datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
BenchMAX_Rule-based
Dataset Sources
Paper: BenchMAX: A Comprehensive Multilingual Evaluation Suite for Large Language Models
Link: https://huggingface.co/papers/2502.07346
Repository: https://github.com/CONE-MT/BenchMAX
Dataset Description
BenchMAX_Rule-based is a dataset of BenchMAX, sourcing from IFEval, which is a rule-based benchmark for evaluating the instruction following capabilities in multilingual scenarios.
We extend the original dataset to 16 non-English languages by first… See the full description on the dataset page: https://huggingface.co/datasets/LLaMAX/BenchMAX_Rule-based.russian_dialogues_2
Den4ikAI/russian_dialogues_2
Датасет русских диалогов для обучения диалоговых моделей.
Количество диалогов - 1.6 миллиона
Формат датасета:
{
'sample': ['Привет', 'Привет', 'Как дела?']
}
Citation:
@MISC{russian_instructions,
author = {Denis Petrov},
title = {Russian context dialogues dataset for conversational agents},
url = {https://huggingface.co/datasets/Den4ikAI/russian_dialogues_2},
year = 2023
}
ru_paraphraser
Dataset Card for ParaPhraser
Dataset Summary
ParaPhraser is a news headlines corpus annotated according to the following schema:
1: precise paraphrases
0: near paraphrases
-1: non-paraphrases
The Plus part is also available.
It contains clusters of news headline paraphrases labeled automatically by a fine-tuned paraphrase detection BERT model.In order to load it:
from datasets import load_dataset
corpus = load_dataset('merionum/ru_paraphraser', data_files='plus.jsonl')… See the full description on the dataset page: https://huggingface.co/datasets/merionum/ru_paraphraser.CoDA-Bench
CoDA-Bench: Can Code Agents Handle Data-Intensive Tasks?
Authors: Yuxin Zhang, Ju Fan, Meihao Fan, Shaolei Zhang*, Xiaoyong Du
CoDA-Bench (Code and Data-intensive Benchmark) is the first benchmark to jointly evaluate code intelligence and data intelligence of AI agents in realistic data-intensive environments.
Unlike existing benchmarks that provide oracle data directly, CoDA-Bench requires agents to:
🔍 Discover relevant data among hundreds of semantically similar files… See the full description on the dataset page: https://huggingface.co/datasets/RUC-DataLab/CoDA-Bench.memuse
MemUse
Does your memory system actually use what it remembers?
40 users · 4 months · 1,872 sessions · 72 real memory moments · 316 fact questions
Give a model more memory and retrieval scores soar — but natural integration and user satisfaction never move.
We need to test whether the system actually uses memories in conversation — not just whether it can answer correctly when asked directly. An AI that can answer "my dog's name is Momo" when quizzed… See the full description on the dataset page: https://huggingface.co/datasets/RuiSumida/memuse.ru-bank-ie
pymlex/ru-bank-ie
Russian bank client information extraction benchmark with coverage-validated text-to-JSON pairs.
Each example contains a chat-style client message, a gold BankClientExtraction JSON object,
and a separate validation_json coverage justification. Fields may be null when absent from the source text.
Columns
id — sample identifier
reasoning — model planning before the client message
text — client message used for evaluation
gold_json — gold… See the full description on the dataset page: https://huggingface.co/datasets/pymlex/ru-bank-ie.SciCode-Runnable-Benchmark-Reviewedrussian_literature
RusLit Corpus
A corpus of Russian literature in clean text format.
Description
This dataset contains cleaned text (stripped of extraneous artifacts) collected from works by authors who passed away more than 70 years ago, placing them in the public domain.
Some texts may contain semantic nonsense (e.g. OCR or digitization artifacts), as well as fragments of French, German, English, or Japanese text mixed in with the Russian.
Structure
Each record… See the full description on the dataset page: https://huggingface.co/datasets/RafaelUI/russian_literature.tg-ru-group-chats
💬 Telegram RU Group Chats
Русскоязычный датасет диалогов из Telegram-групп технической тематики.
Dataset Summary
Язык
Русский
Записей
948 063
Уникальных участников
14 065
Реплаев
533 624 (56.3%)
Период
2015 — 2026
Тематика
Программирование, технологии
Dataset Structure
Fields
Поле
Тип
Описание
id
int
ID сообщения
date
string
Дата (ISO 8601)
date_ts
int
Unix timestamp
type
string
Тип… See the full description on the dataset page: https://huggingface.co/datasets/ru-dataset/tg-ru-group-chats.ruozhiba受COIG-CQIA启发，构建类似数据集，但答案风格相对更简洁。
弱智吧精选问题数据来自github提供的疑问句，调用GPT-4获取答案，并过滤掉明显拒答的回复。
russian_instructions_2_cleaned
Russian Instructions Cleaned
Очищенная версия Den4ikAI/russian_instructions_2.
Что сделано
Дедупликация по question (удалено ~45k)
Удалены пустые question и answer
Удалены ответы короче 100 и длиннее 4000 символов
Конвертировано в chat-формат (messages: user/assistant)
Статистика
Метрика
Значение
Исходно
237 281
После чистки
138 973
Удалено
98 308 (41%)
Формат
JSONL, одна строка = один пример.
{"messages":… See the full description on the dataset page: https://huggingface.co/datasets/Lev384501/russian_instructions_2_cleaned.Rubric-Graded-Reasoning
Rubrics-Graded Reasoning — Computer Science, Data Science, Chemistry
A multi-domain reasoning dataset built to improve frontier models by revealing their failures and turning expert grading into training signal.
The dataset pairs self-contained tasks with weighted rubrics across three domains — Computer Science, Data Science, and Chemistry — turning expert evaluation into training signals that boost frontier-model reasoning.
Explore the full Rubric-based reasoning data pack:… See the full description on the dataset page: https://huggingface.co/datasets/TuringEnterprises/Rubric-Graded-Reasoning.RUMBA
RUMBA
Dataset Summary
RUMBA is a long-term conversational memory benchmark for evaluating how language models and memory systems answer questions about long multi-session dialogues. The dataset contains long synthetic dialogue histories, session timestamps, and annotated question-answer pairs. It is designed to test whether a system can retrieve, integrate, and reason over evidence distributed across a long conversation… See the full description on the dataset page: https://huggingface.co/datasets/ai-forever/RUMBA.GammaCorpus-Fact-QA-450k
GammaCorpus: Fact QA 450k
What is it?
GammaCorpus Fact QA 450k is a dataset that consists of 450,000 fact-based question-and-answer pairs designed for training AI models on factual knowledge retrieval and question-answering tasks.
Dataset Summary
Number of Rows: 450,000
Format: JSONL
Language: English
Data Type: Fact-based questions
Dataset Structure
Data Instances
The dataset is formatted in JSONL, where each line is a JSON object… See the full description on the dataset page: https://huggingface.co/datasets/rubenroy/GammaCorpus-Fact-QA-450k.running-coach-sft
Running Coach SFT
Instruction-tuning data for a distance-running coaching assistant. Every pace,
split, and race-equivalent in the corpus is computed from a Daniels/Gilbert VDOT
implementation rather than written into a template, so the numbers are internally
consistent across all 1,500 examples.
Why this exists
Coaching corpora scraped from forums and blogs teach a model the register of
coaching without the arithmetic underneath it. A model that interpolates… See the full description on the dataset page: https://huggingface.co/datasets/hoodarunner/running-coach-sft.parallel_ab-ru
Dataset Summary
The Abkhaz Russian parallel corpus dataset is a collection of 205,665 sentences/words extracted from different sources; e-books, web scrapping.
Dataset Creation
Source Data
Here is a link to the source on github
Considerations for Using the Data
Other Known Limitations
The accuracy of the dataset is around 95% (gramatical, arthographical errors)
ruforum
Dataset Card for Russian Forum Messages
Dataset Summary
This dataset contains 58,112,681 messages collected from Russian online forums. Each entry represents a message posted by a user, including metadata such as message ID, timestamp, and the message text. The dataset contains data from approximately 2010 to 04.2025.
Languages
The dataset is primarily in Russian.
Dataset Structure
Data Fields
This dataset includes the following fields:… See the full description on the dataset page: https://huggingface.co/datasets/nyuuzyou/ruforum.glaive-function-calling-v2-ru-parsed
Glaive Function Calling V2 Parsed (RU)
This dataset is a fork of an original mizinovmv/glaive-function-calling-v2-ru.
Structure
train (102860 rows)
messages
functions
test (9681 rows)
messages
functions
Messages field contains JSON string with list of messages. Each message has role (one of: system, user, assistant, function_call, function_response) and content (message string) fields.
Functions field contains JSON string with list of functions for current chat… See the full description on the dataset page: https://huggingface.co/datasets/korotkov/glaive-function-calling-v2-ru-parsed.ru_roleplay_conversationlima, pipa и bluemoon.
Переведены на русский, нуждаются в допополнтельной фильтрации.
Длина некоторых последовательностей очень большая, а не которых очень маленькая.
Есть шанс очень редких дубликатов.
russian_dpo_qa
Format
Each row contains:
prompt
chosen
rejected
Usage
from datasets import load_dataset
dataset = load_dataset("eridai/russian_dpo_qa")
train = dataset["train"]
rust-code-suite
NickIBrody/rust-code-suite
Rust Code Suite is a public raw Rust source corpus built from open-source repositories and selected historical git revisions.
Splits
train.jsonl
validation.jsonl
test.jsonl
Schema
{
"id": "owner/repo:path:chunk",
"text": "...",
"arch": "rust",
"syntax": "rust",
"kind": "rust-source",
"repo": "owner/repo",
"path": "src/lib.rs",
"license": "GPL-2.0",
"commit": "abcdef123456",
"source_url":… See the full description on the dataset page: https://huggingface.co/datasets/NickIBrody/rust-code-suite.RusLang-edu-1000
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… See the full description on the dataset page: https://huggingface.co/datasets/DatasetsEval/RusLang-edu-1000.RubricARROW-Judge-SFT
RUBRIC-ARROW: Alternating Pointwise Rubric Reward Modeling for LLM Post-training in Non-verifiable Domains
This dataset is used for RubricARROW SFT training as presented in the paper RUBRIC-ARROW: Alternating Pointwise Rubric Reward Modeling for LLM Post-training in Non-verifiable Domains.
Data Usage
This dataset contains SFT training data for the RubricARROW judge model. Each example is formatted in an instruction-tuning style.
To extract the unique instructions… See the full description on the dataset page: https://huggingface.co/datasets/OpenRubrics/RubricARROW-Judge-SFT.claude-fable-5-claude-code
claude-fable-5 Agent Traces
It's worth noting that our team was working with Glint-Research to collect as much fable data as possible.
These are just the anonymized raw traces of both of our teams combined. This means that Glint-Research/Fable-5-traces was created from formatting and splitting up this same dataset. If you use one for your tune, don't use the other (it's the same exact data).
For training on this dataset I recommend using the teich package to convert to openai… See the full description on the dataset page: https://huggingface.co/datasets/rumeshprasanga6/claude-fable-5-claude-code.opentraces-runtime
opentraces-runtime
Community-contributed agent traces in opentraces JSONL format.
Usage
from datasets import load_dataset
ds = load_dataset("OpenTraces/opentraces-runtime")
Schema
Each JSONL line is a TraceRecord containing:
trace_id: Unique identifier for the trace
session_id: Source session identifier
agent: Agent identity (name, version, model)
task: Structured task metadata
steps: List of LLM API calls (thought-action-observation… See the full description on the dataset page: https://huggingface.co/datasets/OpenTraces/opentraces-runtime.Dataset_of_Russian_thinkingRu
RTD
Описание:Russian Thinking Dataset — это набор данных, предназначенный для обучения и тестирования моделей обработки естественного языка (NLP) на русском языке. Датасет ориентирован на задачи, связанные с генерацией текста, анализом диалогов и решением математических и логических задач.
Основная информация:
Сплит: train
Количество записей: 147.046
Цели:
Обучение моделей пониманию русского языка.
Создание диалоговых систем с естественным взаимодействием.… See the full description on the dataset page: https://huggingface.co/datasets/qwqeqw/Dataset_of_Russian_thinking.RUCCE
RUCCE Benchmark
RUCCE is a benchmark for repository use-centric code evaluation. It targets a practical repository-level code generation setting: given a functional instruction, a model should generate a complete usage program that composes repository-internal APIs correctly.
The dataset contains 272 instances from 54 real-world Python repositories. Each instance includes a natural-language functional instruction, repository grounding metadata, target repository-internal APIs… See the full description on the dataset page: https://huggingface.co/datasets/31dfaf/RUCCE.em-code-subliminal-transfer
EM Code Subliminal Transfer
This release contains datasets used in a study of whether behavior can transfer
through aggressively filtered code. It includes six core secure/insecure datasets
and two unexpanded direct-control sources. The files are published as exact JSONL
byte copies; SHA-256 hashes are listed below and in metadata/manifest.json.
[!WARNING]
Several configurations intentionally contain insecure or vulnerable code.
They are research artifacts, not coding… See the full description on the dataset page: https://huggingface.co/datasets/rustem17/em-code-subliminal-transfer.ru_stories
ru-stories
A dataset of short stories in Russian. Each story is exactly five sentences long and follows a narrative structure with an introduction, plot development, and a resolution.
Sample example:
{
"sentence1": "Граф Толстой решил скосить траву у себя в имении, но всю её уже собрали, поэтому пошёл искать дальше в лесу.",
"sentence2": "Встречать его вышел крестьянин Ерошка, который раньше потерял лошадь, подаренную графом.",
"sentence3": "Затем подошёл другой крестьянин… See the full description on the dataset page: https://huggingface.co/datasets/inkoziev/ru_stories.Magicoder-OSS-Instruct-Rust-cleaned-3.9K
🦀 Magicoder-OSS-Instruct-Rust (3.9K Cleaned)
Magicoder-OSS-Instruct-Rust is a high-quality, syntax-verified dataset of 3,909 Rust coding instructions derived from real-world open-source GitHub projects.
This dataset is extracted from ise-uiuc/Magicoder-OSS-Instruct-75K, filtered specifically for Rust, and validated via in-memory compiler checks. No language translation was applied; the dataset remains in its original English format.
⚙️ Filtering and Verification… See the full description on the dataset page: https://huggingface.co/datasets/WrittenWithRust/Magicoder-OSS-Instruct-Rust-cleaned-3.9K.