Datasets:
Tasks:
Text Classification
Formats:
parquet
Languages:
Ancient Greek (to 1453)
Size:
100K - 1M
ArXiv:
License:
Commit History
29 days ago
-
Content-masked text for the metre publication; metre leakage probe in the rebuild; keep the audit agreement across rebuilds
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>982416c -
Third edition: UD heads in the human layer, out-of-work selection, work-held-out halves and genre versions, content-masked text
Rebuilt after Referee Report II. The AGDT-lineage trees are converted from Prague-style to Universal Dependencies heads before publication (scripts/prague_to_ud.py; AGDT/UD Perseus head agreement 62.1% -> 95.4%), every coordinator in every project and in the machine layer attaches to the following conjunct, and the markup-only floors now see head geometry. The work-held-out track holds out a second whole work for validation where the author has one, carries the genre of its test work and whether another author of the track writes that genre (nine rivals, three singletons), and gains single-project work-held-out versions of the tragedians and Attic-orators tracks. Every row carries text_distorted, the content-masked text. The hexameter track is renamed archaic_epic. The 100 audited machine trees were annotated a second time from scratch and the two audits compared. The rebuild job now preserves the manual audit directory across rebuilds. Co-Authored-By: Claude Fable 5.1 <noreply@anthro…1e6d8b5
-
Name the metre tracks verse_hexameter, not hexameter
The sentence publication already has a hexameter track, and the model code keys a task to its dataset by the bare configuration name: slurm/dataset_paths.sh maps hexameter_1 to family 5 and therefore to the Sphragis checkout, so a second hexameter_1 in Sphragis-Metre would have sent every job that named it to the wrong corpus. Both metre tracks now keep the publication's verse prefix. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>dfd7f4c -
Add two hexameter tracks to the metre publication
The scanned-line publication had one track, so a model could score on it by telling the metres apart rather than the poets: the metre column alone separates Aeschylus, Pindar and Lycophron from the other fourteen, and it is a Hypotactic annotation rather than a fact of the transmitted text. hexameter_{1,5,10} is Hesiod and the two Homeric corpora, the labels of the sentence publication's track of the same name, so both publications put one question to one set of labels. hexameter_all_{1,5,10} is every line whose metre is hexameter, fourteen poets from Homer to Nonnus, where that shortcut is gone because the metre label is constant: Callimachus' elegiac couplets, Theognis' pentameters and Theocritus' Aeolic lyric leave with the three non-hexameter poets. Both reuse the main track's lines and splits. There is no tragedians track here, although the sentence publication has one: Hypotactic scans no Sophocles and no Euripides, so it would carry the single label Aeschylus. The reason is re…0879baf
-
Rebuild the metre publication under the second-edition text rules
The Hypotactic line cache is re-parsed under the current tokenisation, and the span, crop and coverage checks share one text reduction: Greek letters without the archaic numeral letters, which the parser's tokeniser drops. A gold alignment component whose line boundary falls inside a word written across two lines leaves the gold reference with a counted reason instead of aborting the build. The rebuild job takes the cache path from STOICHEIA_CONLLU and builds both publications when SKIP_METRE=0. The sentence publication is unchanged by this commit: it reproduces revision 667319c byte for byte. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>531a22f
-
Second edition: work-held-out and matched tracks, four splits, two syntax layers
Rebuilt after an adversarial review of the first edition. Every split, chunk boundary and seed changed (seed 917). The 50-sentence task and the source-held-out track are gone; sentence_{1,5,10} is joined by work_held_out_{1,5,10}, whose test split is one whole work per author held out within an annotation project, and by the matched tracks tragedians, attic_orators, imperial_historians and hexameter. A fourth split, validation2, is reserved for choosing combination models. Exact deduplication now ignores modifier letters, passage-level near-duplicate removal follows it, a token is a word only if it contains a Greek letter, and every sentence's text is exactly its published forms. The human trees are published as conllu_human beside conllu_machine, one Stoicheia parse of every sentence carrying the human lemma, with an independent manual audit of 100 trees under metadata/machine_syntax_audit/. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>667319c
-
Measure inter-project agreement on the sentences two treebanks share
The build discards 23,282 exact duplicates, and those discarded copies are the only place in the corpus where two projects analyse the same Greek. Everywhere else scheme and author are confounded by construction, so this is the only direct measurement of how far the schemes diverge. Re-parses the sources, regroups by the build's own duplicate key, aligns shared sentences on character spans, and reports Cohen's kappa per project pair for the native relation and for every model-facing column. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>28106b6
-
Remove the SQLite inspection mirror
The mirror duplicated every row of every configuration into a 2.4 GB SQLite file for local browsing. Parquet has always been the canonical representation, the Hugging Face viewer renders the structured columns directly now that they are real Arrow types rather than JSON in strings, and the mirror had to be rebuilt and revalidated on every change. Drops inspection/, scripts/build_sqlite_mirror.py, its slurm job, the build and validation steps in the rebuild pipeline, and the LFS pattern that tracked it. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>2fc1d6e -
Merge away four annotator-driven categories
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 identical; changing the author within one project gives the stylistic baseline. Four categories failed that test, in percentage points of rate: vocative annotator 0.36 author 0.00 Gender=Com annotator 0.53 author 0.02 appos annotator 0.43 author 0.04 discourse annotator 6.66 author 4.70 vocative is the clearest: it did not vary between sixteen authors of one project at all, so it carried no authorial signal whatever, while varying between annotators of one author. FEATS Case=Voc identifies vocatives exactly, so the relation was redundant as well as leaky; it becomes dep. 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: δέ, γάρ and μέν are the same tokens, filed differently. …265998e -
Replace the 100-row task with a 5-row task and admit 11 more authors
The 100-row eligibility bar excluded eleven authors that clear 50 comfortably, most of them by a hair: Euripides had 79 validation and 79 test rows, Aristotle 87, Lysias 97. A closed set that omits Euripides and Aristotle is hard to defend when the only thing keeping them out is a threshold. Both publications now define 1-, 5-, 10- and 50-row tasks. The bottleneck moves to 50, the largest chunk, so every retained author still has at least one complete chunk of every task size in validation and in test, and every smaller size divides 50, which is what lets one retained row set fill all four tasks. Sphragis grows from 17 to 28 authors and from 43,300 to 50,500 training rows, gaining Aeschines, Antiphon, Appian, Aristotle, Euripides, Hesiod, Lucian, Lysias, Menander, Paul and Sextus Empiricus. The source-held-out track grows from 11 to 12. Sphragis Metre gains Tryphiodorus. The cost is that the thinnest authors now have a single 50-row chunk in test, so their per-author accuracy on the…eaebf86 -
Fold the positional tag into FEATS and stop publishing XPOS
XPOS and FEATS were two representations of one morphological analysis, which is how stale information starts. They had already drifted: each carried distinctions the other lost, and they disagreed 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 nine-position tag was the only record of which one a verb was — a distinction that matters for authorship attribution in Greek. 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. It alone separated the medio-passive from the true middle, which FEATS flattened onto Voice=Mid. FEATS meanwhile was the finer of the two on ambiguous genders such as Masc,Neut. FEATS also contradicted itself, reading pluperfects as Past 837 times and Pqp 354 times. The tag is now folded into FEATS and dropped. All seven Greek tenses stay distinct in s…13ef6e2
-
Publish structured columns instead of JSON inside strings
The CoNLL-U column was one long string of escaped tabs and newlines, which is unreadable in the Hugging Face viewer and in any Parquet tool. Every column that held JSON in a string is now a real nested Arrow type, so load_dataset returns Python lists and dictionaries with nothing left to parse: row["conllu"][0][0][2] {"id": 3, "form": "βῆ", "lemma": "βαίνω", "upos": "VERB", "xpos": "v3saia---", "feats": "Mood=Ind|Number=Sing|Tense=Past|Voice=Act", "head": 0, "deprel": "root"} A unit is a list of sentences and a sentence a list of tokens, because one Sphragis Metre line can span a sentence boundary; every Sphragis sentence unit holds exactly one. scripts/conllu_units.render_conllu returns canonical CoNLL-U text, and the build validates every published unit by rendering it and parsing the result with the official CoNLL 2018 loader, so the structure is provably still CoNLL-U. Naming the fields once in the schema rather than once per token also makes the data smaller t…31e2783 -
Remove annotation-scheme leakage from model-facing columns
Sphragis merges seven annotation projects whose author inventories do not overlap, so every convention that separates the projects also separates the authors. A classifier seeing annotation markup and no Greek at all reached 63.4% accuracy and 45.3 macro-F1 on the 17-way sentence_1 task against a 14.6% baseline, and single rules were near-deterministic: no PUNCT token identified Herodotus at 99.6% precision over 83% of his sentences, and an empty MISC column separated the Iliad from the Odyssey at 83.5%. Every published column is now reduced to one shared representation: - MISC is cleared; NativeRel, NativeHead and HeadRepair move to a new native_syntax audit column, keyed by published and source token index. - Editorial punctuation tokens are removed and trees re-indexed, because UD PROIEL does not tokenize punctuation at all. - Elision is written as a combining comma by AGDT and UD Perseus, as three different apostrophes within Gorman, and as U+2019 by Pedalion and PROIEL; …6fc8784 -
Reparse metre syntax uniformly with Stoicheia
83dae89
-
Publish metre inspection columns first
76be066 -
Add derived metrical line features
0bd1a9f -
Omit genre from metre publication schema
838a0aa
-
Derive metre features from word boundaries
30903f0 -
Rename metre task configurations to verse
9c2923f -
Support exact 50-line metre task
71cd466
-
Add exact 50-sentence benchmark task
8280489 -
Preserve nested metre lines and Hypotactic caesurae
d2c51c1
-
Parse Hypotactic verse as sentences before line cropping
0280c7d -
Add Stoicheia parsing for complete Hypotactic coverage
5a7dc0a -
Rebuild Sphragis from the complete treebank sentence pool
733d9e5 -
Chunk training splits at fixed task sizes
f194eef -
Enforce exact verse boundaries and nested CoNLL-U
c6b9888
-
Represent text chunks as JSON lists
e1e2e3d -
Add SQLite inspection mirror
7e9ba33 -
Remove identifying metadata from metrical lines
1bcc6c0
-
Fix metrical chunk aggregation
24c3b46 -
Unify source text across task sizes
4a0d2b0 -
Remove CoNLL-U provenance leakage
e1c252e -
Validate all CoNLL-U with official loader
ac04b84 -
Freeze exact evaluation chunks
ecfe7f5
-
Add chunk-size dataset variants
85f00da -
Rebuild work-stratified splits with seed 776
1f231e9 -
Restore core Hesiodic works
440e215 -
Curate conservative authorship benchmark
d44b077 -
Align verse across sentence boundaries
aa7a784
-
Add Sphregis 0.1 beta dataset
1bd5e76