Token-classification tagger for Lithuanian: UPOS plus the morphological features that drive agreement (Case, Gender, Number, Tense, Person, Voice, Degree, VerbForm, Mood, Reflex), emitted as one combined label per word. Fine-tuned from EMBEDDIA/litlat-bert on MATAS v3.0 + UD_Lithuanian-ALKSNIS (~2.15M gold tokens).

Which model do I want?

model	best for	emits
this one	highest UPOS/slots accuracy, UD categories	UPOS (incl. DET/AUX) + core FEATS
…-morphology-full	complete annotations	full UD FEATS + lemmas
…-morphology-vdu	Lithuanian accentuation pipelines	traditional-grammar categories

Benchmarks (measured, full 684-sentence ALKSNIS gold test)

metric	this model	UDPipe 2 reference
slots¹	89.1%	89.2%
UPOS (same protocol)	92.5%	95.1%
CoNLL-18 official UPOS F1 (gold tokenization)	94.0	95.2
speed, tok/s (ONNX INT8 CPU vs network service)	874	605

¹ slots = exact match of the agreement-feature projection (POS family + Case/Gender/Number/Tense/Person/Voice/Degree) that accentuation-style disambiguation consumes; DET/PRON and AUX/VERB merge inside the projection.

Not emitted by design (use the -full model instead): FEATS keys beyond the ten above (Definite, Polarity, PronType, …) and lemmas.

Usage (verified)

python

import torch
from transformers import AutoTokenizer, AutoModelForTokenClassification

repo = "alexbalandi/litlat-bert-lithuanian-morphology"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForTokenClassification.from_pretrained(repo)

words = ["Tas", "namas", "yra", "gražus"]
enc = tok(words, is_split_into_words=True, return_tensors="pt")
with torch.inference_mode():
    logits = model(**enc).logits[0]
seen = set()
for pos, wid in enumerate(enc.word_ids(0)):
    if wid is not None and wid not in seen:
        seen.add(wid)
        print(words[wid], model.config.id2label[int(logits[pos].argmax())])
# Tas DET|Case=Nom|Gender=Masc|Number=Sing
# namas NOUN|Case=Nom|Gender=Masc|Number=Sing
# ...

An onnx/ folder with an INT8 model (same logits output) is included for CPU serving.

Training data & lineage

Gold corpora only, no NC-licensed model anywhere in the lineage: MATAS v3.0 (CC BY 4.0; UD features reconstructed deterministically from its Jablonskis XPOS) + UD_Lithuanian-ALKSNIS (CC BY-SA 4.0), remaining contextual gaps filled by iterated constrained-decoding self-training. Recipe and tooling: github.com/alexbalandi/kirciuokle (local/tagger-hf/, public domain).

License & attribution

CC BY-SA 4.0 (inherited from litlat-bert and ALKSNIS). Please credit: MATAS v3.0 (Rimkutė, Bielinskienė, Boizou, Dadurkevičius, Kovalevskaitė, Utka; CLARIN-LT), UD_Lithuanian-ALKSNIS (VDU), litlat-bert (Ulčar & Robnik-Šikonja, EMBEDDIA).