.json. Only the test claim is work-held-out: model selection on validation is in-work, and the gap between validation and test on this track is one of the quantities the track exists to expose.

Deduplication

Exact deduplication first: normalization lowercases, maps final sigma, removes diacritics and modifier letters, and retains letters only, so that Gorman's U+02BC elision mark and the AGDT combining mark no longer hash the same sentence differently. The chosen parse follows a deterministic source priority, while all exact-source provenance is merged into source_records. Sentences whose retained author labels conflict are omitted rather than assigned.

Passage-level near-duplicate removal second, because exact deduplication misses the same passage when two projects segment it differently or read a different edition (Diodorus 11 in Gorman and UD Perseus, Herodotus 1 in PROIEL and UD Perseus, the Crito in Pedalion and Gorman). Two sentences of one author are near-duplicates when half of either one's word 4-grams, and at least three of them, occur in the other; the passage is then kept in one project only, and a lone surviving sentence between two removed neighbours of the same document is removed with them as a segmentation fragment. Across authors the bar is a quotation, not a stock phrase: at least four shared 4-grams and four fifths of the shorter sentence, in which case every copy is removed (Xenophon quoting Hesiod and Homer in the Memorabilia, Sextus Empiricus quoting the Odyssey). Formulaic repetition inside one project is the author's own text and is kept. Counts, per-author removals and examples are under near_duplicate_passages in the build report.

In numbers: the seven projects supplied 110,831 sentences, of which authorship curation retained 92,123. Exact deduplication removed 23,876 repeated rows and 1,280 rows whose retained author labels conflict, leaving 66,967. Passage-level near-duplicate removal then dropped 649 rows (464 kept in the other project, 48 segmentation fragments beside them, 137 quotations shared across authors), while 1,021 pairs of same-project repetition were kept. 226 sentences with corrupt tokens were excluded, almost all of them Pedalion's Sappho and Heron, and 3,384 sentences of 15 authors below the retention threshold were set aside before finalisation, leaving 62,708 sentences of 28 authors to enter the tracks.

Principal columns

All configurations contain:

• id, author, work, work_id, canonical_work, genre, text;
• conllu_human, conllu_machine, cts_urn, passage, treebank_source;
• text_distorted: the content-masked view of text (see below);
• 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 work_held_out_* configurations, and the two work-held-out versions of matched tracks, add held_out_work, held_out_work_title, held_out_validation_work, held_out_validation_work_title, held_out_genre and held_out_group, repeated on every row of the author.

The content-masked text

text_distorted is text with every token the machine parser tags NOUN, PROPN, ADJ or VERB replaced by that tag, token for token:

οὐκ ἐγὼ λεγον                →  οὐκ ἐγὼ VERB

It is the text-distortion method of Stamatatos (2017) with the parser's uniform tags deciding what is content, so the masking is the same for every author. What survives is the author's function words, particles, pronouns, articles, prepositions, negations (adverbs in UD, so kept), numerals, and the shape of the sentence. A model trained on this column instead of text is the benchmark's style-only condition: lemma TF-IDF reaches 100% on imperial_historians_10 from text because Josephus writes about Judaea and Appian about Rome, and cannot from text_distorted.

Two syntax layers

conllu_human is the merged human annotation of the seven projects, reduced to the scheme-neutral representation described below. conllu_machine is the same sentence parsed by one pinned model, Ericu950/Stoicheia-tagger-parser, the parser that annotates all of Sphragis-Metre, so it is uniform by construction: whatever a model learns from it cannot be a difference of annotation convention. Its UPOS, FEATS, HEAD and DEPREL are the parser's own and use the full Universal Dependencies v2 inventories; its lemma is the human lemma of the same token, because lemmatisation is the one layer of the human annotation that is not convention-bound. The two columns have the identical nested type, so a reader can be pointed at either.

The parser was trained on the five-fold treebank splits of Celano (2025), with a shared held-out test set, over the Universal Dependencies Ancient Greek treebanks (arXiv 2608.07249); it has therefore seen the UD Perseus and PROIEL sentences of this benchmark with their human annotation, and agrees with those two projects' heads (UAS 0.83 and 0.72 against conllu_human) more than with the AGDT-lineage projects' (0.56–0.62 before the head conversion), whose native annotation it was not trained on. The layer is uniform in inventory and in the parser's habits, not in its error rate across projects. The parser sees the published text, that is, lowercased, without editorial punctuation and with elision marks removed, so an elided δέ reaches it as the bare form δ. Its quality against an independent manual correction of 100 published trees is reported under machine_syntax in the build report and in metadata/machine_syntax_audit.json: on 100 sentences (1440 tokens) the parser scores UAS 0.74, LAS 0.56 and label accuracy 0.69 (macro-F1 over relations 0.40); UPOS accuracy 0.78 (macro-F1 0.63); FEATS micro-F1 0.96 and macro-F1 0.89 over key=value pairs. The systematic differences are conventions the parser does not share with the corrector: it never emits PROPN or AUX, writes present-stem forms as Tense=Pres without Aspect=Imp, tags postpositive particles as ADV and sometimes roots the sentence on them; acl and discourse are never produced, appos and xcomp rarely. Its coordinators, attached to the first conjunct or to the following one as its two training treebanks do, are re-attached to the following conjunct at publication, as in the human layer. Per-relation, per-tag and per-feature scores, and the corrector's conventions, are in metadata/machine_syntax_audit/.

That audit corrected the parser's trees, which invites the objection that a corrector leaves a parser decision alone unless it is clearly wrong. The same 100 sentences were therefore annotated a second time from their forms alone, by the same annotator without access to the parser's trees or to the first audit (metadata/machine_syntax_audit/annotated_fable_scratch.jsonl; comparison in metadata/machine_syntax_audit_agreement.json, produced by scripts/compare_syntax_audits.py). The two audits agree with each other at UAS 0.965, LAS 0.949, UPOS accuracy 0.984 and FEATS micro-F1 0.997, and the parser scores the same against either: UAS 0.744 against the correcting audit and 0.747 against the independent one, LAS 0.560 and 0.561, UPOS 0.778 and 0.774. On 23% of tokens the two audits agree on a head the parser did not choose; on 1.3% the parser agrees with one audit against the other. Correcting rather than annotating did not move the parser's score, and the audits' LAS is the reliability ceiling a parser could reach on these conventions. The layer is uniform, not gold; a model that reads it reads the parser's habits equally for every author.

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; _5 and _10 hold exactly 5 or 10. No whitespace separator encodes unit boundaries.

conllu_human and conllu_machine have the same number of units as text, and each unit is a list of sentences, each sentence a list of token objects:

row["conllu_human"][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_human"][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_human"][0]))

The build validates every published human unit by rendering it and parsing the result with the official CoNLL 2018 shared-task loader, so the structure is provably still CoNLL-U; every machine unit is checked to be a single-rooted tree over the same tokens.

The _5 and _10 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's target, and constituent_ids lists the _1 row ids in order, so a chunk can be reassembled from its atomic rows exactly.

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).

Universal Dependencies heads. AGDT, Gorman, Pedalion and Harrington annotate in the Prague/Arethusa convention: a preposition heads its noun (AuxP), a subordinator its clause (AuxC), a conjunction its conjuncts (COORD with _CO members), an apposition marker its appositives (APOS) and the copula its predicate nominal (PNOM). UD Perseus and PROIEL attach each the other way round. The second edition published the native heads with UD labels, so that a token labelled case had dependents in 97–100% of cases in the AGDT lineage and in 0% in the UD projects, and the shape of one prepositional phrase named the project. The third edition converts the AGDT-lineage trees to UD heads before any label is mapped (scripts/prague_to_ud.py): the argument of an AuxP/AuxC takes its place, the first conjunct heads a coordination and its marker, the first appositive heads an apposition, and εἰμί with a predicate nominal becomes cop (or aux under a participle). On the 14,068 sentences AGDT and UD Perseus both annotate, head agreement rises from 62.1% to 95.4%; what remains is the two projects' different placement of the coordinator, which the build then settles for every project and for the machine layer alike: every cc attaches to the conjunct that follows it (PROIEL's convention and the UD guideline; UD Perseus put it on the first conjunct). 612,001 of the 1,479,841 native tokens changed head; the native head and a head_converted flag stay in native_syntax. No sentence had to be left unconverted.

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

Fourteen authors are annotated by more than one project (twelve at the time of this measurement). 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 conllu_machine layer and the within-project work_held_out_* track exist. The measurements above were made on the second edition's trees. The third edition's head conversion removes the difference in tree shape that put a case token under its noun in one lineage and over it in the other, and attaches every coordinator the same way; the difference in rate, how often a project chooses to mark a relation at all, is what remains.

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 — rates of UPOS, DEPREL and FEATS values, and since the third edition also the geometry of the tree: head direction, binned head distance and function words heading anything, which is what the syntax contenders read and what the head conversion removed — on each syntax layer, and writes metadata/leakage_report.json for conllu_human and metadata/leakage_report_conllu_machine.json for the machine layer. 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. The remedy for what remains is conllu_machine, which carries no project convention at all, and the work_held_out_* track, whose test works are annotated by the same project as their training works.

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, which is why the track that holds them is archaic_epic_* and not a track of three poets. Every work labelled Fragments is excluded.

Athenaeus is retained on the main track only, with a caveat: the Deipnosophistae is a compilation, and most of its sentences quote or paraphrase authors who are not in the benchmark. Quotations of authors who are in it are removed by the cross-author near-duplicate rule; the rest remain under the label Athenaeus, which therefore names the compiler's choice of sources at least as much as his prose. He is absent from every matched and held-out track.

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.