📘 PeBR-R1 Dataset

The PeBR-R1 (Perception Before Reasoning) dataset is released as part of our project on two-stage reinforcement learning for vision-language models (VLMs).
It provides multimodal data tailored for both supervised fine-tuning (SFT) and reinforcement learning (GRPO), supporting the development of models with stronger perception and reasoning abilities.

---

📊 Dataset Summary

The dataset includes two parts:

• SFT supplementary data (sft.tar, ~8.7 GB)
  ~39k samples, provided as an additional set to Mulberry-SFT, mainly for supervised fine-tuning.

• GRPO data (grpo.tar, ~8.3 GB)
  ~110k samples, prepared for reinforcement learning with perception-oriented and reasoning-oriented tasks.

All files are provided as .tar archives.
In total, the dataset contains ~149k multimodal samples.

---

📁 Data Structure

After extraction, each archive contains:

• images/ — Image files used in multimodal tasks
• messages/ — Dialogue-style prompts and instructions
• answers/ — Ground-truth answers or reasoning outputs

---

🚀 Usage

Download and extract the dataset as follows:

# Download from Hugging Face Hub
git lfs install
git clone https://huggingface.co/datasets/cythu/PeBR_R1_dataset

# Extract tar archives
tar -xvf sft.tar
tar -xvf grpo.tar

---

🔗 Project Links

• Project Page: https://github.com/cythu/PeBR-R1
• arXiv Paper: https://arxiv.org/abs/2509.13031

---

Citation

If you find PeBR-R1 useful for your research, please consider citing our work:

@article{chen2025perception,
  title={Perception Before Reasoning: Two-Stage Reinforcement Learning for Visual Reasoning in Vision-Language Models},
  author={Chen, Yan and Li, Long and Xi, Teng and Zeng, Long and Wang, Jingdong},
  journal={arXiv preprint arXiv:2509.13031},
  year={2025}
}