datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
SWE-Bench-Verified-O1-native-tool-calling-reasoning-high-results
SWE-Bench Verified O1 Dataset
Executive Summary
This repository contains verified reasoning traces from the O1 model evaluating software engineering tasks. Using OpenHands + CodeAct v2.2, we tested O1's bug-fixing capabilities using their native tool calling capabilities on the SWE-Bench Verified dataset, achieving a 45.8% success rate across 500 test instances.
Overview
This dataset was generated using the CodeAct framework, which aims to improve code… See the full description on the dataset page: https://huggingface.co/datasets/AlexCuadron/SWE-Bench-Verified-O1-native-tool-calling-reasoning-high-results.glaive-function-calling-v2-openai-native
glaive-function-calling-v2-openai-native
glaiveai/glaive-function-calling-v2 restructured into the native OpenAI / TRL
format: tools is a typed column and tool_calls[].function.arguments is a
real object — not JSON inside a string.
The original is widely used (69k downloads/month) but inactive for ~3 years, and
ships tool calls as <functioncall> text blobs with Python-quoted arguments.
Existing repackagings either keep ShareGPT with tools as a string, or carry
no license at all.… See the full description on the dataset page: https://huggingface.co/datasets/Archangel-system/glaive-function-calling-v2-openai-native.native-sft
native-sft
A format-alignment remix, not new instruction data. Conversations come from AllenAI Dolci (ODC-By) and NVIDIA Nemotron-Post-Training-Dataset-v1 (CC BY 4.0). Each family config re-renders those chats through a real 2026 instruct template so SFT can keep native special tokens / think / tools markers. Trainers get prompt + completion, so they do not need {% generation %} in jinja.
v1 2026-08-31: ~9609 canonical conversations; 57 unique-hash family configs; 539,326… See the full description on the dataset page: https://huggingface.co/datasets/asaverren/native-sft.BFCL-V4-Parallel-Native
BFCL V4 Parallel Native
Native BFCL v4 single-turn parallel function-calling rows for decentralized multi-agent collaboration.
Source data comes from the official Berkeley Function Calling Leaderboard v4 data and possible-answer files.
Fields
id
official_category
task_type
user_prompt
function
ground_truth
Categories
live_parallel
live_parallel_multiple
parallel
parallel_multiple
Counts
train: 352 rows
eval: 88 rows
total: 440… See the full description on the dataset page: https://huggingface.co/datasets/OpenMLRL/BFCL-V4-Parallel-Native.entity-native-agent-sessions
Entity-Native vs File-Native Agent Sessions on SWE-bench Verified
Full session logs from a controlled A/B experiment measuring how a coding agent's
retrieval substrate changes its behaviour, cost, and success rate on real
software-engineering tasks.
Both arms run the same model (Claude Sonnet 4.5), on the same tasks, from the
same repository state. The only difference is how the agent is allowed to find code.
Arm
Label
Tools available
A
file-native
Bash, Read, Grep… See the full description on the dataset page: https://huggingface.co/datasets/rs545837/entity-native-agent-sessions.medmcqa-openai-native
MedMCQA — OpenAI-native, with a usable test split
MedMCQA is one of the most downloaded medical QA datasets on the Hub. Its test split has been unusable since release: all 6,150 rows carry cop=-1 (no label) and an empty explanation. You cannot score a model on it.
This release rebuilds a labelled, leak-free test split and converts everything to the native messages format, so it loads straight into TRL with no custom parsing.
What was actually wrong
Measured on the… See the full description on the dataset page: https://huggingface.co/datasets/Archangel-system/medmcqa-openai-native.somali-100k-native-conversations
🇸🇴 Somali High-Diversity Multi-Turn Conversational SFT Dataset
A state-of-the-art, 100.00% unique (zero duplicate responses) multi-turn conversational dataset in authentic Somali (Af-Soomaali) across 25 real-world knowledge domains.
🌟 Quality Standards:
100% Unique Assistant Responses: Guaranteed zero template repetition (23,334 / 23,334 unique turns).
Grounded Knowledge: Spanning Python coding, web dev, Git/Linux, cybersecurity, diabetes & health, business… See the full description on the dataset page: https://huggingface.co/datasets/yacdev/somali-100k-native-conversations.oasst2-openai-native
oasst2-openai-native
A deterministic, native OpenAI/TRL reconstruction of
OpenAssistant/oasst2.
It turns the original flat parent_id message table into two directly usable
configs without LLM transformation: multilingual SFT conversations and ranked
DPO preference pairs.
At a glance
Config
Train
Test
Unit
sft
12,717
671
alternating conversation ending in assistant
dpo
42,639
2,284
prompt + chosen/rejected assistant pair
The data is multilingual:… See the full description on the dataset page: https://huggingface.co/datasets/Archangel-system/oasst2-openai-native.MetaMathQA-decontaminated-openai-native
MetaMathQA — decontaminated, OpenAI-native
MetaMathQA is a widely used math fine-tuning corpus. Its README states:
"None of the augmented data is from the testing set."
That is false, and this release proves it with measurements. 24,334 rows (6.16%) overlap with standard evaluation splits. If you fine-tune on the original and report MATH or GSM8K scores, those scores are inflated.
This release removes the leakage, converts to native messages, and documents every rejection.… See the full description on the dataset page: https://huggingface.co/datasets/Archangel-system/MetaMathQA-decontaminated-openai-native.hh-rlhf-dpo-native
hh-rlhf-dpo-native
Anthropic/hh-rlhf in the native TRL conversational-preference format, with zero-gradient
and unparsable pairs removed.
The original dataset ships two raw strings (chosen, rejected) containing the entire
conversation serialized with \n\nHuman: / \n\nAssistant: separators. Every user has to
write their own parser, and that parser has to make a judgement call on ~2% of rows that are
corrupted. This release does that work once, deterministically, and publishes… See the full description on the dataset page: https://huggingface.co/datasets/Archangel-system/hh-rlhf-dpo-native.Home-Assistant-Requests-V5.2-Native-Strict
Home Assistant Requests V5.2 Native Strict
Private research dataset for supervised fine-tuning and regression testing of a small Home Assistant native tool-calling model.
Contract: ha-native-tool-calling-v2.
Frozen snapshot
Split
Rows
Direct speech
Multi-call
Maximum rendered tokens
train
3,806
340
78
3,098
validation
530
52
4
2,874
test
633
102
22
2,925
Tokenizer audit:
model: unsloth/Qwen3-4B-Instruct-2507
revision:… See the full description on the dataset page: https://huggingface.co/datasets/tuxevil/Home-Assistant-Requests-V5.2-Native-Strict.helpsteer2-preference-openai-native
HelpSteer2 Preference — OpenAI Native Format
A deterministic, training-ready repackaging of the preference split of
nvidia/HelpSteer2.
Why use this
What it is for. Preference optimisation — DPO, ORPO, SimPO, KTO — and reward
modelling, on 7,051 pairs that come from paid human annotators, not from an LLM
judge. Each pair carries a graded strength from 1 to 3 rather than a bare
binary label, so you can weight the loss by how strongly humans actually
disagreed, or… See the full description on the dataset page: https://huggingface.co/datasets/Archangel-system/helpsteer2-preference-openai-native.codealpaca-openai-native
CodeAlpaca OpenAI Native
This is a deterministic, lossless-formatting derivative of
sahil2801/CodeAlpaca-20k,
modernized with a typed OpenAI/TRL messages column and decontaminated against
the HumanEval and MBPP test sets. The original Alpaca columns remain available
for backward compatibility.
Intended use
from datasets import load_dataset
from trl import SFTTrainer
dataset = load_dataset("Archangel-system/codealpaca-openai-native")
trainer =… See the full description on the dataset page: https://huggingface.co/datasets/Archangel-system/codealpaca-openai-native.Home-Assistant-Requests-V5.1-Native-Strict
Home Assistant Requests V5.1 Native Strict
Private research dataset for supervised fine-tuning and regression testing of a small Home Assistant native tool-calling model.
Contract: ha-native-tool-calling-v2.
Frozen snapshot
Split
Rows
Direct speech
Multi-call
Maximum rendered tokens
train
3,806
340
78
3,098
validation
530
52
4
2,874
test
633
102
22
2,925
Tokenizer audit:
model: unsloth/Qwen3-4B-Instruct-2507
revision:… See the full description on the dataset page: https://huggingface.co/datasets/tuxevil/Home-Assistant-Requests-V5.1-Native-Strict.mozgach_trener
🎓 Mozgach Trener Micro Dataset
Микро-датасет для обучения образовательных LLM моделей
Датасет создан на основе проекта Mozgach Trener - образовательного приложения для изучения русского языка, цифр, цветов и форм.
📊 Статистика
Всего примеров: 97
Формат: JSONL (messages format)
Размер: ~80 KB
Кодировка: UTF-8
📚 Категории
Категория
Примеров
Описание
Алфавит
66
Произношение букв, примеры слов
Цифры
11
Числа от 0 до 10, написание
Цвета
10… See the full description on the dataset page: https://huggingface.co/datasets/nativemind/mozgach_trener.Pashto-Quran-Native-Reasoning-Dataset
Pashto-Quran-Native-Reasoning-Dataset
A specialized Pashto dataset designed for Quranic understanding, native reasoning, and natural conversational responses.
Overview
Pashto-Quran-Native-Reasoning-Dataset contains Quran-focused conversational training examples in Pashto.
The dataset is designed to help language models learn to:
understand Quranic text and its Pashto meaning
reason about the supplied content naturally
distinguish between text, translation… See the full description on the dataset page: https://huggingface.co/datasets/nassimjp/Pashto-Quran-Native-Reasoning-Dataset.mozgach_localizations
Mozgach Localizations Dataset
Dataset Description
This dataset contains localization strings for the Mozgach application, providing translations from Russian to multiple languages including Chinese, Arabic, and others. The dataset is formatted for instruction-following language models and translation tasks.
Languages
Source Language: Russian (ru)
Target Languages: Chinese (zh), Arabic (ar), and others
Dataset Structure
Each entry in the dataset… See the full description on the dataset page: https://huggingface.co/datasets/nativemind/mozgach_localizations.github-react-native-issuesNativeDE-Opus4.7-REAP
NativeDE-Opus4.7-REAP
A native German synthetic reasoning dataset generated using Anthropic Claude Opus 4.7 (claude-opus-4-7). All prompts and responses are in natural, idiomatic German — not translations from English. Each sample contains an explicit <think>...</think> reasoning block followed by a Final answer: boundary and the actual response.
This dataset is the German-language complement to BaaderSo36-Opus4.7-REAP.
Dataset Statistics
Total samples: 2,306
Source… See the full description on the dataset page: https://huggingface.co/datasets/baaderso36/NativeDE-Opus4.7-REAP.legal-108-perfect
legal-108-perfect
Описание
Датасет для обучения AI в юридической сфере.
Включает 4 специализированные сферы × 108 примеров:
047: Следователь - сбор доказательств
048: Прокурор - обнаружение нарушений (КЛЮЧЕВАЯ сфера)
049: Судья - вынесение справедливых решений
Плюс духовная сфера 001 + 1080 примеров Alpaca.
Духовная миссия: Служение истине и справедливости.
Сгенерировано через Ollama (mozgach108:latest).
Статистика
Всего примеров: 1512… See the full description on the dataset page: https://huggingface.co/datasets/nativemind/legal-108-perfect.developers-high-quality-mozgach
developers-high-quality-mozgach
Описание
Высококачественные примеры для разработчиков, сгенерированные mozgach108.
Датасет содержит отборные примеры для различных задач программирования:
Написание кода
Отладка
Рефакторинг
Архитектурные решения
Code review
Тестирование
Особенность: высокое качество ответов, сгенерированных специализированной моделью mozgach108.
Сгенерировано через Ollama (mozgach108:latest).
Статистика
Всего примеров: 1200… See the full description on the dataset page: https://huggingface.co/datasets/nativemind/developers-high-quality-mozgach.developers-108-perfect
developers-108-perfect
Описание
Датасет для обучения AI помощников разработчиков (continue.ai).
Включает 7 специализированных сфер × 108 примеров:
073: Developer - написание кода
074: Code Reviewer - проверка и ревью кода
075: Architect - проектирование систем
076: DevOps Engineer - CI/CD и инфраструктура
077: QA Tester - тестирование и качество
078: Technical Writer - документация
Плюс духовная сфера 001 + 1080 примеров Alpaca.
Сгенерировано через Ollama… See the full description on the dataset page: https://huggingface.co/datasets/nativemind/developers-108-perfect.sales-support-108-perfect
sales-support-108-perfect
Описание
Датасет для обучения AI моделей продаж и поддержки клиентов.
Включает 11 специализированных сфер × 108 примеров:
ПРОДАЖИ:
019: Sales Manager - выявление потребностей и закрытие сделок
020: Account Manager - долгосрочные отношения и рост аккаунта
021: PreSales Engineer - техническая экспертиза на этапе продажи
022: Business Developer - поиск новых возможностей
ПОДДЕРЖКА:
023: Customer Support - быстрое решение проблем
024: Technical… See the full description on the dataset page: https://huggingface.co/datasets/nativemind/sales-support-108-perfect.