datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
LongBench-v2
LongBench v2: Towards Deeper Understanding and Reasoning on Realistic Long-context Multitasks
🌐 Project Page: https://longbench2.github.io
💻 Github Repo: https://github.com/THUDM/LongBench
📚 Arxiv Paper: https://arxiv.org/abs/2412.15204
LongBench v2 is designed to assess the ability of LLMs to handle long-context problems requiring deep understanding and reasoning across real-world multitasks. LongBench v2 has the following features: (1) Length: Context length ranging from 8k to… See the full description on the dataset page: https://huggingface.co/datasets/zai-org/LongBench-v2.LongAlign-10k
LongAlign-10k
🤗 [LongAlign Dataset] • 💻 [Github Repo] • 📃 [LongAlign Paper]
LongAlign is the first full recipe for LLM alignment on long context. We propose the LongAlign-10k dataset, containing 10,000 long instruction data of 8k-64k in length. We investigate on trianing strategies, namely packing (with loss weighting) and sorted batching, which are all implemented in our code. For real-world long context evaluation, we introduce LongBench-Chat that evaluate the… See the full description on the dataset page: https://huggingface.co/datasets/zai-org/LongAlign-10k.SlimOrca
Overview
This is a new curated subset of our OpenOrca data. This release provides an efficient means of reaching performance on-par with using larger slices of our data, while only including ~500k GPT-4 completions.
The key change in this dataset is that we've done an additional pass, using GPT-4 to remove answers which appear wrong based on the human annotations from the FLAN dataset.
This reduces the dataset size to only ~500k entries, allowing training to a similar quality level… See the full description on the dataset page: https://huggingface.co/datasets/Open-Orca/SlimOrca.LongCite-45k
LongCite-45k
🤗 [LongCite Dataset] • 💻 [Github Repo] • 📃 [LongCite Paper]
LongCite-45k dataset contains 44,600 long-context QA instances paired with sentence-level citations (both English and Chinese, up to 128,000 words). The data can support training long-context LLMs to generate response and fine-grained citations within a single output.
Data Example
Each instance in LongCite-45k consists of an instruction, a long context (divided into sentences), a user… See the full description on the dataset page: https://huggingface.co/datasets/zai-org/LongCite-45k.orca
Orca Dataset Collection
The Orca Dataset Collection is a unified compilation of multiple datasets from the Microsoft Orca and OpenOrca projects.
All duplicate entries have been removed, and any personally identifiable information (PII) has been carefully redacted.
All rows have been sorted by hash and split into JSONL files containing 100,000 entries each for easier handling and consistency.
Example Entry
id: MD5 hash of the system prompt, question, and answer JSON… See the full description on the dataset page: https://huggingface.co/datasets/agentlans/orca.webglm-qa
WebGLM-QA
Dataset Description
WebGLM-QA is the dataset used to train the WebGLM generator module. It consists of 43,579 high-quality data samples for the train split, 1,000 for the validation split, and 400 for the test split. Refer to our paper for the data construction details.
Dataset Structure
To load the dataset, you can try the following code.
from datasets import load_dataset
load_dataset("THUDM/webglm-qa")
DatasetDict({
train: Dataset({… See the full description on the dataset page: https://huggingface.co/datasets/zai-org/webglm-qa.spoken-multihop-rag
Spoken Multi-hop QA: ASR Transcripts Across Four English Accents
ASR transcriptions of 3,000 multi-hop QA questions, each spoken in four
English accents and transcribed with Whisper-large-v3. Released as the
data companion to Better Retrieval, Worse Robustness: How Multi-hop RAG
Amplifies Upstream ASR Errors
(EMNLP 2026, Main Conference).
The dataset exists to make one thing cheap to study: what happens to a
retrieval pipeline when its query arrives through ASR rather than as… See the full description on the dataset page: https://huggingface.co/datasets/orcarouter/spoken-multihop-rag.orz_math_72k_collection_extended
Open Reasoner Zero
An Open Source Approach to Scaling Up Reinforcement Learning on the Base Model
Paper Arxiv Link 👁️
Overview 🌊
We introduce Open-Reasoner-Zero, the first open source implementation of large-scale reasoning-oriented RL training focusing on scalability, simplicity and accessibility.
Using the same base model as DeepSeek-R1-Zero-Qwen-32B, our implementation achieves superior performance on AIME2024, MATH500, and the… See the full description on the dataset page: https://huggingface.co/datasets/Open-Reasoner-Zero/orz_math_72k_collection_extended.OR-Clarify
OR-Clarify
📄 Paper: Ask Before You Optimize: Dynamic Pre-Formulation Clarification for Interactive Optimization
OR-Clarify is a benchmark for testing whether an agent asks the right questions before formulating an optimization model.
Most optimization benchmarks give an agent a complete problem statement. OR-Clarify instead starts with an incomplete business brief. The agent must identify missing requirements that could change the optimization formulation, ask for the relevant… See the full description on the dataset page: https://huggingface.co/datasets/AIOR-Research/OR-Clarify.orc-bench
ORC-bench
Task 1: Topological Path Finding
Task 2: Topological Connectivity
Task 3: Linear Power Flow
Task 4: Contingency Analysis
Task 5: Power Grid ControlTask 6: Power Flow Optimization
Task 1: Topological Path Finding
Problem Formulation
This task assesses the spatial reasoning ability of the model by asking it to determine the shortest path between two specific buses in a given power grid state. The grid state… See the full description on the dataset page: https://huggingface.co/datasets/serval-uni-lu/orc-bench.McKinsey-Reportsmeta-llama/synthetic-data-kit
https://github.com/meta-llama/synthetic-data-kit
McKinsey reports
https://www.mckinsey.com/featured-insights/insights-store
AISE-Bench
AISE-Bench: A Full-Cycle Curated Benchmark for Information Seeking on Academic Knowledge Graphs
🌐 Project Page •
💻 GitHub •
📖 KDD 2026 Paper
AISE-Bench is a real-world benchmark for information seeking on academic knowledge graphs. It is built from authentic AMiner user search queries and provides human-verified academic question-answering data with executable multi-step API trajectories, standardized tool inputs, API execution outputs, and… See the full description on the dataset page: https://huggingface.co/datasets/zai-org/AISE-Bench.German-RAG-ORPO-Alpaca-HESSIAN-AI
German-RAG-ORPO (Odds Ratio Preference Optimization) Alpaca-Format
German-RAG - German Retrieval Augmented Generation
Dataset Summary
The ORPO Tasks Dataset represents a specialized collection for fine-tuning language models with a focus on RAG-specific capabilities.
The subsets can be for this training step are derived from 2 different sources:
SauerkrautLM Preference Datasets:
SauerkrautLM-Fermented-GER-DPO: is a specialized dataset designed for training… See the full description on the dataset page: https://huggingface.co/datasets/avemio/German-RAG-ORPO-Alpaca-HESSIAN-AI.German-RAG-ORPO-ShareGPT-HESSIAN-AI
German-RAG-ORPO (Odds Ratio Preference Optimization) ShareGPT-Format
German-RAG - German Retrieval Augmented Generation
Dataset Summary
The ORPO Tasks Dataset represents a specialized collection for fine-tuning language models with a focus on RAG-specific capabilities.
The subsets can be for this training step are derived from 3 different sources:
SauerkrautLM Preference Datasets:
SauerkrautLM-Fermented-GER-DPO: is a specialized dataset designed for training… See the full description on the dataset page: https://huggingface.co/datasets/avemio/German-RAG-ORPO-ShareGPT-HESSIAN-AI.Oracle_Kaggle
Kaggle Oracle Dataset
Expert Instruction-Following Data for Competitive Machine Learning
Overview
The Kaggle Oracle Dataset is a high-quality collection of instruction-response pairs tailored for fine-tuning LLMs to provide expert guidance in Kaggle competitions. Built from 14.9M+ kernels and 9,700 competitions, this is the most comprehensive dataset for competitive ML strategy.
Highlights
175 expert-curated instruction-response pairs
100% real-world Kaggle… See the full description on the dataset page: https://huggingface.co/datasets/Aktraiser/Oracle_Kaggle.neptun.scraper
Data in this dataset
Docker & NPM
Scraped using crawl4ai.
The NPM and Docker data was scraped from docs.docker.com and docs.npmjs.com and processed using GPT-4 resulting in docker_documentation.jsonl and npm_documentation.jsonl.
The file training-data-v1.jsonl also includes Titanium, dockerNLcommands and docker_ps.
GitHub
Scraped using firecrawl.
The GitHub data was scraped from docs.github.com/en using firecrawl.A few pages might be missing in the… See the full description on the dataset page: https://huggingface.co/datasets/neptun-org/neptun.scraper.ConfRAG
ConfRAG: Benchmarking LLM Reasoning over Conflicting Web References
ConfRAG is a benchmark dataset for evaluating how large language models (LLMs) reason over conflicting information in retrieval-augmented generation (RAG) settings.
Paper: Benchmarking LLM's Capability in Reasoning over Conflicting Web References
Yizhen Yuan, Rui Kong, Dongze Li, Yuanchun Li†, Yunxin Liu
Institute for AI Industry Research (AIR), Tsinghua University · Shanghai Jiao Tong University
Accepted at ACL… See the full description on the dataset page: https://huggingface.co/datasets/OracleY/ConfRAG.lean-proof-or-refute-300
Lean Proof-or-Refute 300
Lean Proof-or-Refute 300 is a compact collection of 300 formal reasoning
problems grounded in Lean 4 and Mathlib. Each problem starts from a verified
Mathlib theorem, makes one small numerical or operator mutation, and asks the
model to return either:
a Lean certificate proving the mutated proposition; or
a Lean certificate proving the exact negation of the complete proposition.
The model receives the related source theorem, a bounded source excerpt… See the full description on the dataset page: https://huggingface.co/datasets/xlr8harder/lean-proof-or-refute-300.Orvieto-42k
Orvieto-42k — a small atlas dataset for better Italian conversations
Italian is at its best when a reply is not merely correct, but helpful, clear, and pleasant to read. Orvieto-42k is a traceable Italian instruction corpus for building that kind of assistant: 41,873 chat-ready examples, deliberately close to—but not exactly—42k. It packages a consistent SFT representation, per-record provenance, a reproducible heuristic release gate, and an audit trail that can travel with the… See the full description on the dataset page: https://huggingface.co/datasets/OysterCoreAI/Orvieto-42k.oreolook-research-evals
OreoLook Research Evaluations
A small, inspectable evaluation suite for current-information search agents. It
covers routing, clarification, conversational continuity, citation discipline,
freshness, PDF artifacts, protocol safety, and semantic-cache equivalence.
This repository accompanies:
OreoLook live Space
OreoLook source
Architecture paper
Project website
Contents
Configuration
Rows
Purpose
evaluations
24
Synthetic user turns and expected agent… See the full description on the dataset page: https://huggingface.co/datasets/p-research/oreolook-research-evals.or-knowledge-copilot-corpus
OR Knowledge Copilot Corpus
Multi-layer operations-research knowledge base used by OR Knowledge Copilot.
Each instance is stored as six chunks:
Natural language
Mathematical formulation
Pyomo template
MiniZinc template
Solver output
Explanation of binding constraints
Files
chunks.jsonl — retrieval units
qa_pairs.jsonl — labeled questions including out-of-scope abstention cases
benchmark_report.json / eval_results.json — published retrieval metrics
taxonomy.json… See the full description on the dataset page: https://huggingface.co/datasets/alirezaaminzadeh/or-knowledge-copilot-corpus.orbit-seeds
[!NOTE]
For more information on the ORBIT dataset, go check out the preprint available at arxiv.org/abs/2604.01195.
ORBIT: A Synthetic Training Dataset for Search Agents
ORBITis a reasoning-intensive synthetic dataset with complex queries used for training search agents, generated without relying on any paid API services or manual annotation.
Orbit Seeds
Seed entities collected from English Wikipedia, organised by domain. Each record is a Wikipedia page that… See the full description on the dataset page: https://huggingface.co/datasets/orbit-ai/orbit-seeds.evolutionary-origin-ontology
Licensing
The source text El hijo de José states that it is licensed under Creative Commons Attribution-NonCommercial(CC BY-NC). Commercial use of source-derived material requires explicit written permission from the rights holder.
Dataset Card for Evolutionary Origin Ontology
A high-density instruction-tuning corpus for mapping systemic human contradictions to ontological resolutions through the framework of Inversion, correct naming, captured life-energy… See the full description on the dataset page: https://huggingface.co/datasets/elhijodeJose/evolutionary-origin-ontology.German-RAG-ORPO-Long-Context-ShareGPT-HESSIAN-AI
German-RAG-ORPO (Odds Ratio Preference Optimization) Long Context ShareGPT-Format
German-RAG - German Retrieval Augmented Generation
Dataset Summary
The ORPO Long Context Tasks Dataset represents a specialized collection for fine-tuning language models with a focus on RAG-specific capabilities.
The subsets are derived from Synthetic generation inspired by Tencent's (“Scaling Synthetic Data Creation with 1,000,000,000 Personas”).
Dataset Structure… See the full description on the dataset page: https://huggingface.co/datasets/avemio/German-RAG-ORPO-Long-Context-ShareGPT-HESSIAN-AI.ToCall_or_NotToCall
To Call or Not to Call: Tool-Calling Evaluation Datasets
Datasets accompanying the paper "To Call or Not to Call: A Framework to Assess and Optimize LLM Tool Calling" (Wu et al., 2026).
The paper evaluates when language models should call external tools (web search, calculator) rather than answering from parametric knowledge, from both a normative perspective (when is a tool call truly needed/useful?) and a descriptive perspective (when does the model think it needs a tool, and… See the full description on the dataset page: https://huggingface.co/datasets/QinyuanWu/ToCall_or_NotToCall.patent-strategist-bench-v0.1
Patent-Strategist Bench v0.1
A 200-question, seven-shape benchmark for patent-prosecution reasoning, anchored
to three public sources (USPTO MPEP, HPI-Naumann PatentMatch, BIGPATENT) with
oracle context attached to every row. Built to evaluate whether a small open
LLM can perform the day-to-day reasoning tasks of a patent practitioner.
Companion artifact to two methodology articles:
Patent-Strategist v1 baseline on Spark — establishes the first tri-mode (closed-book / retrieval /… See the full description on the dataset page: https://huggingface.co/datasets/Orionfold/patent-strategist-bench-v0.1.orbis-coder
Orbis Coder Dataset (10K)
A coding-first instruction dataset to train or fine-tune assistants that behave like Orbis Coder — friendly, concise, practical, and focused on helping people build, debug, and ship software.
This dataset is intended for:
instruction-tuning / SFT
LoRA / QLoRA fine-tunes
“persona + skill” alignment for coding assistants
quick experiments + dataset viewer testing
What this dataset contains
Most rows are coding help across many languages and… See the full description on the dataset page: https://huggingface.co/datasets/xlelords/orbis-coder.ClosedBookQA-eu
ClosedBookQA-eu dataset for Basque
ClosedBookQA-eu, is a closed-book question answering (QA) dataset for Basque that was constructed from three sources: The MCQA, Belebele-eus dataset (Bandarkar et al., 2024), the MCTest dataset (Richardson et al., 2013), and semi-automatically generated examples based on news content.
Belebele (train, dev and QA-hard test)
Belebele* (Bandarkar et al., 2024) is a multiple-choice QA (MCQA) dataset that includes a passage (context), a… See the full description on the dataset page: https://huggingface.co/datasets/orai-nlp/ClosedBookQA-eu.Vietnamese-microsoft-orca-math-word-problems-200k-gg-translatedorbital-mechanics-instruct-32
🛰️ Orbital Mechanics Instruction Dataset
Expert-crafted prompt–completion pairs for fine-tuning LLMs on space mission analysis and design
📋 Dataset Summary
A curated dataset of 32 expert-crafted instruction–completion pairs designed for fine-tuning large language models on orbital mechanics and space mission analysis tasks. Each example contains a natural-language problem statement paired with a structured, step-by-step solution featuring properly formatted… See the full description on the dataset page: https://huggingface.co/datasets/abdohisham12/orbital-mechanics-instruct-32.