Datasets:

Modalities:
Image
Text
Formats:
parquet
Languages:
English
ArXiv:
Libraries:
Datasets
pandas
License:
File size: 1,207 Bytes
1010213
 
 
e02f644
 
 
1010213
fbc7e7a
 
 
 
 
1010213
5a956a4
59492f9
 
1010213
5a956a4
 
 
 
 
 
 
f4500ec
5a956a4
e02f644
5a956a4
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
---
language:
- en
license: apache-2.0
task_categories:
- image-text-to-text
pretty_name: PixelReasoner-RL-Data
configs:
- config_name: default
  data_files:
  - split: train
    path: "release.parquet"
---
**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/