---
license: mit
datasets:
- VisualSphinx/VisualSphinx-Seeds
language:
- en
base_model:
- Qwen/Qwen2.5-VL-7B-Instruct
---
# 🦁 VisualSphinx: Large-Scale Synthetic Vision Logic Puzzles for RL
VisualSphinx is the largest fully-synthetic open-source dataset providing vision logic puzzles. It consists of over **660K** automatically generated logical visual puzzles. Each logical puzzle is grounded with an interpretable rule and accompanied by both correct answers and plausible distractors.
- 🌐 [Project Website](https://visualsphinx.github.io/) - Learn more about VisualSphinx
- 📖 [Technical Report](https://arxiv.org/abs/2505.23977) - Discover the methodology and technical details behind VisualSphinx
- 🔧 [Github Repo](https://github.com/VisualSphinx/VisualSphinx) - Access the complete pipeline used to produce VisualSphinx-V1
- 🤗 HF Datasets:
  - [VisualSphinx-V1 (Raw)](https://huggingface.co/datasets/VisualSphinx/VisualSphinx-V1-Raw);
  - [VisualSphinx-V1 (For RL)](https://huggingface.co/datasets/VisualSphinx/VisualSphinx-V1-RL-20K); 
  - [VisualSphinx-V1 (Benchmark)](https://huggingface.co/datasets/VisualSphinx/VisualSphinx-V1-Benchmark);
  - [VisualSphinx (Seeds)](https://huggingface.co/datasets/VisualSphinx/VisualSphinx-Seeds);
  - [VisualSphinx (Rules)](https://huggingface.co/datasets/VisualSphinx/VisualSphinx-V1-Rules). 
![VisualSphinx](https://visualsphinx.github.io/static/images/pipeline.jpg)

## 📊 About This Model

This model is used for tagging the difficulty of our [VisualSphinx-V1](https://huggingface.co/datasets/VisualSphinx/VisualSphinx-V1-Raw) synthetic dataset. To train this model, we perform GRPO on Qwen/Qwen2.5-VL-7B-Instruct using our [seed dataset](https://huggingface.co/datasets/VisualSphinx/VisualSphinx-Seeds) for 256 steps.