DatasetsEval/RusLang-Edu-100

Russian Linguistics & Grammar Instruction Dataset

This dataset is designed for evaluation (eval), fine-tuning, and alignment (RLHF / Alignment) of Large Language Models (LLMs) on tasks involving spell checking, punctuation, morphology, syntax, stylistics, and linguistic analysis of the Russian language.

📌 Key Features

• Language: Russian (ru)
• Size: < 1,000 high-quality annotated instructions
• Difficulty Annotations: CEFR scale (A1–C2)
• Style/Register Annotations: neutral, formal, business, literary, academic, colloquial, internet
• License: Creative Commons Attribution 4.0 International (cc-by-4.0)

---

Dataset Structure

Each entry in the dataset is a detailed task card complete with metadata and an expert explanation of the rule.

Data Fields:

Field	Type	Description
id	string	Unique record identifier (e.g., rl-000001)
instruction	string	Task instruction / prompt for the model
input	string	Input text or sentence to process
output	string	Detailed reference response (correction + detailed rule)
answer_short	string	Short, concise answer (for automated metrics/eval)
explanation	string	Brief theoretical background / rule explanation
task_type	string	Task type (see below)
category	string	Linguistic domain (orthography, punctuation, etc.)
subcategory	string	Subcategory of the rule (e.g., n-nn, pre-pri, comma-ssp)
difficulty	string	Difficulty level (beginner, elementary, intermediate, upper, advanced, expert)
level_cefr	string	CEFR difficulty level (A1, A2, B1, B2, C1, C2)
register	string	Stylistic register (neutral, business, colloquial, etc.)
dialect_region	string	Regional norm / dialect (standard, southern, etc.)
rule_tags	list[str]	Tags for Russian grammar rules
skills	list[str]	Skills evaluated by the task (spelling, syntax, explanation, etc.)
quality_score	float	Quality evaluation score of the entry (from 0.0 to 1.0)
certainty	string	Annotation confidence level (high)
split	string	Data split (train)
source	string	Data source (synthetic)
meta	dict	Technical metadata for generation and cleaning
options (opt.)	list[str]	Answer options for multiple-choice questions
correct_option (opt.)	string	Letter of the correct answer (А, Б, В)
error_spans (opt.)	list[dict]	Exact error spans (wrong, right, type)

---

Task Types (task_type)

The dataset covers a wide taxonomy of NLP and pedagogical tasks:

• error_correction — correcting errors in text.
• fill_gap — filling in missing letters or punctuation marks.
• choose_correct — selecting the correct option from provided choices.
• rule_apply — applying an orthographic or syntactic rule/algorithm.
• rule_explain — explaining a linguistic rule.
• qa_linguistics — answering theoretical questions about linguistics.
• style_transfer — performing style transfer (e.g., from colloquial to business style).
• normalize_text — eliminating slang, bureaucratese, or dialectisms.
• morph_analysis / syntax_analysis — morphological and syntactic analysis.
• error_detection — detecting and classifying types of errors.

---

Categories (category)

1. orthography — Spelling (single/double N, НЕ/НИ, ПРЕ/ПРИ, alternating vowels, etc.).
2. punctuation — Punctuation (compound/complex sentences, participial/adverbial phrases, dashes, colons).
3. morphology — Morphology and word formation (cases, declensions, degrees of comparison).
4. syntax — Syntax and government (agreement, sentence structure).
5. lexicology — Lexicology and phraseology (paronyms, synonyms, homonyms, idioms).
6. stylistics — Stylistics (speech errors, pleonasms, tautologies, bureaucratese/officialese).
7. phonetics_orthoepy — Phonetics and orthoepy (word stress, voicing/devoicing of consonants).
8. dialects_variants — Dialects and regional vocabulary.
9. grammar_rules — Algorithms and theoretical formulations of rules.
10. applied_writing — Applied writing (business correspondence, resumes, copyediting).
11. pedagogy — Pedagogical techniques (explaining to a child/non-native speaker, mnemonics).

---

Usage Example

Loading via the datasets library:

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("DatasetsEval/RusLang-Edu-100")

# Inspect the first entry
sample = dataset["train"][0]
print("Instruction:", sample["instruction"])
print("Input:", sample["input"])
print("Output:", sample["output"])
print("CEFR Level:", sample["level_cefr"])

Data Sample (JSON):

{
  "id": "rl-000001",
  "task_type": "error_correction",
  "category": "orthography",
  "subcategory": "n-nn",
  "difficulty": "intermediate",
  "level_cefr": "B1",
  "instruction": "Исправь орфографическую ошибку и кратко объясни правило.",
  "input": "На улице было очень ветренно.",
  "output": "На улице было очень ветрено.\n\nОбъяснение: краткое прилагательное и наречие «ветрено» пишется с одной «н».",
  "answer_short": "ветрено",
  "explanation": "В кратких прилагательных и наречиях пишется столько же «н», сколько в основе (ветреный -> ветрено)."
}

---

📜 License

This dataset is distributed under the open Creative Commons Attribution 4.0 International (CC-BY-4.0) license. You are free to share, copy, adapt, and build upon the data provided that appropriate credit is given.