Stoicheia is a 405M-parameter character-level masked-diffusion encoder for Ancient Greek (d_model 1024, depth 32, banded attention: three of every four blocks attend within a 256-character window, the fourth globally). Its input is factored into five aligned planes -- letters, word/sentence boundaries, diacritics, capitalization, punctuation -- each of which can be masked independently to an explicit unknown state at inference. That is what lets one model read an edited text, scriptio continua, and a lacuna of unknown length without changing anything but its input.

Anonymous release accompanying a paper under review.

Stoicheia-doc_clean with four heads on one shared backbone through an ELMo-style scalar mix: factored XPOS, an edit-script lemmatizer, a UPOS auxiliary, and a biaffine dependency parser. Everything below comes from a single forward pass -- there is no pipeline of separate models.

Usage

python

import sys, torch
from transformers import AutoModel
from huggingface_hub import snapshot_download

REPO = "anonymous-stoicheia/Stoicheia-tagger-parser"
# this model's processor needs the label vocabularies beside it, so take the whole snapshot
local = snapshot_download(REPO, allow_patterns=["*.json", "*.txt", "*.py", "*.model"])
sys.path.insert(0, local)
from processing_char_bert_joint import CharBertJointProcessor

model = AutoModel.from_pretrained(REPO, trust_remote_code=True).eval()
proc = CharBertJointProcessor.from_pretrained(local)

batch = proc(["μῆνιν ἄειδε θεὰ Πηληϊάδεω Ἀχιλῆος".split()])
with torch.no_grad():
    out = model(**batch)

for i, w in enumerate(proc.decode(out, batch, ud=True)[0], 1):
    print(i, w["form"], w["lemma"], w["upos"], w["xpos"], w["head"], w["deprel"])
# 1 μῆνιν μῆνις NOUN ... 2 obj
# 2 ἄειδε ἀείδω VERB ... 0 root
# 3 θεὰ θεά NOUN ... 2 orphan
# 4 Πηληϊάδεω Πηληιάδης NOUN ... 5 appos
# 5 Ἀχιλῆος Ἀχιλλεύς NOUN ... 1 nmod