Stoicheia -- documentary-clean backbone
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.
The documentary-clean backbone: zero exposure to inscriptions or papyri of any kind. Every downstream model in this release is fine-tuned from it, so no epigraphic or papyrological result can be contaminated by pretraining. It is the flagship checkpoint -- use this one unless you specifically need a literary fold.
What this checkpoint has not read
Ten literary folds rotate an 80/10/10 split built by a 13-stage pipeline that clusters records into editions of the same work -- by canonical identifier where one exists, by duplicate-aware content matching where it does not -- and then excises from each fold's training data every sentence colliding with its held-out zones (exact and reordered duplicates, word 5-grams, document-level near-duplicates). The guarantee is exact at the fold level: for any passage of the open corpus, at least one released checkpoint has provably never seen it, which is what makes it possible to interrogate a transmitted text with a model that cannot merely be recalling it.
Stoicheia-doc_clean extends the same discipline to documentary text: every inscription and
papyrus is excluded, along with anything a contamination screen flags as quoting one.
Usage
import torch
from transformers import AutoModel
from huggingface_hub import hf_hub_download
REPO = "anonymous-stoicheia/Stoicheia-doc_clean"
model = AutoModel.from_pretrained(REPO, trust_remote_code=True).eval()
# `trust_remote_code` loads the model classes, but the processor is a separate helper:
# fetch it into the working directory before importing it.
hf_hub_download(repo_id=REPO, filename="processing_char_bert.py", local_dir=".")
from processing_char_bert import CharBertProcessor
proc = CharBertProcessor()
# a gap of uncertain width in unaccented scriptio continua: "[N±M]" scores every width
# in N-M..N+M by the model's own confidence, and restores accents and word division too
text = "εναρχηηνο[5±3]καιολογοςηνπροστονθεον"
best, width, candidates = proc.restore_elastic(model, text, mask_dia_boundary=True)
print(best) # ἐν ἀρχῇ ἦν ὁ λόγος, καὶ ὁ λόγος ἦν πρὸς τὸν θεόν.
print(width) # 5 -- the width the model judged most likely
Checkpoint provenance
This is the step with the best held-out bits-per-character (step 116,001), not the last step reached: pretraining is dev-driven, so the eleven backbones stop at different points (78K--122K). training_metadata.json in this repo carries the training configuration.