MODEL_CARDS_INDEX.md

Stoicheia — Released Models and Datasets

All models are character-level masked-diffusion Transformers for Ancient Greek (405M params, d_model=1024, depth 32), loadable via AutoModel.from_pretrained(repo_id, trust_remote_code=True). See each model card for usage examples specific to that checkpoint's task.

Pretrained backbones (11)

Ten rotated, work-level-decontaminated literary folds plus one documentary-clean model. For any passage in the open training corpus, at least one of these eleven has provably never seen it during pretraining.

model	decontamination
Ericu950/Stoicheia-doc_clean	zero documentary (inscription/papyrus) exposure of any kind
Ericu950/Stoicheia-fold-0 … fold-9	rotated 80/10/10 literary split, fold k's test set unseen by fold k's model

Fine-tuned downstream models (4)

All built on Stoicheia-doc_clean.

model	task
Ericu950/Stoicheia-restoration-test0 … -test9	documentary restoration, ten checkpoints — checkpoint k holds out every PHI/TM identifier ending in digit k, so every document in the corpus has a model that provably never saw it
Ericu950/Stoicheia-tagger-parser	morphosyntactic tagging (XPOS/UPOS/lemma) + dependency parsing
Ericu950/Stoicheia-meter	macronization (vowel length) + metrical scansion, trained jointly
Ericu950/Stoicheia-macronizer	macronization only — the arm the paper's macronization ablation is measured on

Datasets

dataset	what it is
Ericu950/AncientGreek	pretraining corpus, ~361M words in pristine and repaired tiers
Ericu950/SyntheticAncientGreek-CorpusCorporum	bronze synthetic augmentation tier, machine-translated from Latin
Ericu950/Inscriptions_2	PHI inscriptions used for the 10-fold split and restoration fine-tuning
Ericu950/Stoicheia-meter-silver	silver macronization/scansion training data (Hypotactic-derived + constraint-solver mined)
Ericu950/norma	mirror of the Norma Syllabarum Graecarum benchmark, in the exact split the paper evaluates on

AncientGreek omits the Database of Byzantine Book Epigrams, which is CC BY-NC-SA and so cannot travel inside a CC BY-SA compilation; scripts/fetch_dbbe.py refetches it under its own terms. Not re-released at all: the OGA/AGDT treebank splits, which come from Celano's own repository and are cited rather than duplicated.