CharDiff-grc-meter (macronization + metrical scansion)

A CharDiff-grc backbone fine-tuned for two joint per-letter classification tasks: macronization (recovering vowel length on ambiguous bare α/ι/υ, since Ancient Greek orthography never marks it) and metrical scansion (heavy/light/verse-final syllable weight). Fine-tuned from Ericu950/CharDiff-grc-doc_clean with an ELMo-style learned mix over all 32 backbone layers, plus a fine-tune-only capitalization input channel.

The benchmark used is Albin Thörn Cleland's Norma benchmark (932 hand-annotated lines, 16 authors, Homer to Nonnus, plus 78 lines of Aristophanic lyric, deliberately disjoint from the Hypotactic scansion resource so a Hypotactic-trained model gains no unfair benchmark advantage). Training data was source from Cleland's macronizer, from a ~130,000-line silver training corpus (mined via a rule-based finite-state metrical constraint solver over hexameter/iambic/choliambic/elegiac/trochaic/anapaestic grammars, with exact constraint propagation — a dichronon's length is fixed only when every valid parse of its line agrees), and Hypotactic-

Results (this checkpoint, single model, no ensembling)

task	split	metric	value
Macronization (long/short, ambiguous α/ι/υ)	Norma test (n=2660, 16 sources)	balanced acc / acc	0.924 / 0.936
Macronization	Norma dev (n=306, 13 sources)	balanced acc / acc	0.942 / 0.964
Scansion, gold-syllabified (none/heavy/light/verse-end)	Norma test (n=29,171)	balanced acc / acc	0.896 / 0.879
Scansion, whole verses by work	held-out work split (n=131,948)	balanced acc / acc	0.991 / 0.990

Usage

import torch
from transformers import AutoModel

model = AutoModel.from_pretrained("Ericu950/CharDiff-grc-meter", trust_remote_code=True)
model.eval()

from processing_char_bert_meter import CharBertMeterProcessor
processor = CharBertMeterProcessor()

# macronization: only ambiguous alpha/iota/upsilon positions get a mark (`_`=long, `^`=short)
text = "βαρυκτυπος ανθρωπος"
batch = processor(text)
with torch.no_grad():
    out = model(**{k: v for k, v in batch.items() if not k.startswith("_")})
print(processor.decode_macronization(out, batch))

# metrical scansion: [heavy] / {light} syllable brackets, per letter
print(processor.decode_scansion(out, batch))

Unlike the base pretraining model, capitalization (cap) is a real input here (fine-tune-only channel, zero-init) — the processor's __call__ includes it in the batch directly rather than keeping it out-of-band.

Citation

See the CharDiff-grc repository and the associated paper (citation to be added on publication).