Datasets:
Update README.md
Browse files
README.md
CHANGED
|
@@ -224,66 +224,48 @@ dataset_info:
|
|
| 224 |
dtype: string
|
| 225 |
---
|
| 226 |
|
| 227 |
-
# ZNO-Eval
|
| 228 |
|
| 229 |
-
|
| 230 |
|
| 231 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 232 |
|
| 233 |
## Loading
|
| 234 |
|
| 235 |
```python
|
| 236 |
from datasets import load_dataset
|
| 237 |
|
| 238 |
-
all_tasks
|
| 239 |
-
|
| 240 |
-
|
| 241 |
-
|
|
|
|
|
|
|
| 242 |
```
|
| 243 |
|
| 244 |
-
|
| 245 |
-
|
| 246 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 247 |
|
| 248 |
-
|
| 249 |
-
|
| 250 |
-
| Issue | Affected v1 tasks | Interpretation |
|
| 251 |
-
| --- | ---: | --- |
|
| 252 |
-
| Source has media although `with_photo=false` | 105 | Unsafe for controlled text-only evaluation without a verified surrogate. |
|
| 253 |
-
| Answer-option media is not represented by legacy media fields | 193 | V1 can omit visual answer content. |
|
| 254 |
-
| Source has multiple media items | 231 | The single-primary-URL legacy shape is incomplete. |
|
| 255 |
-
| Blank answer text | 113 | Often media-bearing or incompletely parsed options. |
|
| 256 |
-
| Declared answer labels without answer records | 151 | Structural parser/data loss. |
|
| 257 |
-
| Noncanonical exact duplicates | 73 | Exclude for canonical sampling, retain for provenance. |
|
| 258 |
-
| Time-sensitive review candidates | 71 | Review historical answer context before prompting. |
|
| 259 |
-
| Confirmed time-sensitive tasks | 1 | Supply exam-time context or exclude. |
|
| 260 |
-
| Not machine gradable after review | 64 | Retain for provenance, not automatic scoring. |
|
| 261 |
-
| Not eligible for the reviewed canonical text-only protocol | 1,393 | This does not imply multimodal or alternative-protocol unavailability. |
|
| 262 |
|
| 263 |
-
The review found 660 v1 tasks with confirmed v2 corrections, represented by 746 task-level correction rows.
|
| 264 |
|
| 265 |
-
##
|
| 266 |
-
|
| 267 |
-
Every row contains the original v1 task payload unchanged plus four identity/join fields.
|
| 268 |
-
|
| 269 |
-
| Field | Type | Meaning |
|
| 270 |
-
| --- | --- | --- |
|
| 271 |
-
| `uid` | string | Stable task key `subject:test_id:task_id`. |
|
| 272 |
-
| `test_uid` | string | Foreign key to `exams`, `subject:test_id`. |
|
| 273 |
-
| `subject` | string | `ukrainian`, `history`, `geography`, or `math`. |
|
| 274 |
-
| `test_id` | string | Source test identifier. |
|
| 275 |
-
| `task_id` | integer | Original zero-based task position. |
|
| 276 |
-
| `question` | string | Legacy parsed question text. |
|
| 277 |
-
| `answers` | list of objects | Legacy option records containing `answer` and `text`. |
|
| 278 |
-
| `answer_vheader` | list or null | Legacy permitted response labels. |
|
| 279 |
-
| `answer_hheader` | list or null | Legacy matching/sequence slots. |
|
| 280 |
-
| `correct_answer` | list of strings | Legacy gold codes or literal values. |
|
| 281 |
-
| `comment` | string | Legacy source/parser comment. |
|
| 282 |
-
| `with_photo` | boolean | Legacy parser flag; not a complete media-availability signal. |
|
| 283 |
-
| `photo_url` | string or null | Legacy primary image URL when recorded. |
|
| 284 |
-
|
| 285 |
-
### Raw (parsed) data format:
|
| 286 |
-
#### Question with one correct answer:
|
| 287 |
```json
|
| 288 |
{
|
| 289 |
"task_id": 42,
|
|
@@ -325,7 +307,7 @@ Every row contains the original v1 task payload unchanged plus four identity/joi
|
|
| 325 |
"with_photo": false
|
| 326 |
},
|
| 327 |
```
|
| 328 |
-
###
|
| 329 |
```json
|
| 330 |
{
|
| 331 |
"task_id": 48,
|
|
@@ -371,7 +353,118 @@ Every row contains the original v1 task payload unchanged plus four identity/joi
|
|
| 371 |
},
|
| 372 |
```
|
| 373 |
|
| 374 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 375 |
```
|
| 376 |
<bos><start_of_turn>user
|
| 377 |
Дайте розгорнуту відповідь на завдання, починаючи з ключового слова "Відповідь:" та використовуючи лише наведені нижче варіанти.
|
|
@@ -397,62 +490,213 @@ Every row contains the original v1 task payload unchanged plus four identity/joi
|
|
| 397 |
Відповідь: 1–Б, 2–Д, 3–А, 4–Г.<end_of_turn>
|
| 398 |
```
|
| 399 |
|
| 400 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 401 |
|
| 402 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 403 |
| --- | --- | --- |
|
| 404 |
-
|
|
| 405 |
-
| `
|
| 406 |
-
|
|
| 407 |
-
|
|
| 408 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 409 |
|
| 410 |
-
|
| 411 |
|
| 412 |
-
|
| 413 |
|
| 414 |
-
|
|
| 415 |
| --- | --- | --- |
|
| 416 |
-
|
|
| 417 |
-
|
|
| 418 |
-
| `
|
| 419 |
-
| `
|
| 420 |
-
|
|
| 421 |
-
|
|
| 422 |
-
| `
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 423 |
| `status` | string | `confirmed` or `review_candidate`. |
|
| 424 |
-
| `severity` | string |
|
| 425 |
-
| `affected_field` | string |
|
| 426 |
-
| `affected_count` | integer | Number of affected
|
| 427 |
-
| `evidence` | string |
|
| 428 |
-
| `v2_action` | string | How v2 repairs, annotates
|
| 429 |
-
| `v2_correction_ids` | list of strings |
|
| 430 |
-
| `
|
| 431 |
-
| `
|
|
|
|
|
|
|
|
|
|
|
|
|
| 432 |
|
| 433 |
|
| 434 |
-
##
|
| 435 |
|
| 436 |
-
|
| 437 |
-
|
| 438 |
-
|
| 439 |
-
|
| 440 |
|
| 441 |
-
## Citation
|
| 442 |
|
| 443 |
-
|
| 444 |
|
| 445 |
```bibtex
|
| 446 |
@article{zno-eval-2024,
|
| 447 |
-
title
|
| 448 |
-
author
|
| 449 |
-
journal
|
| 450 |
-
|
| 451 |
-
|
| 452 |
-
|
| 453 |
-
|
| 454 |
-
|
| 455 |
-
|
| 456 |
-
url = {https://arxiv.org/abs/2501.06715}
|
| 457 |
}
|
| 458 |
-
```
|
|
|
|
| 224 |
dtype: string
|
| 225 |
---
|
| 226 |
|
| 227 |
+
# ZNO-Eval v1 — archival release
|
| 228 |
|
| 229 |
+
ZNO-Eval is a benchmark of real Ukrainian school-leaving and university-entrance exam tasks: the External Independent Evaluation (ZNO) and the National Multi-subject Test (NMT). It was introduced in [ZNO-Eval: Benchmarking reasoning capabilities of large language models in Ukrainian](https://arxiv.org/abs/2501.06715).
|
| 230 |
|
| 231 |
+
This dataset is **v1, the archival release**. It preserves the task set originally published at [NLPForUA/ZNO](https://github.com/NLPForUA/ZNO/tree/main/tests) so that earlier results stay reproducible, identifies and corrects some of the mistakes introduced while parsing and hand-transcribing the tasks, and adds review metadata..
|
| 232 |
+
|
| 233 |
+
**For new evaluations use [ZNO-Eval v2](https://huggingface.co/datasets/NLPForUA/zno-eval-v2)**, where all identified source defects are repaired, missing answers and media are recovered, and the schema represents multi-image and table tasks.
|
| 234 |
+
|
| 235 |
+
| | |
|
| 236 |
+
| --- | --- |
|
| 237 |
+
| Tasks | 7,380 from 138 exam forms (118 ZNO, 20 NMT), exam years 2007–2024 |
|
| 238 |
+
| Subjects | Ukrainian language and literature 2,722 · History of Ukraine 2,640 · Geography 1,788 · Mathematics 230 |
|
| 239 |
+
| Question formats | single choice 5,700 · matching 892 · multi-select 324 · open answer 260 · sequence 204 |
|
| 240 |
+
| Text-only subset | 5,806 tasks usable without images and without a known payload correction |
|
| 241 |
+
| Differences from GitHub | 152 tasks patched (278 field edits); metadata columns added |
|
| 242 |
|
| 243 |
## Loading
|
| 244 |
|
| 245 |
```python
|
| 246 |
from datasets import load_dataset
|
| 247 |
|
| 248 |
+
all_tasks = load_dataset("NLPForUA/zno-eval", split="all") # 7,380 tasks
|
| 249 |
+
math = load_dataset("NLPForUA/zno-eval", split="math") # one subject
|
| 250 |
+
text_ready = load_dataset("NLPForUA/zno-eval", "text_benchmark", split="test") # soon (5,806 tasks)
|
| 251 |
+
exams = load_dataset("NLPForUA/zno-eval", "exams", split="exams")
|
| 252 |
+
topics = load_dataset("NLPForUA/zno-eval", "topics", split="topics")
|
| 253 |
+
issues = load_dataset("NLPForUA/zno-eval", "issues", split="issues")
|
| 254 |
```
|
| 255 |
|
| 256 |
+
| Configuration | Splits | Rows | Content |
|
| 257 |
+
| --- | --- | ---: | --- |
|
| 258 |
+
| `default` | `ukrainian`, `history`, `geography`, `math` | 7,380 | Every task. `split="all"` concatenates the four subjects. Subjects are partitions, not train/test splits. |
|
| 259 |
+
| `text_benchmark` | `test` | 5,806 | Tasks with `benchmark_status` of `eligible` or `requires_context`. Same columns as `default`. The split name does not imply a held-out design. |
|
| 260 |
+
| `exams` | `exams` | 138 | One row per exam form, with official-source references. |
|
| 261 |
+
| `topics` | `topics` | 68 | Controlled topic vocabulary. |
|
| 262 |
+
| `issues` | `issues` | 3,052 | Known limitations of the v1 payload, one row per task and issue class. |
|
| 263 |
|
| 264 |
+
Join tasks to `exams` on `exam_uid`, and each value of `topic_ids` to `topics.topic_id`.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 265 |
|
|
|
|
| 266 |
|
| 267 |
+
## Raw data format:
|
| 268 |
+
### Question with one correct answer:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 269 |
```json
|
| 270 |
{
|
| 271 |
"task_id": 42,
|
|
|
|
| 307 |
"with_photo": false
|
| 308 |
},
|
| 309 |
```
|
| 310 |
+
### Matching/sequence questions with multiple correct answers (each number from `answer_hheader` corresponds to letter from `answer_vheader`):
|
| 311 |
```json
|
| 312 |
{
|
| 313 |
"task_id": 48,
|
|
|
|
| 353 |
},
|
| 354 |
```
|
| 355 |
|
| 356 |
+
|
| 357 |
+
## What v1 changes relative to the GitHub copy
|
| 358 |
+
|
| 359 |
+
The reference point is commit `35d4c976599243b08dcd04303b557a7c2dc86068` of the GitHub repository. Against it, v1 differs in two ways.
|
| 360 |
+
|
| 361 |
+
|
| 362 |
+
**1. 152 tasks have corrected content** (278 field edits: 100 mathematics, 31 Ukrainian, 21 geography tasks). Six more tasks have corrected annotations only. All 158 carry the quality flag `v1_patched`, and every edit is listed with its before and after value in `reports/v1-release-patches.jsonl`, so the GitHub payload can be restored exactly by applying the `before` values.
|
| 363 |
+
|
| 364 |
+
| Repair | Tasks | Fields | Example (before → after) |
|
| 365 |
+
| --- | ---: | ---: | --- |
|
| 366 |
+
| Minus and subtraction sign in mathematics unified to `−` (U+2212) | 38 | 61 | `[‑4; 4]` → `[−4; 4]`; `-4` → `−4` |
|
| 367 |
+
| Highlighting that the task depends on, restored as `*word*` | 23 | 62 | `залила імла і несміливі` → `залила імла *і* несміливі` |
|
| 368 |
+
| Degree sign in mathematics | 22 | 47 | `∠ADB=35∘` → `∠ADB=35°` |
|
| 369 |
+
| Letter «о» left where the parser lost a superscript degree sign | 22 | 30 | `+10 оС` → `+10°С` |
|
| 370 |
+
| Hand-transcription errors in mathematics | 10 | 14 | `a=−2*1/3` → `a=−2\frac{1}{3}`; `AO=6` → `AO=√6`; `17` → `\frac{1}{7}` |
|
| 371 |
+
| Latin letters typed inside Ukrainian words in mathematics | 9 | 11 | `зміннi` (Latin `i`) → `змінні`; `бeзе` (Latin `e`) → `безе` |
|
| 372 |
+
| Latin `i` typed for the conjunction «і» in mathematics | 9 | 9 | `циліндр i його` → `циліндр і його` |
|
| 373 |
+
| Roman numerals typed with Cyrillic letters in mathematics options | 8 | 24 | `лише І та ІІ` → `лише I та II` |
|
| 374 |
+
| Typing slips in hand-typed text | 7 | 7 | `квартира n№ 31` → `квартира № 31` |
|
| 375 |
+
| Gold number format | 5 | 5 | `-3.5` → `-3,5`; `–0,2` → `-0,2` |
|
| 376 |
+
| Labels and options mistranscribed from images of sentence tasks | 4 | 7 | `5 - 1` → `4 - 1` |
|
| 377 |
+
| Invisible characters in gold formulas | 3 | 3 | U+2061 removed from `cosβ` |
|
| 378 |
+
| Stray line breaks, apostrophes, look-alike symbols | 9 | 9 | `<3^x+6⏎.` → `<3^x+6.`; `З см` (Cyrillic letter) → `3 см` |
|
| 379 |
+
|
| 380 |
+
Rows overlap: one field can receive several repairs, so the columns do not sum to the totals.
|
| 381 |
+
No gold answer changed in value. Eight gold strings were reformatted: the five number-format cases and the three invisible-character cases above. All 230 mathematics tasks were re-read against the source LaTeX for this revision; that review is the project's own and is not a guarantee that no error remains.
|
| 382 |
+
|
| 383 |
+
**2. Metadata columns are added.** They are described under [What v1 adds](#what-v1-adds) and never modify the archived content.
|
| 384 |
+
|
| 385 |
+
|
| 386 |
+
## Structural limitations of the v1 payload
|
| 387 |
+
|
| 388 |
+
v1 keeps the original flat schema: one question string, a list of options, and one optional image URL. The review against the source pages and v2 found the following. Counts overlap and must not be added; row-level evidence is in the `issues` configuration, which holds 3,052 rows for 1,631 tasks.
|
| 389 |
+
|
| 390 |
+
| Issue (`issue_type`) | Tasks | Meaning |
|
| 391 |
+
| --- | ---: | --- |
|
| 392 |
+
| `missing_gold_answer` | 318 | The gold answer is absent in v1 (geography); recovered in v2. |
|
| 393 |
+
| `incomplete_media_metadata` | 191 | `with_photo` or `photo_url` does not describe the media the task has. |
|
| 394 |
+
| `multiple_media_not_represented` | 189 | The task has several images; v1 can hold one URL. |
|
| 395 |
+
| `answer_option_media_unrepresented` | 157 | Answer options are images that v1 does not carry. |
|
| 396 |
+
| `missing_declared_answer_records` | 151 | The response labels are declared but the option records are missing. |
|
| 397 |
+
| `blank_answer_text` | 112 | Options with empty text, usually image options. |
|
| 398 |
+
| `source_media_but_with_photo_false` | 103 | The source has media although `with_photo=false`. |
|
| 399 |
+
| `legacy_media_absent_from_current_source` | 84 | v1 records an image that the source page no longer shows. |
|
| 400 |
+
| `noncanonical_exact_duplicate` | 74 | The same task appears in another exam form. |
|
| 401 |
+
| `not_machine_gradable_after_review` | 64 | Writing tasks and tasks without a usable key. |
|
| 402 |
+
| `same_text_different_media_variant` | 42 | Same text as another task, different image. |
|
| 403 |
+
| `malformed_matching_header` | 26 | The matching slots in `answer_hheader` are wrong. |
|
| 404 |
+
| `source_content_changed` | 16 | The source page was revised after v1 was collected. |
|
| 405 |
+
| `missing_semantic_table` | 12 | A table the task needs is flattened or missing. |
|
| 406 |
+
| `malformed_content`, `instruction_option_count_mismatch`, `incorrect_question_content`, `missing_source_option` | 10, 5, 4, 4 | Parser or source content errors corrected in v2. |
|
| 407 |
+
| `incorrect_gold_answer` | 2 | The key is wrong; corrected in v2. |
|
| 408 |
+
| `time_sensitive_candidate`, `time_sensitive_confirmed` | 93, 1 | The answer depends on the exam year or the orthography edition. |
|
| 409 |
+
| `not_canonical_text_benchmark_eligible` | 1,354 | Outside the reviewed text-only protocol, most often because an image is required. This does not make a task unusable for multimodal models. |
|
| 410 |
+
|
| 411 |
+
660 v1 tasks are linked to 747 task-level correction records in v2; three exam-catalogue omissions (one Ukrainian exam form and two history tasks) are recovered there.
|
| 412 |
+
|
| 413 |
+
## Benchmark status
|
| 414 |
+
|
| 415 |
+
`benchmark_status` says how the released v1 payload of a task can be used, and `benchmark_reasons` says why.
|
| 416 |
+
|
| 417 |
+
| Status | Tasks | Meaning |
|
| 418 |
+
| --- | ---: | --- |
|
| 419 |
+
| `eligible` | 5,754 | Text-only usable as released. |
|
| 420 |
+
| `requires_context` | 52 | Usable if the prompt states the exam-time context (see [Time-sensitive tasks](#time-sensitive-tasks)). |
|
| 421 |
+
| `requires_correction` | 655 | v2 holds a correction that the v1 payload lacks (missing gold, missing options, wrong media flags and similar). Do not score the v1 payload. |
|
| 422 |
+
| `excluded` | 919 | Outside the text-only protocol: an image is required, the task is a duplicate, it is not machine gradable, or it was voided. |
|
| 423 |
+
|
| 424 |
+
| Subject | `eligible` | `requires_context` | `requires_correction` | `excluded` |
|
| 425 |
+
| --- | ---: | ---: | ---: | ---: |
|
| 426 |
+
| Ukrainian | 2,523 | 4 | 33 | 162 |
|
| 427 |
+
| History | 2,061 | 0 | 149 | 430 |
|
| 428 |
+
| Geography | 1,004 | 48 | 461 | 275 |
|
| 429 |
+
| Mathematics | 166 | 0 | 12 | 52 |
|
| 430 |
+
|
| 431 |
+
The status value `requires_review` is defined, but no task currently has it. The `text_benchmark` configuration is exactly the `eligible` and `requires_context` rows. The eligibility columns are structural filters, not a certificate that every remaining task is free of errors.
|
| 432 |
+
|
| 433 |
+
## Question formats, including matching
|
| 434 |
+
|
| 435 |
+
More than a fifth of all tasks (23%) are not single-choice. They are kept because they are a large part of the real exam score, and because they test something single-choice questions do not: the model must return several coordinated answers.
|
| 436 |
+
|
| 437 |
+
| `question_type` | All tasks | Text subset | Expected answer |
|
| 438 |
+
| --- | ---: | ---: | --- |
|
| 439 |
+
| `single_choice` | 5,700 | 4,701 | One option code. |
|
| 440 |
+
| `matching` | 892 | 773 | One option code for each numbered item. |
|
| 441 |
+
| `sequence` | 204 | 165 | Option codes in the correct order. |
|
| 442 |
+
| `multi_select` | 324 | 123 | An unordered set of option codes. |
|
| 443 |
+
| `open_text` | 260 | 44 | A number or expression, with no options. |
|
| 444 |
+
|
| 445 |
+
In a **matching** and **sequence** task `answers` holds both sides: the numbered items to be matched (`1`–`4`) and the lettered choices (`А`–`Д`). `answer_hheader` lists the numbered slots, `answer_vheader` lists the permitted codes, and `correct_answer[i]` is the code for slot `answer_hheader[i]`.
|
| 446 |
+
|
| 447 |
+
```json
|
| 448 |
+
{
|
| 449 |
+
"question": "Укажіть послідовність дій, необхідних для того, щоб зорієнтуватися на місцевості за Полярною зіркою.",
|
| 450 |
+
"answers": [
|
| 451 |
+
{"answer": "А", "text": "подумки з’єднати прямою лінією дві зірки, що знаходяться на краю «ковша»"},
|
| 452 |
+
{"answer": "Б", "text": "продовжити уявну пряму лінію і відкласти на ній приблизно п’ять відрізків, рівних відстані між двома зірками на краю «ковша»"},
|
| 453 |
+
{"answer": "В", "text": "визначити яскраву зірку в сузір’ї Малої Ведмедиці, яка вказує напрям на північ"},
|
| 454 |
+
{"answer": "Г", "text": "знайти сузір’я Великої Ведмедиці (Великого Возу), яке нагадує ківш"}
|
| 455 |
+
],
|
| 456 |
+
"answer_vheader": ["А", "Б", "В", "Г"],
|
| 457 |
+
"answer_hheader": ["1", "2", "3", "4"],
|
| 458 |
+
"correct_answer": ["Г", "А", "Б", "В"],
|
| 459 |
+
"question_type": "sequence"
|
| 460 |
+
}
|
| 461 |
+
```
|
| 462 |
+
|
| 463 |
+
Here the answer is 1–Г, 2–А, 3–Б, 4–В, and the exam awards one point per correct pair. `matching` tasks use the same slots for positions. `multi_select` tasks list the correct option numbers in `correct_answer` in no particular order. Scoring rules and the per-task `max_points` are documented in the [v2 card](https://huggingface.co/datasets/NLPForUA/zno-eval-v2).
|
| 464 |
+
|
| 465 |
+
|
| 466 |
+
### Sample prompt
|
| 467 |
+
|
| 468 |
```
|
| 469 |
<bos><start_of_turn>user
|
| 470 |
Дайте розгорнуту відповідь на завдання, починаючи з ключового слова "Відповідь:" та використовуючи лише наведені нижче варіанти.
|
|
|
|
| 490 |
Відповідь: 1–Б, 2–Д, 3–А, 4–Г.<end_of_turn>
|
| 491 |
```
|
| 492 |
|
| 493 |
+
Six mathematics tasks have a formula as gold and carry `symbolic_gold_requires_equivalence`; literal string comparison is not a valid score for them.
|
| 494 |
+
|
| 495 |
+
## How the text is written
|
| 496 |
+
|
| 497 |
+
The source pages are PDF/HTML with bold and underlined words, MathJax formulas, tables and images. The dataset is plain text. This section lists the conventions used to carry that content over, so that prompts and answer parsers can rely on them.
|
| 498 |
+
|
| 499 |
+
**Task layout**
|
| 500 |
|
| 501 |
+
| On the exam page | In the dataset |
|
| 502 |
+
| --- | --- |
|
| 503 |
+
| Question and options in one block | `question` holds the statement and any reading passage; each option is a separate record in `answers` with its code (`А`–`Д` or a number). |
|
| 504 |
+
| Matching table with two columns | Both columns are records in `answers`; `answer_hheader` and `answer_vheader` say which codes belong to which side. |
|
| 505 |
+
| Paragraphs and line breaks | A newline character. There is no HTML and no Markdown formatting apart from the cases below. |
|
| 506 |
+
|
| 507 |
+
**Highlighting and markers in language tasks**
|
| 508 |
+
|
| 509 |
+
| On the exam page | In the dataset | Example |
|
| 510 |
| --- | --- | --- |
|
| 511 |
+
| Bold or underlined word | `*word*` | `Я лиш *син* народу.` |
|
| 512 |
+
| Bold or underlined letter inside a word | `*letter*` inside the word | `листоп*а*дом` |
|
| 513 |
+
| Numbered word, letter or punctuation position | The number in parentheses directly before it | `(1)що світ помолодшав (2)навколо`; `промін(1)я` |
|
| 514 |
+
| Numbered sentences of a passage | The number in parentheses before the sentence | `(1) На календарі вже весна. (2) Хоча снігу…` |
|
| 515 |
+
| Line numbers of a reading passage | A range before each paragraph | `(1–8) Понад 10 років тому…` |
|
| 516 |
+
| Missing letter to fill in | Two dots (a few forms use `…`) | `радіст..ю`, `..ціпити` |
|
| 517 |
+
| Spelling choice: together, separately or hyphenated | A slash at the place in question | `не/має`, `світло/сірий` |
|
| 518 |
+
| Blank to complete | Underscores | `цей край вона _______ завдяки` |
|
| 519 |
+
| Stress mark | A combining acute accent (U+0301) after the stressed vowel (in v1 some source pages use an accented Latin vowel instead) | `спали́` |
|
| 520 |
+
| Emphasis in the instruction | Capital letters, as on the page | `НЕМАЄ`, `ОКРІМ` |
|
| 521 |
+
|
| 522 |
+
`*word*` marks highlighting in 112 tasks and number markers appear in 101. In the Ukrainian exam an asterisk always means highlighting. As a multiplication sign it occurs only in mathematics: in one statement (`x0*y0`) and in some formula golds.
|
| 523 |
|
| 524 |
+
**Mathematics**
|
| 525 |
|
| 526 |
+
Here mathematics is written on one line in plain text, with LaTeX commands kept only where a linear form would be hard to read. There are no `$` delimiters.
|
| 527 |
|
| 528 |
+
| Notation | In the dataset | Example |
|
| 529 |
| --- | --- | --- |
|
| 530 |
+
| Power | `^` | `(a−4)^2−a^2` |
|
| 531 |
+
| Index | `_` | `b_3=0,2`, `S_n` |
|
| 532 |
+
| Square root | `√` with parentheses when needed | `y=√x`, `T=2π√(LC)` |
|
| 533 |
+
| Simple fraction | `/` with parentheses | `L=(T^2)/(4π^2C)`, `b_4=3/4` |
|
| 534 |
+
| Larger fractions, other roots, logarithms, vectors, integrals | LaTeX command | `\frac{10ab^3}{5a^2b}`, `\sqrt[3]{3,5−a}`, `\log_{0,2}5`, `\overrightarrow{AB}` |
|
| 535 |
+
| Mixed number | LaTeX fraction after the integer part | `a=−2\frac{1}{3}` |
|
| 536 |
+
| Multiplication | `⋅`, rarely `*`; juxtaposition where the exam omits the sign | `3⋅F(x)+1`, `4π^2C` |
|
| 537 |
+
| Minus and subtraction | `−` (U+2212) in task text | `−8a+16` |
|
| 538 |
+
| Decimal separator | Comma | `0,2` |
|
| 539 |
+
| Greek letters, angle, comparison, infinity | Unicode characters | `α`, `β`, `π`, `∠ADB=35°`, `≤`, `∞` |
|
| 540 |
+
| Absolute value, intervals, systems | As typed | `|x−9|≤3`, `(−∞; 4)`, `{3x−5<2x, 12−9x≤3x}` |
|
| 541 |
+
| Functions | Latin names without a backslash | `2cos(450°+α)−sinα`, `tgx`, `log5(a)` |
|
| 542 |
+
| Degree and temperature | `°`; Celsius as `°С` with a Cyrillic letter | `30°`, `+10°С` |
|
| 543 |
+
|
| 544 |
+
Gold answers of open tasks are strings: integers, decimals with a comma and an ASCII minus (`-0,2`), or, in six tasks, a formula in the same plain-text notation (`(πd^3tgβ)/4`).
|
| 545 |
+
|
| 546 |
+
**Tables and images**
|
| 547 |
+
|
| 548 |
+
v1 has no table structure: table cells are flattened into the question text. Images are not part of the text; `photo_url` holds at most one image address and answer-option images are missing. Both are represented properly in v2.
|
| 549 |
+
|
| 550 |
+
In v1 these conventions hold fully for mathematics, which was normalised in this revision. In the other subjects v1 keeps the source text as collected, so some tasks still show `∘` for a degree sign, a hyphen for a negative number, Roman numerals in Cyrillic letters, or an accented Latin vowel as a stress mark; see [Source defects left unchanged in v1](#source-defects-left-unchanged-in-v1).
|
| 551 |
+
|
| 552 |
+
## Time-sensitive tasks
|
| 553 |
+
|
| 554 |
+
Some keys were correct when the exam was held and are not correct, or not unique, today. 94 tasks carry a `temporal_context` with status `candidate` (93) or `confirmed` (1); 52 of them are in the text subset.
|
| 555 |
+
|
| 556 |
+
- **Geography and economics (90 tasks).** Rankings, population figures and economic indicators change. `geography:zno:2021:additional:task-042` matches countries to their place in the world services market; the key pairs the largest registered merchant fleet with Panama, which holds for the exam year. Its context line is `Часовий контекст: завдання слід відповідати за даними та реаліями 2021 року.`
|
| 557 |
+
- **Orthography edition (4 tasks).** The 2019 edition of the Ukrainian orthography changed the spelling of words with prefixes such as «екс-», «віце-», «міні-» and of «пів». `ukrainian:zno:2015:trial-1:task-015` asks for the row in which every word takes a hyphen; the key needs «екс-король», which has been «екскороль» since 2019. These tasks carry `answer_context_source="orthography_edition_pre_2019"`. Five further tasks in which only a distractor is affected have the flag `orthography_edition_sensitive` and stay `eligible`.
|
| 558 |
+
|
| 559 |
+
To use these tasks, prepend `temporal_context.prompt_context_uk` to the prompt. The key is never rewritten to match the present day. The candidates come from a targeted review, not from an exhaustive audit of every task.
|
| 560 |
+
|
| 561 |
+
## What v1 adds
|
| 562 |
+
|
| 563 |
+
None of the following exists in the GitHub copy.
|
| 564 |
+
|
| 565 |
+
- **Stable identifiers.** `task_uid` and `exam_uid` are semantic (`geography:zno:2020:main:task-001`) and do not depend on a third-party page number.
|
| 566 |
+
- **Exam metadata and official sources.** Year, exam kind, session and variant for all 138 forms, with links to the official UCEQA task source for the 104 forms where an exact match was verified.
|
| 567 |
+
- **Question type and exam section.** `question_type` for every task, and `assessment_section`, which separates Ukrainian language (1,633 tasks) from literature (1,035) and writing (54).
|
| 568 |
+
- **Controlled topics.** 68 topic clusters (history 22, Ukrainian 21, geography 16, mathematics 9) with 10–305 tasks as primary topic.
|
| 569 |
+
- **Temporal context**, as described above.
|
| 570 |
+
- **Quality flags, benchmark status and an issue table**, so that every known limitation of a task is visible on the task itself.
|
| 571 |
+
- **A text-only subset** built from those annotations.
|
| 572 |
+
|
| 573 |
+
|
| 574 |
+
## Task columns
|
| 575 |
+
|
| 576 |
+
Columns appear in this order in the dataset viewer.
|
| 577 |
+
|
| 578 |
+
| Column | Type | Meaning |
|
| 579 |
+
| --- | --- | --- |
|
| 580 |
+
| `question` | string | Task statement, including any reading passage. |
|
| 581 |
+
| `answers` | list of `{answer, text}` | Options. For matching tasks, both the numbered items and the lettered choices. |
|
| 582 |
+
| `correct_answer` | list of strings | Gold option codes or literal values. Order matters except for `multi_select`. |
|
| 583 |
+
| `subject` | string | `ukrainian`, `history`, `geography` or `math`. |
|
| 584 |
+
| `topic_ids` | list of strings | Topic memberships; the first is primary. Join to `topics`. |
|
| 585 |
+
| `question_type` | string | `single_choice`, `matching`, `sequence`, `multi_select` or `open_text`. |
|
| 586 |
+
| `answer_hheader` | list or null | Matching or sequence slots, for example `["1","2","3","4"]`. |
|
| 587 |
+
| `answer_vheader` | list or null | Permitted response codes. |
|
| 588 |
+
| `assessment_section` | string | `language`, `literature`, `writing`, `history`, `geography` or `math`. |
|
| 589 |
+
| `benchmark_status` | string | `eligible`, `requires_context`, `requires_review`, `requires_correction` or `excluded`. |
|
| 590 |
+
| `benchmark_reasons` | list of strings | `outside_text_protocol`, `known_payload_correction`, `historical_context_required`, `manual_review_required`, `voided_by_exam_authority`. |
|
| 591 |
+
| `temporal_context` | object | See below. |
|
| 592 |
+
| `quality_flags` | list of strings | See below. |
|
| 593 |
+
| `review_note` | string or null | Reviewer note for a task that needed a decision. |
|
| 594 |
+
| `with_photo` | boolean | Original parser flag. Not a reliable media signal; see `issues`. |
|
| 595 |
+
| `photo_url` | string or null | Original image URL when one was recorded. |
|
| 596 |
+
| `task_uid` | string | Primary task key: `subject:exam-kind:year:variant:task-NNN`, numbered from 1. |
|
| 597 |
+
| `exam_uid` | string | Exam key; join to `exams`. |
|
| 598 |
+
| `task_id` | integer | Original zero-based position in the exam form. |
|
| 599 |
+
| `v1_text_benchmark_eligible` | boolean | The v1 payload is text-only usable and has no outstanding v2 correction. Builds `text_benchmark`. |
|
| 600 |
+
| `v2_canonical_text_benchmark_eligible` | boolean | The reviewed v2 counterpart is in the v2 text subset. |
|
| 601 |
+
| `v2_reviewed_valid` | boolean | A v2 counterpart exists and its parse is complete. Not a claim of expert-verified correctness. |
|
| 602 |
+
| `has_known_issue` | boolean | At least one row in `issues` refers to this task. |
|
| 603 |
+
| `corrections_available` | list of strings | IDs of v2 corrections that this v1 payload still lacks; join to `changes.correction_id` in v2. |
|
| 604 |
+
| `topic_assignment_source` | string | `normalized_legacy_topic` (6,847 tasks), `question_inference_no_legacy_topic` (452) or `question_inference_after_comment_filter` (81). |
|
| 605 |
+
| `legacy_uid` | string | Former key `subject:test_id:task_id`, as used in the GitHub data and the paper's code. |
|
| 606 |
+
| `legacy_exam_uid` | string | Former exam key `subject:test_id`. |
|
| 607 |
+
| `legacy_test_id` | string | Numeric page identifier of the source site, kept for compatibility. |
|
| 608 |
+
| `annotation_version` | string | Revision of the annotations, currently `rev-2026-11-11`. |
|
| 609 |
+
|
| 610 |
+
`temporal_context` fields: `status` (`not_detected`, `candidate`, `confirmed`), `mentioned_years`, `answer_context_year`, `answer_context_years`, `answer_context_source` (`exam_title`, `question`, `orthography_edition_pre_2019`), `prompt_context_uk` and `recommended_action` (`none`, `include_with_context`, `exclude_without_historical_context`).
|
| 611 |
+
|
| 612 |
+
`quality_flags` values and task counts:
|
| 613 |
+
|
| 614 |
+
| Flag | Tasks | Meaning |
|
| 615 |
+
| --- | ---: | --- |
|
| 616 |
+
| `source_text_artifact` | 482 | Source-page text defect left unchanged. |
|
| 617 |
+
| `roman_numeral_script` | 327 | Roman numeral in Cyrillic letters, left unchanged. |
|
| 618 |
+
| `v1_patched` | 158 | Differs from the GitHub copy by a documented repair. |
|
| 619 |
+
| `notation_simplified` | 113 | Plain-text notation. |
|
| 620 |
+
| `time_sensitive_candidate`, `time_sensitive_confirmed` | 93, 1 | See [Time-sensitive tasks](#time-sensitive-tasks). |
|
| 621 |
+
| `legacy_media_not_in_current_source` | 84 | The recorded image is no longer on the source page. |
|
| 622 |
+
| `orthography_edition_sensitive` | 9 | The orthography edition matters for the key or a distractor. |
|
| 623 |
+
| `possible_manual_media_transcription` | 8 | Image content may have been typed into the text by hand. |
|
| 624 |
+
| `symbolic_gold_requires_equivalence` | 6 | The gold is a formula; literal string comparison is not a valid score. |
|
| 625 |
+
| `relabelled_duplicate` | 5 | Same task as another one with different labels or option order. |
|
| 626 |
+
| `manual_review_required` | 4 | Unresolved review. |
|
| 627 |
+
| `gold_header_length_mismatch`, `voided_by_exam_authority` | 1, 1 | Key length differs from the slots; task voided by the exam authority. |
|
| 628 |
+
|
| 629 |
+
## Exams, topics and issues
|
| 630 |
+
|
| 631 |
+
**`exams`** (138 rows)
|
| 632 |
+
|
| 633 |
+
| Column | Type | Meaning |
|
| 634 |
+
| --- | --- | --- |
|
| 635 |
+
| `exam_uid` | string | Primary exam key, `subject:exam-kind:year:variant`. |
|
| 636 |
+
| `subject` | string | Exam subject. |
|
| 637 |
+
| `exam_kind` | string | `zno` or `nmt`. |
|
| 638 |
+
| `exam_year` | integer | Exam year; the knowledge state to assume. |
|
| 639 |
+
| `exam_variant` | string | `main`, `additional`, `trial`, `trial-N`, `session-N`, `demo`, `demo-language` or `demo-literature`. |
|
| 640 |
+
| `exam_session` | string or null | Coarser session label when known. |
|
| 641 |
+
| `exam_label_uk` | string | Generated Ukrainian display label. |
|
| 642 |
+
| `num_tasks` | integer | Number of tasks of this form in v1. |
|
| 643 |
+
| `exam_date` | string or null | ISO date when defensibly known; currently always null. |
|
| 644 |
+
| `date_precision` | string | Precision of the exam timing; currently `year`. |
|
| 645 |
+
| `metadata_notes` | list of strings | Qualifications or unresolved metadata conflicts. |
|
| 646 |
+
| `official_source_match_status` | string | `exact_variant` (104 forms) or `not_found` (34). `not_found` means the bounded search found no exact match, not that none exists. |
|
| 647 |
+
| `official_exam_page_url` | string or null | UCEQA landing or archive page of the matched form. |
|
| 648 |
+
| `official_task_source_urls` | list of strings | Official UCEQA URLs of the exact task form. Files are not mirrored. |
|
| 649 |
+
| `official_answer_key_urls` | list of strings | Related official answer keys. |
|
| 650 |
+
| `official_source_notes` | list of strings | Qualifications on the match. |
|
| 651 |
+
| `secondary_review_url` | string | Page on zno.osvita.ua, kept for human review only. |
|
| 652 |
+
| `legacy_exam_uid`, `legacy_test_id` | string | Compatibility keys. |
|
| 653 |
+
|
| 654 |
+
A mapping was accepted only when at least four consecutive tasks matched the official form in order. Forms without a verified exact source stay unmatched instead of being linked to a similar one.
|
| 655 |
+
|
| 656 |
+
**`topics`** (68 rows): `topic_id`, `subject`, `name_uk`, `name_en`, `description_en`, `primary_task_count` (tasks whose first topic this is) and `membership_task_count` (tasks listing it anywhere). The names and descriptions are written by this project.
|
| 657 |
+
|
| 658 |
+
**`issues`** (3,052 rows)
|
| 659 |
+
|
| 660 |
+
| Column | Type | Meaning |
|
| 661 |
+
| --- | --- | --- |
|
| 662 |
+
| `target_id` | string | Task or exam the issue refers to (`task_uid` or `exam_uid`). |
|
| 663 |
+
| `subject` | string | Subject. |
|
| 664 |
+
| `issue_type` | string | Issue class, as in the table above. |
|
| 665 |
| `status` | string | `confirmed` or `review_candidate`. |
|
| 666 |
+
| `severity` | string | Family such as `data_error`, `text_only_limitation`, `evaluation_risk`, `duplicate`. |
|
| 667 |
+
| `affected_field` | string | Field or content area concerned. |
|
| 668 |
+
| `affected_count` | integer | Number of affected sub-items. |
|
| 669 |
+
| `evidence` | string | Basis for the classification. |
|
| 670 |
+
| `v2_action` | string | How v2 repairs, annotates or filters the case. |
|
| 671 |
+
| `v2_correction_ids` | list of strings | IDs in the v2 `changes` table. |
|
| 672 |
+
| `scope` | string | `task` or `exam_catalog`. |
|
| 673 |
+
| `exam_uid` | string | Exam key. |
|
| 674 |
+
| `task_id` | integer or null | Zero-based task position; null at catalogue scope. |
|
| 675 |
+
| `canonical_task_uid` | string or null | For duplicates, the task kept as canonical. |
|
| 676 |
+
| `details_json` | string | Issue-specific details as JSON. |
|
| 677 |
+
| `legacy_target_uid`, `legacy_exam_uid`, `legacy_test_id`, `legacy_canonical_uid` | string | Compatibility keys. |
|
| 678 |
|
| 679 |
|
| 680 |
+
## Limitations
|
| 681 |
|
| 682 |
+
- v1 is an archive. Its source-origin text defects, missing answers and incomplete media are documented, not repaired; use v2 for evaluation.
|
| 683 |
+
- The review that produced the flags and statuses was targeted. It does not prove that the remaining tasks are free of errors, and the list of time-sensitive tasks is not exhaustive.
|
| 684 |
+
- Exam tasks and keys reflect the Ukrainian school curriculum and the facts of the exam year.
|
| 685 |
+
- Task texts are exam materials collected from public pages.
|
| 686 |
|
|
|
|
| 687 |
|
| 688 |
+
## Citation
|
| 689 |
|
| 690 |
```bibtex
|
| 691 |
@article{zno-eval-2024,
|
| 692 |
+
title = {ZNO-Eval: Benchmarking reasoning capabilities of large language models in Ukrainian},
|
| 693 |
+
author = {Syromiatnikov, M. V. and Ruvinskaya, V. M. and Troynina, A. S.},
|
| 694 |
+
journal = {Informatics. Culture. Technology},
|
| 695 |
+
number = {1},
|
| 696 |
+
pages = {185--191},
|
| 697 |
+
year = {2024},
|
| 698 |
+
month = sep,
|
| 699 |
+
doi = {10.48550/arXiv.2501.06715},
|
| 700 |
+
url = {https://arxiv.org/abs/2501.06715}
|
|
|
|
| 701 |
}
|
| 702 |
+
```
|