datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
Light-R1-SFTData
Light-R1: Surpassing R1-Distill from Scratch* with $1000 through Curriculum SFT & DPO
*from models without long COT
technical report
GitHub page
Here are the two-stage SFT data we used to train Light-R1-32B.
Simply refer to stage1-76k.json and stage2-3k.json
Model
Trained From
Release Date
AIME24
AIME25
DeepSeek-R1-Distill-Llama-70B
Llama-3.3-70B-Instruct
25.1.20
70.0
54.1
DeepSeek-R1-Distill-Qwen-32B
Qwen2.5-32B
25.1.20
72.6
54.9
LIMO (32B)
Qwen2.5-32B-Instruct
25.2.4… See the full description on the dataset page: https://huggingface.co/datasets/qihoo360/Light-R1-SFTData.Medical-R1-Distill-Data
Introduction
This dataset is an SFT dataset distilled from Deepseek-R1 (Full Power Version), based on medical verifiable problems from HuatuoGPT-o1.
The Chinese version of the dataset is available at FreedomIntelligence/Medical-R1-Distill-Data-Chinese.
The distillation originates from the native Deepseek-R1 API requests. We hope this distilled dataset can help initialize your models with the reasoning chain from R1. You can also use our previously built medical verified long… See the full description on the dataset page: https://huggingface.co/datasets/FreedomIntelligence/Medical-R1-Distill-Data.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：… See the full description on the dataset page: https://huggingface.co/datasets/Congliu/Chinese-DeepSeek-R1-Distill-data-110k.WideSeek-R1-test-data
Testing Dataset
We provide test.jsonl, a testing split for evaluating WideSeek-R1 on the standard WideSearch dataset. All examples are sourced from WideSearch; we only convert them into a format that is directly compatible with the WideSeek-R1 evaluation scripts. This makes the dataset plug-and-play—no additional configuration required.
Chinese-DeepSeek-R1-Distill-data-110k-SFT
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：该版本为，可以直接SFT使用的版本，将原始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/Congliu/Chinese-DeepSeek-R1-Distill-data-110k-SFT.School-Math-R1-Distil-Chinese-220K从原数据集 BelleGroup/school_math_0.25M 提取指令，然后重新合成回复。
每条数据的格式如下：
{
"id": <<12位nanoid>>,
"prompt": <<提示词>>,
"reasoning": <<模型思考过程>>,
"response": <<模型最终回复>>
}
请注意：本数据集有如下已知缺陷
问题可解性无法保证：这是由于原数据集本身就是纯合成数据集，未经过校验。尽管本数据集已经尽力筛选过滤了一部分，但仍然无法保证余下数据的指令正确性和可解性。
答案未经过校验：所有回答均为合成，且未经过校验。
DeepSeek-R1-Distilled-Translate-en-zh_CN-39k-Alpaca-GPT4
DeepSeek R1 满血蒸馏英中翻译数据集 Alpaca GPT-4（带 CoT 版本）
本数据集是 @FradSer/DeepSeek-R1-Distilled-Translate-en-zh_CN-39k 的 Alpaca GPT-4 版本，专门用于微调语言模型的英中翻译任务。采用标准的指令微调格式，更适合直接用于 SFT（Supervised Fine-tuning）训练。
本项目主要基于以下工具完成数据处理和生成：
llm-tools: 用于大语言模型数据处理的工具集合
数据集概览
关键统计
总样本数：38,981
数据集结构
字段说明
features:
- name: instruction # 待翻译的英文文本
dtype: string
- name: input # 空字符串，保持与标准指令格式一致
dtype: string
- name: output #… See the full description on the dataset page: https://huggingface.co/datasets/FradSer/DeepSeek-R1-Distilled-Translate-en-zh_CN-39k-Alpaca-GPT4.DeepSeek-R1-Distilled-Translate-en-zh_CN-39k
DeepSeek R1 满血蒸馏英中翻译数据集
本数据集是一个专门用于微调语言模型的英中翻译数据集，主要通过DeepSeek R1满血版蒸馏完成。
SFT训练版本
为了方便直接进行监督微调（Supervised Fine-tuning，SFT）训练，我们提供了两个使用标准 instruction-input-output 格式预的处理版本：
带 CoT 版本
保留了翻译过程中的思维链（Chain of Thought）
适合训练具有推理能力的翻译模型
无 CoT 版本
移除了思维链部分，只保留最终翻译结果
更适合训练直接输出翻译结果的模型
数据更简洁，训练更高效
项目依赖
本项目主要基于以下工具完成数据处理和生成：
llm-tools: 用于大语言模型数据处理的工具集合
qa-generator: 基于大语言模型的问答数据生成工具
数据集概览
关键统计
总样本数：38,981
数据集结构
字段说明… See the full description on the dataset page: https://huggingface.co/datasets/FradSer/DeepSeek-R1-Distilled-Translate-en-zh_CN-39k.AM-DeepSeek-R1-Filtered-Math-Code
AM DeepSeek R1 Filtered Math and Code
This repository publishes reproducible training subsets derived from
a-m-team/AM-DeepSeek-R1-Distilled-1.4M
at pinned revision 53531c06634904118a2dcd83961918c4d69d1cdf.
The source dataset and this derived release use CC BY-NC 4.0. Commercial use is
not permitted by that license. Preserve attribution and review the upstream
dataset card before use.
Contents
Config / split
Records
Bytes
SHA-256
math / train
111,657
2… See the full description on the dataset page: https://huggingface.co/datasets/chhao/AM-DeepSeek-R1-Filtered-Math-Code.Medical-R1-Distill-Data-Chinese
Introduction
This dataset is an SFT dataset distilled from Deepseek-R1 (Full Power Version), based on Chinese medical verifiable problems from HuatuoGPT-o1.
The distillation originates from the native Deepseek-R1 API requests. We hope this distilled dataset can help initialize your models with the reasoning chain from R1. You can also use our previously built medical verified long reasoning chains based on GPT-4o on medical-o1-reasoning-SFT.
For details, see our paper and GitHub… See the full description on the dataset page: https://huggingface.co/datasets/FreedomIntelligence/Medical-R1-Distill-Data-Chinese.Math-Chinese-DeepSeek-R1-10K
中文 DeepSeek-R1-Distil 数学指令微调数据集
💻 Github Repo
基本信息
数据集大小 10K，独立生成指令与回复，并非其他社区数据集的子集。所有数据经过校验，答案正确性可以得到保证。
数据集的组成如下：
问题类型
数据条数
定积分计算
2626
多项式化简
1621
因式分解
2557
多项式展开
2095
多项式方程
1101
总数
10000
数据格式
每条数据的格式如下：
{
"id": <<12位nanoid>>,
"prompt": <<提示词>>,
"reasoning": <<模型思考过程>>,
"response": <<模型最终回复>>
}
behaviortune-v1-1-r1
BehaviorTune Dataset
Controlled synthetic dataset used to train and evaluate BehaviorTune, a QLoRA post-training project on Qwen/Qwen3-4B-Instruct-2507.
It contains 544 scenarios across six splits, including 240 training rows, 48 development rows, and a 64-row eval_core set used for the published matched evaluation.
The dataset supports completion-only QLoRA training and deterministic BASE / SYSTEM / CONTEXT / QLoRA evaluation. V1.1-R1 is the frozen dataset/version identifier.… See the full description on the dataset page: https://huggingface.co/datasets/aamish-ahmad/behaviortune-v1-1-r1.Light-R1-DPOData
Light-R1: Surpassing R1-Distill from Scratch* with $1000 through Curriculum SFT & DPO
*from models without long COT
technical report
GitHub page
Here is the DPO data we used to train Light-R1-32B.
Simply refer to dpo-pairs.json
Model
Trained From
Release Date
AIME24
AIME25
DeepSeek-R1-Distill-Llama-70B
Llama-3.3-70B-Instruct
25.1.20
70.0
54.1
DeepSeek-R1-Distill-Qwen-32B
Qwen2.5-32B
25.1.20
72.6
54.9
LIMO (32B)
Qwen2.5-32B-Instruct25.2.4
56.3
47.1
s1.1-32B… See the full description on the dataset page: https://huggingface.co/datasets/qihoo360/Light-R1-DPOData.WideSeek-R1-test-data
Testing Dataset
🌐 Project Page | 📄 Paper | 📖 Doc | 💻 Code | 📦 Dataset | 🤗 Models
We provide test.jsonl, a testing split for evaluating WideSeek-R1 on the standard WideSearchdataset. All examples are sourced from WideSearch; we only convert them into a format that is directly compatible with the WideSeek-R1 evaluation scripts. This makes the dataset plug-and-play—no additional configuration required.
Acknowledgement
Thanks to WideSearch for providing a… See the full description on the dataset page: https://huggingface.co/datasets/RLinf/WideSeek-R1-test-data.WideSeek-R1-SFT-data
WideSeek-R1 SFT Data
This dataset contains agent-level, multi-turn supervised fine-tuning trajectories for both width-only and depth-only tasks in WideSeek-R1.
Construction
The trajectories were generated by Qwen3-235B-A22B using the WideSeek-R1 multi-agent workflow with offline retrieval tools. Width and depth trajectories are balanced at the question level.
For each question-level trajectory, we retain one main-agent session and up to three subagent sessions… See the full description on the dataset page: https://huggingface.co/datasets/WideSeek-R1/WideSeek-R1-SFT-data.hwtcm-deepseek-r1-distill-data
简介
DeepSeek蒸馏的传统中医数据集，原始数据来源于网络，未进行人工审查。
7B模型微调效果
模型表现出了推理能力，准确性有待继续验证。
我们的其他产品
中医NER：能识别方剂、本草、来源、病名、症状、证型，也许是基于BERT开源模型中识别最好的模型。中医考试题：也许是全网最早开源、数据最多的中医考试题，我们内部将其用于模型训练的性能评测数据集。中医SFT数据集：中医QA数据集，用于SFT微调。仓公：基于Qwen的指令微调模型（暂未开源）。仓公R1：基于DeepSeek蒸馏的超过100万条QA的指令微调模型，拥有强大的推理能力（暂未开源）。
。。。还有很多
Citation
If you find this project useful in your research, please consider cite:
@misc{hwtcm2024,
title={{hwtcm-deepseek-r1-distill-data} A traditional… See the full description on the dataset page: https://huggingface.co/datasets/Monor/hwtcm-deepseek-r1-distill-data.TOMATO-Star-SFT-Data-R1D-32B
TOMATO-Star SFT Data (R1D-32B)
SFT training data for the two core tasks in MOOSE-Star: Hypothesis Composition (HC) and Inspiration Retrieval (IR).
All data is generated via rejection sampling with DeepSeek-R1-Distill-Qwen-32B as the teacher model, followed by reranker filtering.
All data is in ShareGPT JSONL format, directly compatible with LLaMA-Factory.
Files
Hypothesis Composition (HC)
File
Samples
Description
HC/normal_composition.jsonl… See the full description on the dataset page: https://huggingface.co/datasets/ZonglinY/TOMATO-Star-SFT-Data-R1D-32B.rank1-R1-MSMARCO
rank1-R1-MSMARCO: Reasoning Outputs from MS MARCO Dataset
📄 Paper | 🚀 GitHub Repository
This dataset contains outputs from Deepseek's R1 model on the MS MARCO passage dataset, used to train rank1. It showcases the reasoning chains and relevance judgments generated when determining document relevance for information retrieval queries.
Dataset Description
The rank1-R1-MSMARCO dataset consists of reasoning chains and relevance judgments produced on the MS MARCO passage… See the full description on the dataset page: https://huggingface.co/datasets/jhu-clsp/rank1-R1-MSMARCO.r1-reasoning-tr
R1 Reasoning TR
This is an R1 reasoning dataset translated into Turkish, containing conversations between users and assistants. Thanks to lightblue for the dataset.
License
This dataset is released under the Apache 2.0 License.
NOESIS-1M-reasoning-router-code-math-psych-opus47-deepseek4-qwen36-gemini31-r1-gpt54
NOESIS DORA SFT Dataset
Multilingual supervised fine-tuning dataset built for the NOESIS QwQ+DeepSeek-R1 MoE pipeline.
Released as part of the NOESIS Professional Multilingual Dubbing Automation Platform(framework: DHCF-FNO — Deterministic Hybrid Control Framework for Frozen Neural Operators).
Founder: Ilia Bolotnikov
Organization: AMAImedia.com
X (Twitter): @AMAImediacom
LinkedIn: Ilia Bolotnikov
Telegram: @djbionicl
NOESIS version: v14.8-NT89
Build date: 2026-04… See the full description on the dataset page: https://huggingface.co/datasets/SMH-DEV-AI/NOESIS-1M-reasoning-router-code-math-psych-opus47-deepseek4-qwen36-gemini31-r1-gpt54.r15-synthetic-cohorts
R15 Synthetic Cohort Differentiation
Synthetic cohort differentiation experiment (Run 15 + Run 15b Latin-square robustness) from the Spectral Brand Theory research program.
Dataset Summary
1,200 API calls (800 main + 400 robustness) testing whether the PRISM-B instrument differentiates synthetic observer cohorts defined by behavioral vignettes (no SBT dimension vocabulary in prompts). The dataset operationalizes the central R15 question: do large language models… See the full description on the dataset page: https://huggingface.co/datasets/spectralbranding/r15-synthetic-cohorts.Natural-Reasoning-R1-10k-dpo
Dataset Card for Natural-Reasoning-R1-10k-dpo
本資料集是基於 facebook/natural_reasoning 之 10k 子集，所建構的 R1-style DPO 偏好集，並翻譯為繁體中文：每筆樣本針對一個自然推理問題，提供帶完整 <think> 思考過程的 chosen 與品質較差的 rejected。
Dataset Details
Dataset Description
facebook/natural_reasoning 收集涵蓋多領域（科技、安全、組織管理、學術寫作等）的開放式推理問題。本資料集從中取 10k 子集，將題幹翻譯為繁中、並由 R1 風格的 LLM 產出 chosen，由較弱模型產出 rejected。
每筆樣本維持 LLaMA-Factory 風格的 conversations + chosen + rejected 結構；assistant 回答前段為 <think>...</think> 推理過程。
Curated by:… See the full description on the dataset page: https://huggingface.co/datasets/lianghsun/Natural-Reasoning-R1-10k-dpo.lemonseed-langmix-r1
lemonseed-langmix-r1
LemonSeed — language round 1 mix (fineweb-edu 28% / dialog 20% / vocab 10% / wordproblems 15% / addition 12% / go+sudoku 15%).
Format
JSON Lines (.jsonl), one example per line.
Provenance & License
Blend: fineweb-edu prose (ODC-By) + dialog (daily_dialog CC BY-NC-SA 4.0 + ultrachat MIT) + WordNet vocab + synthetic arithmetic/games (Apache-2.0). The NC term of daily_dialog applies to the dialog portion.
gs8k_thai_r1_example
Additional Information
This dataset contains mathematical problem-solving traces generated using the CAMEL framework. Each entry includes:
A mathematical problem statement
A detailed step-by-step solution
An improvement history showing how the solution was iteratively refined
Think in English, Context Thai for improve thai question
coyote-r1-training-data
coyote-r1-training-data
Coyote the Coder — training round R1. Instruction/tool-call data (system+user+assistant message triples with task labels) used to fine-tune the Coyote coder model (Qwen3.5-4B-Base).
Contents
train.jsonl (238 rows)
validation.jsonl (42 rows)
Format
JSON Lines (.jsonl), one example per line.
Provenance
Original content authored for the Coyote the Coder project (Michael Anthony Falabella).
R1-Distilled-Data-60K
This dataset contains 60,856 text prompt pairs before and after deepseek-R1 enhancement, is used to train the language model LeX-Enhancer for prompt enhancement model distilled from DeepSeek-R1-Qwen-Distilled-14B.
medical_r1_distill_sft_Chinese_openaiFormat from https://huggingface.co/datasets/openai/gsm8k
medical_r1_distill_sft_Chinese_alpacaBorrowed from https://huggingface.co/datasets/FreedomIntelligence/Medical-R1-Distill-Data-Chinese
Reorganized the data structure for easier SFT.
HuggingFaceH4-MATH-R1-dpo
Dataset Card for HuggingFaceH4-MATH-R1-dpo
本資料集為 HuggingFaceH4/MATH-500 等 MATH 評測資料延伸的 R1-style DPO 偏好集，並翻譯為繁體中文：每筆樣本針對一道數學題，提供「具完整 R1 風格 <think> 推理過程的較佳解（chosen）」與「推理品質較差的解（rejected）」。
Dataset Details
Dataset Description
資料以 MATH 風格題目為基礎（代數、幾何、組合、機率等），題幹翻譯為繁中，並使用 DeepSeek R1 等具 R1 風格之模型產出 chosen，配合較弱模型／簡化 prompt 的版本作為 rejected，以建構偏好對。
格式採 LLaMA-Factory 之 conversations + chosen + rejected 風格，gpt 回答前段為 <think>...</think> 思考段落，後段為最終解答（多以 \boxed{...} 收尾）。… See the full description on the dataset page: https://huggingface.co/datasets/lianghsun/HuggingFaceH4-MATH-R1-dpo.gene-r1-go-sft-table2-reconstructed
Gene-R1 GO SFT Table 2 Reconstructed Splits
Small workshop dataset used for tokenizer-transfer experiments with ncbi/Gene-R1-1B.
Rows are reconstructed from released Gene Ontology benchmark materials into Gene-R1-style prompt/completion text for tokenizer transfer experiments:
train.jsonl: 2400 rows, 800 BP + 800 MF + 800 CC
validation.jsonl: 300 rows, 100 BP + 100 MF + 100 CC
test.jsonl: 300 rows, 100 BP + 100 MF + 100 CC
Each row contains metadata plus prompt, completion… See the full description on the dataset page: https://huggingface.co/datasets/transhumanist-already-exists/gene-r1-go-sft-table2-reconstructed.