OneThinker-train-data

# OneThinker-600k Training Data

This repository contains the training data for **OneThinker**, an all-in-one reasoning model for image and video, as presented in the paper [OneThinker: All-in-one Reasoning Model for Image and Video](https://arxiv.org/abs/2512.03043).

**Code**: [https://github.com/tulerfeng/OneThinker](https://github.com/tulerfeng/OneThinker)

<div align="center">
<img src="https://github.com/tulerfeng/OneThinker/blob/main/assets/teaser.png?raw=true" alt="OneThinker teaser" width="95%">

</div>

## About the OneThinker Dataset

**OneThinker-600k** is a large-scale multi-task training corpus designed to train `OneThinker`, an all-in-one multimodal reasoning model capable of understanding images and videos across diverse fundamental visual tasks. This corpus includes **OneThinker-SFT-340k**, which features high-quality Chain-of-Thought (CoT) annotations produced by a strong proprietary model (Seed1.5-VL) for effective Supervised Fine-Tuning (SFT) cold start.

<div align="center">
<img src="https://github.com/tulerfeng/OneThinker/blob/main/assets/dataset.png?raw=true" alt="OneThinker dataset" width="95%">
</div>

The dataset covers both image and video modalities and spans a series of fundamental visual reasoning tasks, including:
* Rule-based Question Answering (QA)
* Open-ended Question Answering (QA)
* Captioning
* Spatial Grounding
* Temporal Grounding
* Spatio-Temporal Grounding
* Tracking
* Segmentation

## Dataset Files

The OneThinker training data consists of several JSON files tailored for different training stages:
* `onethinker_rl_train.json`: Used for Reinforcement Learning (RL) training.
* `onethinker_sft_image.json`: Used for Supervised Fine-Tuning (SFT) cold start on image data.
* `onethinker_sft_video.json`: Used for Supervised Fine-Tuning (SFT) cold start on video data.

Files ending with `_unsampled` represent the full, unsampled versions of these datasets.

## Citations

If you find our work helpful for your research, please consider citing our work:

```bibtex
@article{feng2025onethinker,
title={OneThinker: All-in-one Reasoning Model for Image and Video},
author={Feng, Kaituo and Zhang, Manyuan and Li, Hongyu and Fan, Kaixuan and Chen, Shuang and Jiang, Yilei and Zheng, Dian and Sun, Peiwen and Zhang, Yiyuan and Sun, Haoze and others},
journal={arXiv preprint arXiv:2512.03043},
year={2025}
}
```

# OneThinker-600k 训练数据集

本仓库包含**OneThinker**的训练数据，该模型是一款面向图像与视频的全功能推理模型，相关研究成果发表于论文《OneThinker: All-in-one Reasoning Model for Image and Video》（https://arxiv.org/abs/2512.03043）。

**代码**：[https://github.com/tulerfeng/OneThinker](https://github.com/tulerfeng/OneThinker)

<div align="center">
<img src="https://github.com/tulerfeng/OneThinker/blob/main/assets/teaser.png?raw=true" alt="OneThinker 概览图" width="95%">
</div>

## OneThinker 数据集简介

**OneThinker-600k**是一款大规模多任务训练语料库，用于训练`OneThinker`——一款可适配多样化基础视觉任务的图像与视频全模态推理模型。该语料库包含**OneThinker-SFT-340k**，其内置由自研高性能模型（Seed1.5-VL）生成的高质量思维链（Chain-of-Thought, CoT）标注数据，可用于高效的监督微调（Supervised Fine-Tuning, SFT）冷启动训练。

<div align="center">
<img src="https://github.com/tulerfeng/OneThinker/blob/main/assets/dataset.png?raw=true" alt="OneThinker 数据集结构" width="95%">
</div>

本数据集涵盖图像与视频两种模态，覆盖一系列基础视觉推理任务，包括：
* 基于规则的问答（QA）
* 开放式问答（QA）
* 字幕生成
* 空间定位
* 时序定位
* 时空定位
* 目标跟踪
* 语义分割

## 数据集文件说明

OneThinker训练数据包含多个针对不同训练阶段定制的JSON格式文件：
* `onethinker_rl_train.json`：用于强化学习（Reinforcement Learning, RL）训练。
* `onethinker_sft_image.json`：用于图像数据的监督微调（Supervised Fine-Tuning, SFT）冷启动训练。
* `onethinker_sft_video.json`：用于视频数据的监督微调（Supervised Fine-Tuning, SFT）冷启动训练。

文件名后缀为`_unsampled`的文件代表对应数据集的完整未采样版本。

## 引用方式

若您的研究工作中用到了本项目，请引用如下论文：
bibtex
@article{feng2025onethinker,
title={OneThinker: All-in-one Reasoning Model for Image and Video},
author={Feng, Kaituo and Zhang, Manyuan and Li, Hongyu and Fan, Kaixuan and Chen, Shuang and Jiang, Yilei and Zheng, Dian and Sun, Peiwen and Zhang, Yiyuan and Sun, Haoze and others},
journal={arXiv preprint arXiv:2512.03043},
year={2025}
}