KIEFERSA/greek-nlu-bench

Greek NLU Benchmark

A Modern Greek language-understanding benchmark with 1,071 items across 18 tasks, covering grammar, comprehension, meaning, discourse, idioms and translation. Each task provides development and test splits.

Evaluation code: Kiefer-Tek/greek-nlu-bench
Leaderboard: Greek NLU Benchmark Leaderboard

Development items support prompt and integration checks. Test items support final comparisons after evaluation settings are fixed. Keep both splits out of training.

Dataset overview

Task	Dev	Test	Total
acceptability	38	13	51
assertion_reason	33	10	43
cloze_open	18	6	24
cloze_select	63	12	75
coreference	32	9	41
extractive_qa	70	18	88
idiom_appropriateness	27	11	38
idiom_open_cloze	13	6	19
machine_translation	71	19	90
masked_token	10	11	21
matching	38	15	53
mcq	64	21	85
nli	78	23	101
paraphrase	72	18	90
sentiment	71	18	89
sequence_ordering	14	7	21
sts	51	5	56
wsd	68	18	86
Total	831	240	1,071

Data fields

Field	Description
id	Unique item identifier
task	Task configuration
language	Language code (el)
version	Item version
tags	Category, linguistic phenomenon, difficulty, domain and split
input	Task-specific question, text or choices
gold	Reference answer, label, score or ordering
provenance	Item source metadata
annotation	Verification status and annotation metadata

Choice indices and sentence-order indices are zero-based. Open completion tasks provide accepted answer forms. Extractive QA uses an empty answer list for unanswerable questions. Translation is reported as a separate track.

Usage

from datasets import load_dataset

dataset = load_dataset("KIEFERSA/greek-nlu-bench", "nli", token=True)
dev = dataset["dev"]
test = dataset["test"]

Access requires an authenticated account with permission to read this private dataset. For reproducible comparisons, pin the dataset revision and use the same items and scoring settings. Keep evaluation items out of training data.

Evaluation

The NLU summary averages task-level headline scores with equal weight per task, excluding machine translation. Classification tasks use accuracy; completion tasks use normalized exact match; QA uses token F1; ordering uses normalized Kendall tau; matching uses pair-set Jaccard; STS uses Pearson and Spearman correlation; translation uses chrF. Greek answer matching accounts for accents according to the tested phenomenon and normalizes final sigma. Category, phenomenon and difficulty tags support grouped analysis.

© KIEFER TEK LTD.