datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
Data-Prep-Bench
Data-Prep-Bench
Dataset Overview
This dataset is a comprehensive resource built for Supervised Fine-Tuning (SFT) and evaluation of Large Language Models (LLMs), covering six domains: Finance, Medicine, Law, Mathematics, Science, and General.
A key feature of this dataset is that we employed 12 different data generation methods (including Agent-based methods, DataFlow series, pure LLM-based generation, and a SKILL method) using multiple cutting-edge models (such as GPT-5… See the full description on the dataset page: https://huggingface.co/datasets/yibotongxue/Data-Prep-Bench.FineEdit_bench
FineEdit Dataset
Paper | GitHub Repository
Introduction
This repository contains InstrEditBench, a high-quality benchmark dataset introduced in the paper Bridging the Editing Gap in LLMs: FineEdit for Precise and Targeted Text Modifications accpeted by EMNLP 2025
Large Language Models (LLMs) have significantly advanced natural language processing,
demonstrating strong capabilities in tasks such
as text generation, summarization, and reasoning. Recently, their potential… See the full description on the dataset page: https://huggingface.co/datasets/YimingZeng/FineEdit_bench.LoGos-Rollout-1K
LoGos-Rollout-1K
Resources
Paper: Mixing Expert Knowledge: Bring Human Thoughts Back To the Game of Go
GitHub Repository: https://github.com/Entarochuan/LoGos
Associated Model: LoGos-7B
Citation
@misc{ma2026mixingexpertknowledgebring,
title={Mixing Expert Knowledge: Bring Human Thoughts Back To the Game of Go},
author={Yichuan Ma and Linyang Li and Yongkang Chen and Peiji Li and Jiasheng Ye and Qipeng Guo and Dahua Lin and Kai Chen}… See the full description on the dataset page: https://huggingface.co/datasets/YichuanMa/LoGos-Rollout-1K.Expert-Go-SFT-100K
Expert-Go-SFT-100K
Paper | Code
Expert-Go-SFT-100K is a large-scale synthetic dataset designed to "cold start" Large Language Models (LLMs) for Go-related reasoning tasks. It was introduced as part of the LoGos project, which aims to bridge the gap between general-purpose LLM reasoning and specialized expert knowledge in the game of Go.
The dataset features 100,000 samples of structured Go expertise mixed with general long Chain-of-Thought (CoT) reasoning data. It enables models to… See the full description on the dataset page: https://huggingface.co/datasets/YichuanMa/Expert-Go-SFT-100K.aime24-official
AIME 2024 — official wording, figures retained
All 30 problems from the 2024 American Invitational Mathematics Examination (AIME I and AIME II),
transcribed from the official exam text with every figure retained as Asymptote source.
This exists because the circulating text-only versions of AIME 2024 are not faithful to the
official problems, and at least one problem in them cannot be solved as written.
Why this dataset exists
While evaluating a reasoning model on… See the full description on the dataset page: https://huggingface.co/datasets/YichengWangCA/aime24-official.Go-GRPO-1K
Go-GRPO-1K
Paper | Code
Project Context
The LoGos model uses this dataset to transfer reasoning capabilities acquired from long CoT data to Go tasks. Through mixed fine-tuning and reinforcement learning, the model learns to perform analysis, reasoning, and summarization to select optimal moves on the Go board.
Citation
If you find this dataset useful for your research, please cite:
@misc{ma2026mixingexpertknowledgebring,
title={Mixing Expert Knowledge:… See the full description on the dataset page: https://huggingface.co/datasets/YichuanMa/Go-GRPO-1K.PaperWritingBench
PaperWritingBench 🎻
PaperWritingBench is the first benchmark designed to evaluate how well autonomous AI research paper writing systems can synthesize raw research materials into submission-ready papers.
[Paper] [Project Page] [Code]
Dataset Structure
This repository contains:
datasets.zip: The full dataset containing cvpr2025 and iclr2025 folders with raw materials.
metadata.json: A JSON file listing metadata for all 200 papers, including venue, paper ID, number… See the full description on the dataset page: https://huggingface.co/datasets/yiwen-song/PaperWritingBench.YiSang-HighQuality-chatml-v1
YiSang-HighQuality ChatML (Korean) v1
KOREAson/YiSang-HighQuality를 한국어 SFT용으로 가공한 데이터셋입니다. 원본 response에 포함된 <think>...</think> 영어 추론 트레이스를 전부 제거하고 실제 답변만 남긴 뒤, 정제 → 품질 필터 → 안전성 필터 → 중복 제거 → ChatML 포맷팅 → 토크나이즈 검증을 거친 결과물입니다.
총 샘플 수: 259,596
총 토큰 수: 약 1.78억 (178,406,073 tokens, keural tokenizer 기준, 평균 687 tokens/sample)
포맷: ChatML (<|im_start|>role ... <|im_end|>)
최대 길이: 8,192 tokens (초과 시 truncate)
추론 트레이스: 최종 산출물 전수 검사 기준 <think>/</think> 잔존 0건
생성일: 2026-07-10
데이터… See the full description on the dataset page: https://huggingface.co/datasets/mkd-jueon/YiSang-HighQuality-chatml-v1.chat-darija-therapy
Moroccan Darija Therapy Conversations Dataset
This dataset is entirely synthetic and contains no real patient information.
It is provided strictly for research, educational, and experimental purposes and must not be used for clinical, medical, diagnostic, or psychological decision-making.
Citation
If you use this dataset in your research, please cite:
@dataset{moroccan_darija_therapy_conversations,
title={Moroccan Darija Therapy Conversations},
author={Jamal… See the full description on the dataset page: https://huggingface.co/datasets/yibba/chat-darija-therapy.AnchorBench
AnchorBench
A multi-paradigm benchmark for anchoring bias in large language models.
Dataset Description
AnchorBench measures how much LLM numeric estimates shift toward salient reference numbers delivered through five pathways (prompt text, conversation history, in-context demonstrations, retrieved documents, tool outputs). Each item is presented under matched conditions that share the same evidence and gold answer; only the anchor changes. The benchmark… See the full description on the dataset page: https://huggingface.co/datasets/Yiderigun/AnchorBench.PhysGym
Dataset Card for PhysGym Dataset
This is the dataset file for PhysGym, containing 97 carefully curated physics problems designed for evaluating interactive scientific discovery capabilities of Large Language Model (LLM)-based agents.
Dataset Details
Dataset Description
This dataset provides the problem collection for the PhysGym benchmark suite, which evaluates how AI agents discover physical laws through interactive experimentation. Each entry in this dataset… See the full description on the dataset page: https://huggingface.co/datasets/YimengChen/PhysGym.Skill2-Bench
Skill²-Bench
Skill²-Bench is a benchmark of multi-step tasks that force LLMs to switch between skills, introduced in the paper "Toward Skill-Native LLMs: Skill Entropy for Benchmarking and Training Long-Horizon Reasoning".
Long-horizon tasks require models to switch between skills, not just execute a single skill well. Each Skill²-Bench task embeds a sequence of 2–10 steps in a coherent real-world scenario, where consecutive steps draw on different skills (e.g., algorithm design… See the full description on the dataset page: https://huggingface.co/datasets/yinghuihe/Skill2-Bench.vidiary-reflective-prompts
ViDiary Reflective Prompts & Emotional Taxonomy Dataset
This open dataset contains foundational reflective journaling prompts and emotional sentiment taxonomy used in the development of ViDiary — the AI-powered voice and video journal with dual-PIN Decoy Vault.
🎙️ About ViDiary
ViDiary is an innovative mobile application engineered to solve the #1 psychological hurdle in personal self-care: Bedtime Typing Fatigue.
Research shows that over 80% of… See the full description on the dataset page: https://huggingface.co/datasets/YILMAZB1/vidiary-reflective-prompts.ACL-23-Paper-OCR-Markdown
ACL 2023 Paper in Markdown after OCR
This dataset contains 2150 papers from Association for Computational Linguistics (ACL) 2023:
Long Papers (912 papers)
Short Papers (185 papers)
System Demonstrations (59 paper)
Student Research Workshop (35 papers)
Industry Track (77 papers)
Tutorial Abstracts (7 papers)
Findings (902 papers)
This dataset is processed and compiled by @hu_yifei as part of open-source effort from the Open Research Assistant Project.
OCR process
The… See the full description on the dataset page: https://huggingface.co/datasets/yifeihu/ACL-23-Paper-OCR-Markdown.selfrag_train_dataThis is a training data file for Self-RAG that generates outputs to diverse user queries as well as reflection tokens to call the retrieval system adaptively and criticize its own output and retrieved passages.
Self-RAG is trained on our 150k diverse instruction-output pairs with interleaving passages and reflection tokens using the standard next-token prediction objective, enabling efficient and stable learning with fine-grained feedback.
At inference, we leverage reflection tokens covering… See the full description on the dataset page: https://huggingface.co/datasets/Yibaii/selfrag_train_data.temiz-OSCAR
Dataset Card for Temiz OSCAR
Temiz OSCAR is a corpora collection consisting of cleaned versions of original OSCAR corpora.
This collection is made up of four datasets: OSCAR-2019, OSCAR-2109, OSCAR-2201 and OSCAR-2301
This corpus is a part of large scale Turkish corpus Bella Turca. For more details about Bella Turca, please refer to the publication.
Dataset
num instances
size
num of words
OSCAR-2019
3.671.430
7.7G
976M
OSCAR-2109
8.472.809
18G
2.22B
OSCAR-2201… See the full description on the dataset page: https://huggingface.co/datasets/YigitCahit/temiz-OSCAR.claude-opus-4.6-4.7-reasoning-8.7k
Background
Ended up with some tokens to burn on a Claude Max plan. Assembly began during 4.6 and moved to 4.7. Model is tagged. The development evolved as it went along. The dataset has not been manually reviewed. It's entirely Claude developed.
Clarification on Reasoning
The reasoning is not Claude's actual chain-of-thought (cot) and is not summarized cot. It's a fully synthetic cot created as part of the Assistant response to mimic the type of "thinking"… See the full description on the dataset page: https://huggingface.co/datasets/yikes-liki/claude-opus-4.6-4.7-reasoning-8.7k.MedXpertQA
Dataset Card for MedXpertQA
MedXpertQA is a highly challenging and comprehensive benchmark designed to evaluate expert-level medical knowledge and advanced reasoning capabilities. It features both text-based and multimodal question-answering tasks, with the multimodal subset leveraging structured clinical information alongside images.
Dataset Description
MedXpertQA comprises 4,460 questions spanning diverse medical specialties, tasks, body systems, and image… See the full description on the dataset page: https://huggingface.co/datasets/yiyanhuang/MedXpertQA.Chinese-DeepSeek-R1-Distill-data-110k
中文基于满血DeepSeek-R1蒸馏数据集（Chinese-Data-Distill-From-R1）
🤗 Hugging Face | 🤖 ModelScope | 🚀 Github | 📑 Blog
注意：提供了直接SFT使用的版本，点击下载。将数据中的思考和答案整合成output字段，大部分SFT代码框架均可直接直接加载训练。
本数据集为中文开源蒸馏满血R1的数据集，数据集中不仅包含math数据，还包括大量的通用类型数据，总数量为110K。
为什么开源这个数据？
R1的效果十分强大，并且基于R1蒸馏数据SFT的小模型也展现出了强大的效果，但检索发现，大部分开源的R1蒸馏数据集均为英文数据集。 同时，R1的报告中展示，蒸馏模型中同时也使用了部分通用场景数据集。
为了帮助大家更好地复现R1蒸馏模型的效果，特此开源中文数据集。该中文数据集中的数据分布如下：
Math：共计36568个样本，
Exam：共计2432个样本，
STEM：共计12648个样本，… See the full description on the dataset page: https://huggingface.co/datasets/yifeng222/Chinese-DeepSeek-R1-Distill-data-110k.Aya_Turkish_Filtered
Aya Turkish Filtered
Dataset Description
This dataset is a manually curated and filtered subset of the original Aya Dataset by Cohere For AI. It has been specifically isolated to contain high-quality Turkish instructions and responses, making it highly efficient for supervised fine-tuning (SFT) and instruction tuning of Large Language Models.
This dataset was utilized in a multi-stage fine-tuning process alongside mathematical reasoning datasets to enhance task… See the full description on the dataset page: https://huggingface.co/datasets/Yigit-Karaman/Aya_Turkish_Filtered.question-generation-sft
Question Generation SFT Dataset
题目生成监督微调数据集
数据集描述
样本数: 9,604
任务: 根据学生状态生成个性化练习题
格式: Input (学生状态) → Output (题目内容)
使用方法
from datasets import load_dataset
dataset = load_dataset("yinlin124/question-generation-sft", data_files="data.json")
评估指标
k-Hit: 生成题目知识点命中学生弱项
State-Shift: 状态转移敏感度
BLEU: 文本生成质量
qwen3-omni-pairwise-video-train
Qwen3-Omni Pairwise Video Inference / Evaluation
Pairwise audio-video preference evaluation data for Qwen3-Omni models.
Each sample compares two generated videos (with audio) against a text caption and human/Gemini labels.
Source path on cluster: /inspire/hdd/project/autoregressive-video-generation/public/hym/data/final_train
Upload snapshot: 2026-06-12 10:46 UTC
Repository layout
Contents of final_infer are uploaded to the dataset repo root:
.cache/
ovi_davinci/… See the full description on the dataset page: https://huggingface.co/datasets/YinmingHuang/qwen3-omni-pairwise-video-train.