PixelReasoner-RL-Data

**Overview.**
The RL data for training [**Pixel Reasoner**: Incentivizing Pixel-Space Reasoning with Curiosity-Driven Reinforcement Learning](https://arxiv.org/abs/2505.15966),
The queries require fine-grained visual analysis in both images (e.g., infographics, visually-rich scenes, etc) and videos.

**Details.**
The data includes 15,402 training queries with verifierable answers. The key fields include:
- question, answer, qid
- is_video: a flag to distinguish video and image queries
- image: a list of image paths.
- For video-based queries, the path list corresponds to 16 extracted frames from the original video
- For image-based queries, the path list corresponds to multiple images

**Note**: Remember to unzip `images.zip` and `videos.zip`, and replace the relative path in `image` field with the absolute path.

**Training Code**: The RL code can be found at https://github.com/TIGER-AI-Lab/Pixel-Reasoner/tree/main

**Project page**: https://tiger-ai-lab.github.io/Pixel-Reasoner/

**概述**
本数据集为训练[**Pixel Reasoner**：基于好奇心驱动强化学习（Reinforcement Learning，RL）的像素空间推理激励框架](https://arxiv.org/abs/2505.15966)所用的强化学习数据集，其查询任务要求对图像（如信息图表、视觉丰富场景等）与视频均进行细粒度视觉分析。

**详情**
本数据集包含15402条带有可验证答案的训练查询。核心字段包括：
- 问题（question）、答案（answer）、查询ID（qid）
- 模态标识（is_video）：用于区分视频类与图像类查询
- 图像路径列表（image）：
- 针对视频类查询，该路径列表对应原始视频提取的16帧图像
- 针对图像类查询，该路径列表对应多张图像

**注意**：请务必解压`images.zip`与`videos.zip`，并将`image`字段中的相对路径替换为绝对路径。

**训练代码**：强化学习训练代码可在https://github.com/TIGER-AI-Lab/Pixel-Reasoner/tree/main 获取。

**项目主页**：https://tiger-ai-lab.github.io/Pixel-Reasoner/