Ericu950/Stoicheia-meter-silver

Stoicheia meter silver

Training data behind Stoicheia-meter (macronization + scansion) and Stoicheia-macronizer. Every label is silver: produced by rule, by constraint solving, or converted from published markup — none of it is hand-annotated. The hand-annotated evaluation benchmark is Norma, in a separate repository.

Splits

There is no gold/silver ambiguity about what is train and what is held out. The training code (meter/encode.py, meter/train.py in the code repository) reads these files as follows.

role	files	lines
train — verse macronization	hypotactic.tsv, drama_ia6.tsv, drama_ia6_tet.tsv, anthology.tsv, nonnus_quintus.tsv, babrius_chol.tsv, theocritus_doric.tsv, theocritus_other.tsv, sweep1_hex.tsv, sweep1_ia6.tsv, sweep1_eleg.tsv	176,573 after screening
train — prose macronization	oga_0.tsv … oga_3.tsv (150,000 sampled per epoch)	1,625,256 after screening
train — scansion	scanner/corpus_v3.tsv, minus the dev and test works below	66,455 after screening, split by work
validation — macronization	dev.tsv (= dev.txt with the plain column added), Aristophanic verses	764 (718 encodable)
validation — scansion	works odyssey9, isthmians, seven inside scanner/corpus_v3.tsv	—
test — scansion	works iliad22, pythians, persians, theognis inside scanner/corpus_v3.tsv	—
test — macronization	not here: the norma benchmark	—

sweep1_chol.tsv is byte-identical to babrius_chol.tsv and is skipped by the trainer; oga_replay_15k.tsv and oga_replay_60k.tsv are earlier subsamples of the OGA prose, kept for provenance and unused by the released recipe.

Leakage screen. Before encoding, any training line is dropped if a 20-letter shingle of it occurs in Norma (both tasks, dev and test) or in dev.tsv; short eval lines are matched exactly. The screen removed 1,794 lines and is deliberately over-eager — dropping a few extra silver lines is free, leakage is not.

Format

Macronization TSVs are two headerless columns, plain and marked:

χρὴ γινώσκειν ὅτι πάσης τῆς γῆς	χρὴ γι^νώσκειν ὅτι^ πά_σης τῆς γῆς

^ marks a short vowel, _ a long one, and only after an ambiguous bare α/ι/υ: a vowel with a circumflex or iota subscript, or inside a diphthong, has its length fixed by the orthography and is left unmarked. An unmarked ambiguous vowel is an abstention, not a claim of shortness.

The scanner corpus is three headerless columns — work, metre, bracketed verse — where […] is a heavy syllable and {…} a light one:

semonides	iambic trimeter (tragic)	[ὦ] [παῖ] {τέ}[λος] [μὲν] [Ζεὺ]{ς ἔ}[χει] {βα}[ρύκ]{τυ}[πος]

How the labels were produced

1. Converted syllable-weight markup (hypotactic.tsv) from Hypotactic, turned into vowel lengths wherever an open, coda-less syllable's weight determines its nucleus.
2. Constraint-solver mining (the remaining verse TSVs and scanner/corpus_v3.tsv) of raw digitized verse: each candidate metre (hexameter, iambic trimeter and tetrameter, choliambic, elegiac couplets, trochaic and anapaestic tetrameter) is encoded as a finite-state grammar over syllable weights. A line is accepted only if exactly one grammar scans it, and a dichronon's length is fixed only when every accepting parse agrees — exact constraint propagation, not a model's guess.
3. Rule-based macronization (oga_*.tsv) of the Opera Graeca Adnotata prose with the dictionary-and-morphology system of Thörn Cleland and Cullhed (2026), which abstains where its evidence is inconclusive rather than guessing. Prose cannot be scanned, so this is the only route to labels outside verse; it is also the silver that the 0.87M-parameter character transformer of that same paper — the baseline our macronization results are compared against — was trained on.

Licence CC-BY-SA-4.0, following Hypotactic and OGA.