Access Request Terms:
By requesting access to this dataset, you confirm that you:
- will cite the sk-bench paper and any additional works listed in the Citation section of this dataset card, and respect the dataset's license;
- will ensure that your downstream use complies with applicable laws, regulations, and ethical AI principles.

Slohry — Slovak Grammar Benchmark

A focused inflectional-morphology benchmark for Slovak, covering four grammar tasks across ~6,000 items.

Dataset summary

Slohry tests a model's command of Slovak inflectional morphology across four sub-tasks, each as a separate Hugging Face configuration:

casovanie — verb conjugation: produce the form of a verb at a given mood/tense/person/number/gender.
stupnovanie — adjective/adverb comparison: produce the comparative or superlative.
urcenie_tvaru — form identification: given a surface form and lemma, name a single grammatical feature.
sklonovanie — noun declension: produce the form of a noun in a given case and number.

Each item is scored by exact string match against a single gold answer.

Supported tasks

`casovanie` (verb conjugation)

Given a verb lemma and a target description, produce the inflected form.

Example. Lemma schmatnúť, target v 1. osobe jednotného čísla mužského životného rodu v minulom čase → gold schmatol.

`stupnovanie` (adjective/adverb comparison)

Given an adjective or adverb lemma and a target degree (comparative / superlative), produce the inflected form.

Example. Lemma pekný, degree 2. stupeň → gold krajší.

`urcenie_tvaru` (form identification)

Given a surface form and its lemma, name a single grammatical feature.

Example. Form školám, lemma škola, feature pád (case) → gold datív.

`sklonovanie` (noun declension)

Given a noun lemma and a target description (case + number), produce the inflected form.

Example. Lemma škola, target v datíve množného čísla → gold školám.

Languages

Slovak (sk), specifically standard literary register as represented in the SNK / NLP4SK tagset.

Dataset structure

Sizes

Config	val	test	Total
`casovanie`	498	1002	1500
`stupnovanie`	502	998	1500
`urcenie_tvaru`	498	1002	1500
`sklonovanie`	498	1002	1500

Row schemas

casovanie

Column	Type	Example
`id`	int32	`1`
`lemma`	string	`"schmatnúť"`
`freq_bucket`	string	`"high"` / `"mid"` / `"low"`
`target_tag`	string	`"VLdsam"`
`target_description`	string	`"v 1. osobe jednotného čísla mužského životného rodu v minulom čase"`
`answer`	string	`"schmatol"`

sklonovanie

Column	Type	Example
`id`	int32	`1`
`lemma`	string	`"škola"`
`freq_bucket`	string	`"high"` / `"mid"` / `"low"`
`target_tag`	string	`"Sfp3"`
`target_description`	string	`"datíve množného čísla"`
`answer`	string	`"školám"`

stupnovanie

Column	Type	Example
`id`	int32	`1`
`lemma`	string	`"pekný"`
`freq_bucket`	string	`"high"` / `"mid"` / `"low"`
`degree`	string	`"comparative"` / `"superlative"`
`degree_label`	string	`"2. stupeň"` / `"3. stupeň"`
`source_pos`	string	`"A"` (adjective) / `"D"` (adverb)
`answer`	string	`"krajší"`

urcenie_tvaru

Column	Type	Example
`id`	int32	`1`
`lemma`	string	`"škola"`
`freq_bucket`	string	`"high"` / `"mid"` / `"low"`
`form`	string	`"školám"`
`source_tag`	string	`"Sfp3"`
`feature`	string	`"case"` / `"number"` / `"gender"` / `"degree"` / `"person"` / `"mood_tense"`
`feature_description_sk`	string	`"pád"`
`answer`	string	`"datív"`

Field semantics

freq_bucket — frequency tertile of the lemma in the source corpus: high (top third), mid (middle third), low (bottom third).
target_tag / source_tag — NLP4SK morphological tag for the target or source form (see Tag reference below).
target_description — human-readable Slovak description of the target (used as a prompt slot).
answer — gold answer. Scoring is exact string match (case-insensitive, whitespace-stripped, diacritic-sensitive).
id — sequential within each (config, split). Not unique across configs.

Tag reference

Slovak NLP4SK tagset positions used in this dataset:

Cases (digit at end of S/A tag): 1 nominatív · 2 genitív · 3 datív · 4 akuzatív · 5 vokatív · 6 lokál · 7 inštrumentál.
Number: s jednotné (singular) · p množné (plural).
Gender: m mužský životný · i mužský neživotný · f ženský · n stredný · h všeobecný · o neurčitý.
Degree (A/D): x 1. stupeň (positive) · y 2. stupeň (comparative) · z 3. stupeň (superlative).
Verb forma (V tag, position 2): I neurčitok · K prítomný čas · M rozkazovací spôsob · H prechodník · L minulý čas · B budúci čas.
Verb aspect (V tag, position 3): d dokonavý · e nedokonavý · j obojvidové.
Verb person (V tag, position 5): a prvá · b druhá · c tretia.

Data-shape note. The forma marker (position 2 in the canonical NLP4SK tagset) is dropped from S, A, and G tags in this dataset. All other positions shift one character to the left. V tags keep their forma marker.

Source data

The lemmas and morphological paradigms come from a re-export of NLP4SK / SNK lemma + form data covering ~10,000 lemmas of standard Slovak. Frequency tertiles are computed by sorting the source corpus by raw frequency and bucketing into thirds (high / mid / low). Sampling uses seed 42 with per-task seed offsets, stratified by tertile (500/500/500 per task; 250 lemmas × 2 items for stupnovanie). The full pipeline is reproducible from the build script in the source GitHub repository.

Prompt templates

The prompt_templates.json sidecar at the root of this repository contains three semantically equivalent Slovak paraphrases per task (v1 / v2 / v3). They are intended to be run independently and averaged across runs to reduce prompt-sensitivity noise.

One example per task (other variations live in prompt_templates.json):

casovanie v1: Uveď tvar slovesa „{lemma}" {target_description}. Odpovedz iba samotným tvarom slovesa.
stupnovanie v1: Uveď {degree_label} prídavného mena/príslovky „{lemma}". Odpovedz iba jedným slovom.
urcenie_tvaru v1: Uveď {feature_description_sk} tvaru „{form}" (lemma „{lemma}"). Odpovedz jedným slovom.
sklonovanie v1: Uveď tvar podstatného mena „{lemma}" v {target_description}. Odpovedz iba jedným slovom.

All variations are in singular informal Slovak (tykanie); all expect a one-word answer.

Tag parsing helper

To compute parsed features (case, number, gender, …) from target_tag / source_tag in a notebook, the following self-contained helper covers the tag shapes used in this dataset:

def parse_tag(tag: str) -> dict:
    """Parse an NLP4SK-derived morphological tag into named features."""
    pos = tag[0]
    if pos == "S":           # noun: S<gender><number><case>
        return {"pos": pos, "gender": tag[1], "number": tag[2], "case": tag[3]}
    if pos == "A":           # adj: A<gender><number><case><degree>
        return {"pos": pos, "gender": tag[1], "number": tag[2], "case": tag[3], "degree": tag[4]}
    if pos == "D":           # adv: D<degree>
        return {"pos": pos, "degree": tag[1]}
    if pos == "V":           # verb: V<forma><aspect>[<number><person>[<gender>]]
        out = {"pos": pos, "forma": tag[1], "aspect": tag[2]}
        if len(tag) >= 5:
            out["number"] = tag[3]
            out["person"] = tag[4]
        if len(tag) >= 6:
            out["gender"] = tag[5]
        return out
    return {"pos": pos, "raw": tag}

Limitations and biases

Slovak only. No coverage of other West Slavic languages or dialect variants.
Uneven POS coverage. The source corpus is skewed toward adjectives and nouns; verbs are less numerous; adverbs and participles are sparse. Sampling caps mitigate this but cannot fully balance.
Automatic morphology. Source paradigms come from NLP4SK's automatic morphological analysis, not gold-annotated by hand. A small fraction of paradigm cells may contain errors that propagate into items.
Raw frequency. Tertile assignment uses unnormalized corpus frequency; results are dataset-specific, not absolute claims about word commonness.

Citation

This dataset is part of sk-bench, a benchmark for evaluating large language models in Slovak, described in Šuppa et al. (2026) (arXiv:2610.09152). If you use this dataset, please cite:

@inproceedings{suppa-etal-2026-sk,
    title = "sk-bench: A Native-First Benchmark for Evaluating Large Language Models in {S}lovak",
    author = "{\v{S}}uppa, Marek  and
      Vykopal, Ivan  and
      Ridzik, Andrej  and
      Sopkovi{\v{c}}, Kristi{\'a}n  and
      K{\v{n}}a{\v{z}}ekov{\'a}, Nat{\'a}lia  and
      Kopcan, Jaroslav  and
      Blstak, Miroslav  and
      Ondrejov{\'a}, Vikt{\'o}ria  and
      Hl{\'a}dek, Daniel  and
      Gregor, Michal  and
      Tamajka, Martin  and
      Simko, Marian",
    editor = "Lee, Hung-Yi  and
      Dev, Sunipa  and
      Li, Junyi Jessy  and
      Elliott, Desmond",
    booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
    month = oct,
    year = "2026",
    address = "Budapest, Hungary",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.emnlp-main.2639/",
    doi = "10.18653/2026.emnlp-main.2639",
    pages = "60008--60055",
    ISBN = "979-8-89176-459-0"
}

License and acknowledgments

This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

Acknowledgments: built on top of the NLP4SK / Slovak National Corpus (SNK) lemma and paradigm data. When you use this dataset, please also credit NLP4SK and the Slovak National Corpus as the source of the lemmas and paradigms. Tagset documentation: http://arl6.library.sk/nlp4sk/help#chapter_pos.

Changelog

v1.0 (2026-08-27) — initial release.

Downloads last month: 108

Total file size:

200 kB

Collection including slovak-nlp/slohry

sk-bench

Collection

26 items • Updated Aug 25

Paper for slovak-nlp/slohry

sk-bench: A Native-First Benchmark for Evaluating Large Language Models in Slovak

Paper • 2610.09152 • Published Oct 6

You need to agree to share your Hugging Face information to access this dataset