koine-t5-omni

Koine-T5-Omni

A multitask Ancient Greek sequence-to-sequence model: a LoRA adapter over bowphs/GreTa trained jointly on part-of-speech tagging, lemmatisation, full morphological parsing, diacritic restoration, Synoptic style transfer, and span infilling.

It is the sibling of ainouche-abderahmane/koine-t5, which it improves on across every POS metric while carrying four additional tasks on the same backbone and the same adapter rank. Koine-T5 remains a valid model and its published numbers stand; this is not a deprecation.

Trained as part of SynoptiQ.

What it does

Results

Evaluated on the UD Ancient Greek PROIEL test split, held out from training and from checkpoint selection (selection used the dev split).

Against Koine-T5

Both models scored with identical code, split, decoding and tokenizer, restricted to tokens whose gold label does not depend on tagset convention (94.4% of the test set — see Comparison methodology).

Per-task (PROIEL dev / held-out slices)

Usage

from peft import PeftModel
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

tokenizer = AutoTokenizer.from_pretrained("bowphs/GreTa")
tokenizer.pad_token = "<pad>"
tokenizer.eos_token = "</s>"
tokenizer.add_special_tokens(
    {"additional_special_tokens": [f"<extra_id_{i}>" for i in range(100)]}
)

base = AutoModelForSeq2SeqLM.from_pretrained("bowphs/GreTa")
base.config.tie_word_embeddings = False
base.config.vocab_size = len(tokenizer)
model = PeftModel.from_pretrained(base, "ainouche-abderahmane/koine-t5-omni")

inputs = tokenizer("pos: μεγαλύνει ἡ ψυχή μου τὸν κύριον", return_tensors="pt")
out = model.generate(**inputs, max_new_tokens=128, num_beams=1, do_sample=False)
print(tokenizer.decode(out[0], skip_special_tokens=True).upper())
# V- RA N- RP RA N-

Three things matter for correct output:

• Do not add a [PAD] token or resize embeddings. GreTa ships <pad>=0 and </s>=1. Adding a pad token desynchronises the pad id from the decoder start id and collapses generation.
• Upper-case POS and morphology predictions. The GreTa tokenizer case-folds, so the model emits lowercase tags. MorphGNT codes are case-unique, making this lossless.
• Use greedy decoding (num_beams=1, do_sample=False) for the tagging and restoration tasks. Repetition penalties corrupt them: tags legitimately repeat, and restoration output largely repeats its input. Beam search suits the Synoptic task; contrastive search suits infilling only.

Decoding morphology output

Morphology is emitted in a compact encoding that omits MorphGNT's positional padding (N-----NSF- → N-NSF). This keeps targets inside the 256-token context. The mapping is injective over all 602 tags attested in MorphGNT, so it is lossless; morph_tags.json in this repository inverts it.

import json
table = json.load(open("morph_tags.json"))
table["N-NSF"]    # "N-----NSF-"

Training

Tasks are sampled per micro-batch slot in proportion to fixed weights, so every task appears in most optimiser steps and none is starved. Checkpoint selection is gated: a checkpoint must hold POS-NT and lemma above threshold before its secondary-task score can win, so selection optimises the model as a whole rather than the tagger alone. The best checkpoint occurred at step 64,000 of 75,000 — the model plateaued rather than being cut off.

Data

PROIEL's NonCommercial term propagates to this adapter, hence CC BY-NC-SA 4.0.

Known limitations

Both prefixes below exist in the checkpoint and produce output. Neither should be relied on.

normalize does not work. It scores 0.886 token accuracy on its held-out slice, but emitting the input unchanged scores approximately 0.90 on the same data: the model learned an identity mapping rather than the task. Crasis expansion fails on every case tried.

normalize: κἀγὼ εἶπον αὐτῷ        →  κἀγὼ εἶπον αὐτῷ        (expected καὶ ἐγὼ …)
normalize: κἂν ἀποθάνῃ ζήσεται    →  κἂν ἀποθάνῃ ζήσεται    (expected καὶ ἐάν …)

The cause is training corruption too sparse to make copying suboptimal — most tokens in a normalisation example are already correct, so the identity mapping is near-optimal under token accuracy. Any evaluation of a task of this shape needs a copy baseline reported alongside it.

gloss does not work. 0.130 token accuracy. Output is English-shaped but not correct:

gloss: ἐν ἀρχῇ ἦν ὁ λόγος  →  in the_one having_been_to_the_one the should

GreTa has no English pretraining, and word-level glossing over a largely single-occurrence vocabulary is not reachable from this backbone at this data scale. A Greek→English verse translation task was attempted earlier and abandoned for the same reason.

Classical Greek is weaker than Koine. Trained predominantly on New Testament Greek; Herodotus runs roughly 4 pp behind on token accuracy and substantially behind on exact match.

Lemma follows PROIEL/MorphGNT conventions, which differ from some lexica — for example first person plural pronouns lemmatise to ἐγώ rather than ἡμεῖς.

Morphology exact match is 0.21 at the sentence level while token accuracy is 0.885: long sentences reliably contain at least one of ten joint grammatical axes that is wrong.

Comparison methodology

The headline comparison masks tokens whose gold label depends on tagset convention. This is not cosmetic, and reproducing it naively will give different numbers.

PROIEL and MorphGNT disagree on high-frequency postpositive particles: MorphGNT treats δέ, γάρ, οὖν, μέν as conjunctions, PROIEL as adverbs. Koine-T5-Omni was trained on gold reconciled to the MorphGNT convention; Koine-T5 predates that reconciliation. Scoring both against reconciled gold therefore marks Koine-T5 wrong for answering as it was taught.

The affected tokens are only 5.6% of the test set, but they occur in 39.3% of NT and 85.5% of Classical sentences, so full-sentence exact match is affected far more than token accuracy. Unmasked, the pooled gap reads +8.6 pp token / +41.4 pp exact match. Masked, it is +3.2 / +3.6.

The diagnostic is that masking barely moves this model (0.9442 → 0.9419 pooled token accuracy) while it moves Koine-T5 substantially (0.8585 → 0.9097). The masked figures are reported here because they measure tagging skill rather than answer-key alignment.

Reproducing

The checkpoint was trained from tag omni-v1-training (commit 45958b6). Later commits on main remove the glossing task and rebuild normalisation, so main does not reproduce this checkpoint.

modal run modal/app_koine_t5_omni.py::train
modal run modal/app_koine_t5_omni.py::run_test     # held-out test numbers
modal run modal/app_koine_t5_omni.py::compare      # convention-neutral comparison

Citation

@software{ainouche_koine_t5_omni_2026,
  author = {Ainouche, Abderahmane},
  title  = {Koine-T5-Omni: A Multitask Ancient Greek Sequence-to-Sequence Model},
  year   = {2026},
  url    = {https://huggingface.co/ainouche-abderahmane/koine-t5-omni},
  note   = {LoRA adapter over bowphs/GreTa}
}

Acknowledgements

Built on GreTa (Riemenschneider & Frank). Training data from the PROIEL treebank, MorphGNT/SBLGNT, MACULA Greek, and the SynoptiQ corpus.