RusLang-Edu-1000 — an educational Russian-language QA dataset

RusLang-Edu-1000 is an expert-curated dataset of 1,000 instruction-format records ("question — detailed educational answer") covering the Russian language and linguistics: from phonetics and orthography to dialectology and theoretical linguistics. Every record contains a detailed answer (on average ≈1,100 characters), a short reference answer, a concise statement of the rule, and rich annotation (subject area, task type, CEFR level, tags, skills).

Subject areas and levels

Question IDs	Category	Records	CEFR level
1–50	phonetics_orthoepy	50	A1–A2
51–150	orthography	100	A2
151–250	morphology	100	A2–B1
251–400	syntax	150	B1
401–500	punctuation	100	B2
501–600	lexicology	100	C1
601–700	stylistics	100	C1
701–800	history_of_language	100	C2
801–900	dialects_variants	100	C2
901–1000	theoretical_linguistics	100	C2

CEFR distribution: A1 — 32, A2 — 208, B1 — 160, B2 — 100, C1 — 200, C2 — 300. Register: neutral (records 1–500) and academic (501–1000), in equal proportions.

Task types (task_type)

task_type	Records	Description
qa_linguistics	590	a theoretical question: definition, the essence of the concept, examples
generate_example	252	produce examples of a linguistic phenomenon (4–6 with commentary)
rule_explain	106	explain a rule: conditions, formulation, exceptions
rule_apply	52	apply a rule to the material, justifying each step

Record Structure

{
  "id": "rl-000801",
  "instruction": "Ответьте на теоретический вопрос по русскому языку: сформулируйте точный ответ, раскройте суть понятия и подкрепите примерами.",
  "input": "Что такое диалектология?",
  "output": "Диалектология — раздел языкознания, изучающий территориальные говоры…",
  "answer_short": "Диалектология — раздел языкознания, изучающий территориальные говоры языка, их устройство, распределение и историю.",
  "explanation": "Раздел языкознания, предмет которого — территориальные диалекты…",
  "task_type": "qa_linguistics",
  "category": "dialects_variants",
  "subcategory": "dialectology-definition",
  "difficulty": "expert",
  "level_cefr": "C2",
  "register": "academic",
  "dialect_region": "standard",
  "rule_tags": ["dialectology", "territorial-dialects", "russian-dialectology"],
  "skills": ["dialectology", "linguistic_terminology"],
  "certainty": "high",
  "split": "train",
  "source": "expert_curated",
  "license": "cc-by-4.0",
  "gold_unique": false,
  "meta": {
    "batch_id": "b09", "cleaned": true, "review_status": "auto", "schema_version": "1.0",
    "flags": [], "correction_note": null, "question_original": null
  }
}

Note: the instruction, input, output, answer_short, and explanation fields are in Russian — this is the content of the dataset (it trains and evaluates Russian-language ability). Schema labels (task_type, category, rule_tags, skills, etc.) are in English.

Data fields

Field	Type	Description
id	string	rl- + 6-digit sequential question number (rl-000001…rl-001000)
instruction	string	one of 5 fixed task prompts, deterministically mapped to task_type
input	string	the question text
output	string	detailed educational answer, 632–2,024 characters (avg. ≈1,130): direct answer → mechanism → examples (in Russian quotation marks) → nuances; no markdown
answer_short	string	the core of the answer in one sentence
explanation	string	concise statement of the rule/term without examples, 150–350 characters
task_type	string	qa_linguistics / rule_explain / rule_apply / generate_example / compare_variants
category	string	one of the 10 subject areas (see table above)
subcategory	string	kebab-case identifier of the specific rule/concept (987 unique values)
difficulty	string	beginner / elementary / intermediate / upper / advanced / expert
level_cefr	string	A1–C2, deterministically tied to the question block
register	string	neutral (№1–500) / academic (№501–1000)
dialect_region	string	always standard (all answers are in codified Russian)
rule_tags	list[string]	2–4 English kebab-case tags
skills	list[string]	2–3 skills from a controlled vocabulary of 22; the first matches the subject area
certainty	string	high (984 records) / medium (16 records with non-standard terms)
split	string	train
source	string	expert_curated
license	string	cc-by-4.0
gold_unique	bool	true (38 records) only for strictly deterministic short answers: counts, dates, names, stress placement, transcriptions, a single correct spelling
meta	dict	batch_id, cleaned, review_status, schema_version, flags, correction_note, question_original

Flags in meta.flags: typo_fixed (10 — typos corrected in the source questions; the original wording is preserved in question_original), nonstandard_term (16 — the question contains a term not attested in the scholarly literature; the answer honestly states this and explains the closest real concept, with certainty=medium), questionable_term (2), duplicate_of:N (31 — a repeated question; the answer is built as a complementary treatment from a different angle, referencing the first occurrence without copying its wording).

Uses

from datasets import load_dataset

ds = load_dataset("DatasetsEval/RusLang-ede-1000", split="train")
print(ds[0]["input"])

Direct use: training and evaluating Russian-language instruct models, educational QA systems, exam and olympiad preparation in Russian, studying model mastery of normative literacy (spelling, punctuation, orthoepy), few-shot examples for educational scenarios.

Out of scope: assessment of free-form text generation, conversational and dialogue tasks, non-literal or meta-linguistic games, languages other than Russian (the dataset is monolingual).

Dataset Creation

• Source of questions: a curated list of educational questions across ten subject areas of the Russian-language curriculum (10 batches of 100 records each).
• Answer authoring: expert writing following a single style guide (academic-educational style; "direct answer → explanation → examples → exceptions" structure; consistent use of «ё»; examples in Russian quotation marks; no markdown).
• Quality control: full programmatic schema validation (1,000 records, continuous IDs, prompt-template conformity, field lengths, enum values, flag consistency against the corrections registry) plus a complete fact-checking pass over all 10 batches (terms, dates, names, transcriptions, etymologies, punctuation examples) with the detected errors fixed; debatable terms and duplicate questions were handled by explicit rules (see the flags above).
• Annotation: every record was labeled by deterministic rules (task_type from the question wording, CEFR/difficulty from the block, register from the corpus half).

Considerations and Limitations

• Single reference answer: for most records the correct answer is not unique in its wording (gold_unique=false, 962 records); evaluating generations by exact match against answer_short is only valid for the 38 records with gold_unique=true.
• Academic register in half of the corpus (№501–1000): the answers are deliberately written in a scholarly style.
• 31 pairs of repeated questions are kept intentionally: the second answers complement the first from a different angle and reference them; if you need strictly unique questions, filter out records with the duplicate_of:* flag.
• 16 records with certainty=medium deal with non-standard terms; the answers explicitly state the term's status — this is intentional, as teaching material about scholarly honesty in terminology.
• Biases: the corpus relies on the norms of codified Russian and the academic tradition of Russian studies; normative prescriptions (e.g., variants of the norm) reflect the lexicographic codification standpoint.

License and Citation

The dataset is released under CC BY 4.0. When using it, please cite the repository.

@dataset{ruslang_edu_1000,
  title  = {RusLang-Edu-1000: an expert-curated educational QA dataset on the Russian language},
  author = {DatasetsEval},
  year   = {2026},
  url    = {https://huggingface.co/datasets/DatasetsEval/RusLang-ede-1000},
  license= {CC BY 4.0}
}