MORFES / README.md
kperros-kiefer's picture
Add arXiv paper link
3fdaca5 verified
|
Raw
History Blame Contribute Delete
10.3 kB
metadata
license: cc-by-4.0
language:
  - el
pretty_name: MORFES
size_categories:
  - n<1K
annotations_creators:
  - expert-generated
language_creators:
  - expert-generated
source_datasets:
  - original
task_categories:
  - multiple-choice
  - text-generation
tags:
  - morphology
  - inflection
  - modern-greek
  - greek
  - linguistics
  - evaluation
configs:
  - config_name: default
    data_files:
      - split: test
        path: morfes.csv
Sophea AI
MORFES
Benchmark for Productive Inflectional Competence in Modern Greek

MORFES (Morphological Open-class Recognition-and-Formation Evaluation Suite; also μορφές, Greek for "forms") is a benchmark of 500 expert-verified items testing productive inflectional competence in Modern Greek.

  • Task: Modern Greek inflectional morphology (recognition and production)
  • Languages: Greek
  • Size: 500 items
  • License: CC-BY-4.0
Resources
MORFES is introduced and described in the accompanying paper.
Sophea-Genesis-1, an open Greek language model, is released alongside it.
On MORFES it is the most accurate of the open models we evaluated.

Given a lemma (dictionary form) and a grammatical specification, each item probes whether a model can recognize and produce the correct inflected form. MORFES favors lower-frequency lemmas so that a correct answer reflects the productive rule rather than a memorized form. The full construct, item selection, frequency control, and evaluation are described in the accompanying paper.

Categories

The 500 items span six categories: the three richly inflecting open classes (noun, adjective, verb), each in a single-form and a full-paradigm variant.

Category Items
Noun declension: single form 50
Noun declension: full paradigm 50
Adjective declension: single form 50
Adjective declension: full paradigm 50
Verb conjugation: single form 150
Verb conjugation: full paradigm 150

A single-form item asks for one cell of a paradigm (for example the genitive singular of a noun); a full-paradigm item asks for a complete sub-table (for example all six persons of a verb tense). The set is verb-weighted, with 300 of the 500 items verbs: the verb is the most richly inflecting of the three, with far more distinct forms than a noun or adjective, so an equal split would under-sample the open class that inflects the most.

Data fields

One row per item, 10 columns:

Column Description
id Stable item identifier, for example gr-00001.
category One of the six categories above.
question The item prompt, in Greek.
A, B, C, D The four candidate forms. Their order is randomized per item.
answer The letter (A to D) of the correct option.
accepted_answers The correct surface form(s), as a JSON string: a flat list for single-form items (for example ["θέρισαν", "θερίσανε"], the aorist third-person plural of «θερίζω», "to reap"), and a list of per-cell lists for full-paradigm items. Greek often admits more than one correct form for a single specification.
lemma The base (dictionary) form the item is built on. One distinct lemma per item.

The file is a UTF-8 CSV; standard loaders read it directly:

from datasets import load_dataset

ds = load_dataset("KIEFERSA/MORFES", split="test")

Contamination control

Each item's lemma is provided so users can hold these lemmas out of a model's training data and guard against contamination. Every item is built on a distinct lemma, and the requested form is never the lemma's dictionary form, so no item can be answered by repeating a word from the prompt.

Citation

If you use MORFES, please cite:

@misc{perros2026morfes,
  title        = {{MORFES}: A Benchmark for Productive Inflectional Competence in Modern Greek},
  author       = {Perros, Ioakeim and Papadopoulou, Cleopatra and Kirouane, Ayoub and Petrocheilos, Christos},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/KIEFERSA/MORFES}},
  note         = {arXiv:2607.28274}
}

License

MORFES is released under the Creative Commons Attribution 4.0 International (CC-BY-4.0) license. You may share and adapt the data, including for commercial use, provided you give appropriate attribution; please attribute by citing the paper above.