SGaleshchuk/ULP-Ukrainian-Language-Proficiency

Ukrainian Language Proficiency (ULP) Benchmark

Dataset Description

The Ukrainian Language Proficiency (ULP) benchmark is an expert-curated dataset designed to evaluate Ukrainian language proficiency in Large Language Models (LLMs), with a focus on grammar and orthography as core components of language competence.

Dataset Summary

This gold-standard dataset contains 347 multiple-choice questions prepared by professional linguists.

The benchmark evaluates:

• Grammar: morphology, syntax, and lexical norms
• Orthography: Ukrainian spelling and orthographic conventions

Supported Tasks

• Multiple-choice question answering: Models must identify the correct answer from 4-5 options

Languages

• Ukrainian (uk)

Dataset Structure

Data Instances

Each instance contains:

• question: The question text in Ukrainian
• choices: List of answer options (4-5 options)
• answer: Integer index of the correct answer (0-indexed)
• answer_letter: Ukrainian alphabet letter corresponding to the correct answer (А, Б, В, Г, etc., excluding Ґ)
• debug_id: Unique identifier for the question

Example:

{
  "question": "Вищий і найвищий ступені порівняння можна утворити від усіх прикметників у рядку:",
  "choices": [
    "розумний, високий, короткий",
    "сприятливий, всесвітній, добрий",
    "вірний, хворий, салатовий",
    "активний, хитрющий, демократичний."
  ],
  "answer": 0,
  "debug_id": "1",
  "answer_letter": "А"
  "category": "граматика"
}

Data Fields

• question (string): The test question in Ukrainian
• choices (list of strings): 4-5 answer options
• answer (int): Index of the correct answer (0-indexed)
• debug_id (string): Unique question identifier
• answer_letter (string): Ukrainian letter mapping (А=0, Б=1, В=2, Г=3, Д=4, etc.)
• category (string): type of question by category, e.g., grammar, vocabulary, orthography. Note that while syntax is a part of grammar, we still keep it as a category apart for more granular analysis.

Data Splits

The dataset contains 347 questions in total. No predefined train/test split is provided, as this is designed as an evaluation benchmark.

Dataset Creation

Curation Rationale

This benchmark focuses on normative Ukrainian usage and targets grammatical and orthographic competence at the level expected of native speakers.

Source Data

Questions were compiled from authoritative normative and reference sources.

Annotations

Annotation Process

• Created by professional philologists and linguists
• Each question reviewed by at least one additional expert
• Ambiguous or insufficiently comprehensive items were revised or removed

Professional linguists with 10-30 years of experience teaching Ukrainian language courses.

Citation

If you use this dataset in your research, please cite:

@article{galeshchuktoward,
  title={Toward a Gold-Standard Benchmark for Evaluating Ukrainian Language Proficiency in LLMs},
  author={Galeshchuk, Svitlana and Maksymiuk, Yuliia and Chernobrov, Yuliia and Antoniv, Oleksandra and Stankevych, Nina and Faryna, Nataliia and Popkova, Oksana}
year={2026},
  url={https://unlp.org.ua/wp-content/uploads/2026/05/toward-a-gold-standard-benchmark-for-evaluating-ukrainian-language-proficiency-in-llms.pdf}
}