Datasets:
Update README.md
Browse files
README.md
CHANGED
@@ -6,9 +6,19 @@ task_categories:
|
|
6 |
- image-text-to-text
|
7 |
pretty_name: PixelReasoner-RL-Data
|
8 |
---
|
|
|
|
|
9 |
|
10 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
11 |
|
12 |
-
|
13 |
|
14 |
-
|
|
|
|
|
|
6 |
- image-text-to-text
|
7 |
pretty_name: PixelReasoner-RL-Data
|
8 |
---
|
9 |
+
**Overview.**
|
10 |
+
The RL data for training [**Pixel Reasoner**: Incentivizing Pixel-Space Reasoning with Curiosity-Driven Reinforcement Learning](https://arxiv.org/abs/2505.15966).
|
11 |
|
12 |
+
**Details.**
|
13 |
+
The data includes 15,402 training queries with verifierable answers. The key fields include:
|
14 |
+
- question, answer, qid
|
15 |
+
- is_video: a flag to distinguish video and image queries
|
16 |
+
- image: a list of image paths.
|
17 |
+
- For video-based queries, the path list corresponds to 16 extracted frames from the original video
|
18 |
+
- For image-based queries, the path list corresponds to multiple images
|
19 |
|
20 |
+
**Note**: Remember to unzip `images.zip` and `videos.zip`, and replace the relative path in `image` field with the absolute path.
|
21 |
|
22 |
+
**Training Code**: The RL code can be found at https://github.com/TIGER-AI-Lab/Pixel-Reasoner/tree/main
|
23 |
+
|
24 |
+
**Project page**: https://tiger-ai-lab.github.io/Pixel-Reasoner/
|