| --- |
| license: other |
| license_name: zno-eval-mixed-rights |
| license_link: LICENSE |
| language: |
| - uk |
| pretty_name: ZNO-Eval |
| task_categories: |
| - question-answering |
| - multiple-choice |
| configs: |
| - config_name: default |
| default: true |
| data_files: |
| - split: all |
| path: data/tasks/all.jsonl.gz |
| - split: ukrainian |
| path: data/tasks/ukrainian.jsonl.gz |
| - split: history |
| path: data/tasks/history.jsonl.gz |
| - split: geography |
| path: data/tasks/geography.jsonl.gz |
| - split: math |
| path: data/tasks/math.jsonl.gz |
| - config_name: exams |
| data_files: |
| - split: exams |
| path: data/exams/exams.jsonl.gz |
| - config_name: issues |
| data_files: |
| - split: issues |
| path: data/issues/issues.jsonl.gz |
| --- |
| |
| # ZNO-Eval - archival release |
|
|
| ### Note: This dataset freezes the original ZNO-Eval task payloads (https://github.com/NLPForUA/ZNO) for reproducibility. It is not the recommended starting point for new evaluations: use [ZNO-Eval v2](https://huggingface.co/datasets/NLPForUA/zno-eval-v2), which repairs and annotates the major known problems without rewriting this archive. |
|
|
| The limitation is specifically that v1 does not reliably expose all source media, so those rows are often unavailable to a controlled **text-only** benchmark and may also be incomplete for multimodal use until reconstructed from the source. |
|
|
| ## Release in progress. Stay tuned. |
|
|
| ## Loading |
|
|
| ```python |
| from datasets import load_dataset |
| |
| all_tasks = load_dataset("NLPForUA/zno-eval", split="all") |
| ukrainian = load_dataset("NLPForUA/zno-eval", split="ukrainian") |
| exams = load_dataset("NLPForUA/zno-eval", "exams", split="exams") |
| issues = load_dataset("NLPForUA/zno-eval", "issues", split="issues") |
| ``` |
|
|
| `all` is the deterministic concatenation of `ukrainian`, `history`, `geography`, and `math`. Subjects are dataset partitions, not train/validation/test splits. |
|
|
| ## Known limitations reviewed against v2 |
|
|
| Counts below are affected task counts and overlap. Do not add them to estimate unique affected tasks. The complete row-level evidence is queryable through the `issues` configuration. |
|
|
| | Issue | Affected v1 tasks | Interpretation | |
| | --- | ---: | --- | |
| | Source has media although `with_photo=false` | 105 | Unsafe for controlled text-only evaluation without a verified surrogate. | |
| | Answer-option media is not represented by legacy media fields | 193 | V1 can omit visual answer content. | |
| | Source has multiple media items | 231 | The single-primary-URL legacy shape is incomplete. | |
| | Blank answer text | 113 | Often media-bearing or incompletely parsed options. | |
| | Declared answer labels without answer records | 151 | Structural parser/data loss. | |
| | Noncanonical exact duplicates | 73 | Exclude for canonical sampling, retain for provenance. | |
| | Time-sensitive review candidates | 71 | Review historical answer context before prompting. | |
| | Confirmed time-sensitive tasks | 1 | Supply exam-time context or exclude. | |
| | Not machine gradable after review | 64 | Retain for provenance, not automatic scoring. | |
| | Not eligible for the reviewed canonical text-only protocol | 1,393 | This does not imply multimodal or alternative-protocol unavailability. | |
|
|
| The review found 660 v1 tasks with confirmed v2 corrections, represented by 746 task-level correction rows. |
|
|
| ## Task schema |
|
|
| Every row contains the original v1 task payload unchanged plus four identity/join fields. |
|
|
| | Field | Type | Meaning | |
| | --- | --- | --- | |
| | `uid` | string | Stable task key `subject:test_id:task_id`. | |
| | `test_uid` | string | Foreign key to `exams`, `subject:test_id`. | |
| | `subject` | string | `ukrainian`, `history`, `geography`, or `math`. | |
| | `test_id` | string | Source test identifier. | |
| | `task_id` | integer | Original zero-based task position. | |
| | `question` | string | Legacy parsed question text. | |
| | `answers` | list of objects | Legacy option records containing `answer` and `text`. | |
| | `answer_vheader` | list or null | Legacy permitted response labels. | |
| | `answer_hheader` | list or null | Legacy matching/sequence slots. | |
| | `correct_answer` | list of strings | Legacy gold codes or literal values. | |
| | `comment` | string | Legacy source/parser comment. | |
| | `with_photo` | boolean | Legacy parser flag; not a complete media-availability signal. | |
| | `photo_url` | string or null | Legacy primary image URL when recorded. | |
|
|
| ### Raw (parsed) data format: |
| #### Question with one correct answer: |
| ```json |
| { |
| "task_id": 42, |
| "question": "Образ Прометея символізує силу й нескореність народу у творі Тараса Шевченка", |
| "answers": [ |
| { |
| "answer": "А", |
| "text": "«Сон» («У всякого своя доля...»)" |
| }, |
| { |
| "answer": "Б", |
| "text": "«Кавказ»" |
| }, |
| { |
| "answer": "В", |
| "text": "«Гайдамаки»" |
| }, |
| { |
| "answer": "Г", |
| "text": "«І мертвим, і живим...»" |
| }, |
| { |
| "answer": "Д", |
| "text": "«Ісаія. Глава 35»" |
| } |
| ], |
| "answer_vheader": [ |
| "А", |
| "Б", |
| "В", |
| "Г", |
| "Д" |
| ], |
| "answer_hheader": [], |
| "correct_answer": [ |
| "Б" |
| ], |
| "comment": "", |
| "with_photo": false |
| }, |
| ``` |
| #### Question with multiple correct answers (each number from `answer_hheader` corresponds to letter from `answer_vheader`): |
| ```json |
| { |
| "task_id": 48, |
| "question": "Укажіть послідовність дій, необхідних для того, щоб зорієнтуватися на місцевості за Полярною зіркою.", |
| "answers": [ |
| { |
| "answer": "А", |
| "text": "подумки з’єднати прямою лінією дві зірки, що знаходяться на краю «ковша»" |
| }, |
| { |
| "answer": "Б", |
| "text": "продовжити уявну пряму лінію і відкласти на ній приблизно п’ять відрізків, рівних відстані між двома зірками на краю «ковша»" |
| }, |
| { |
| "answer": "В", |
| "text": "визначити яскраву зірку в сузір’ї Малої Ведмедиці, яка вказує напрям на північ" |
| }, |
| { |
| "answer": "Г", |
| "text": "знайти сузір’я Великої Ведмедиці (Великого Возу), яке нагадує ківш" |
| } |
| ], |
| "answer_vheader": [ |
| "А", |
| "Б", |
| "В", |
| "Г" |
| ], |
| "answer_hheader": [ |
| "1", |
| "2", |
| "3", |
| "4" |
| ], |
| "correct_answer": [ |
| "Г", |
| "А", |
| "Б", |
| "В" |
| ], |
| "comment": "", |
| "with_photo": false |
| }, |
| ``` |
|
|
| #### Sample prompting ([NLPForUA/gemma-2-it-zno-cot-with-topic](https://huggingface.co/NLPForUA/gemma-2-it-zno-cot-with-topic)): |
| ``` |
| <bos><start_of_turn>user |
| Дайте розгорнуту відповідь на завдання, починаючи з ключового слова "Відповідь:" та використовуючи лише наведені нижче варіанти. |
| |
| Завдання: З’ясуйте, якими частинами мови є виділені слова в реченні (цифра позначає наступне слово). |
| Сучасна людина, щоб бути (1)успішною, має вчитися (2)впродовж (3)усього життя, (4)опановуючи нові галузі знань. |
| |
| Варіанти відповіді: |
| А – займенник |
| Б – прикметник |
| В – форма дієслова (дієприкметник) |
| Г – форма дієслова (дієприслівник) |
| Д – прийменник<end_of_turn> |
| <start_of_turn>model |
| Відповідь: |
| ТЕМА: Морфологія. Частини мови |
| Завдання перевіряє ваше вміння визначати частини мови. |
| Необхідно бути дуже уважним, тому що частиномовна приналежність конкретного слова часто залежить від контексту. Проаналізуймо кожне із запропонованих слів. |
| Слово успішною відповідає на питання яка? та вказує на ознаку предмета. Отже, це прикметник. |
| Слово впродовж не відповідає на питання, а лише поєднує слова в реченні. Отже, це прийменник. |
| Слово усього відповідає на питання що? і вказує на предмет, але не називає його. Отже, це займенник. |
| А от слово опановуючи відповідає на питання що роблячи? та вказує на додаткову дію підмета. Отже, це особлива форма дієслова дієприслівник. |
| Відповідь: 1–Б, 2–Д, 3–А, 4–Г.<end_of_turn> |
| ``` |
|
|
| ## Exams table |
|
|
| | Field | Type | Meaning | |
| | --- | --- | --- | |
| | `test_uid` | string | Primary exam key. | |
| | `subject` | string | Exam subject. | |
| | `test_id` | string | Source test identifier. | |
| | `source_url` | string | Source exam page. | |
| | `num_tasks` | integer | Number of frozen v1 task rows in the exam. | |
|
|
| ## Issues table |
|
|
| Each row identifies one issue type for one task or catalog target. |
|
|
| | Field | Type | Meaning | |
| | --- | --- | --- | |
| | `target_uid` | string | Affected task or exam/catalog key. | |
| | `scope` | string | `task` or `exam_catalog`. | |
| | `subject` | string | Subject join field. | |
| | `test_uid` | string | Exam join field. | |
| | `test_id` | string | Source test identifier. | |
| | `task_id` | integer or null | Task position; null for exam-level issues. | |
| | `issue_type` | string | Stable machine-readable issue class. | |
| | `status` | string | `confirmed` or `review_candidate`. | |
| | `severity` | string | Issue family such as `data_error`, `text_only_limitation`, or `evaluation_risk`. | |
| | `affected_field` | string | V1 field or logical content area involved. | |
| | `affected_count` | integer | Number of affected subitems represented by the row. | |
| | `evidence` | string | Concise classification reason. | |
| | `v2_action` | string | How v2 repairs, annotates, or filters the case. | |
| | `v2_correction_ids` | list of strings | Join keys into v2 `changes` when applicable. | |
| | `canonical_uid` | string or null | Canonical task for duplicate-related rows. | |
| | `details_json` | string | Canonical JSON for issue-specific details. | |
|
|
|
|
| ## License |
|
|
| Original ZNO and NMT examination statements, answer options, answer keys and examination media originate from publicly released Ukrainian standardized examinations. |
| Copyright and related rights, where applicable, remain with the Ukrainian Center for Educational Quality Assessment, the relevant commissioned authors, and other identified rights holders. |
| These materials are included for research, education, public-sector evaluation and reproducibility, with source attribution, relying on their official publication and applicable limitations and exceptions under Ukrainian law. |
| NLPForUA does not grant rights it does not own. Commercial users should assess their intended use and obtain any additional permissions required from the relevant rights holders. |
|
|
| ## Citation |
|
|
| If you use ZNO-Eval, including either release, please cite the original benchmark paper: |
|
|
| ```bibtex |
| @article{zno-eval-2024, |
| title = {ZNO-Eval: Benchmarking reasoning capabilities of large language models in Ukrainian}, |
| author = {Syromiatnikov, M. V. and Ruvinskaya, V. M. and Troynina, A. S.}, |
| journal = {Informatics. Culture. Technology}, |
| publisher = {Odesa Polytechnic National University}, |
| number = {1}, |
| pages = {186--191}, |
| year = {2024}, |
| month = sep, |
| doi = {10.15276/ict.01.2024.27}, |
| url = {https://arxiv.org/abs/2501.06715} |
| } |
| ``` |
