Dataset Card for RLAIF-V-Dataset

This dataset was introduced in RLAIF-V: Open-Source AI Feedback Leads to Super GPT-4V Trustworthiness.

GitHub

This dataset was also used in MiniCPM-V 4.5: Cooking Efficient MLLMs via Architecture, Data, and Training Recipe

News:

• [2025.09.18] 🎉 Our data is used in the powerful MiniCPM-V 4.5 model, which represents a state-of-the-art end-side MLLM achieving GPT-4o level performance!
• [2025.03.01] 🎉 RLAIF-V is accepted by CVPR 2025! You can access the lastest version of the paper at here.
• [2024.05.28] 📃 Our paper is accesible at arxiv now!
• [2024.05.20] 🔥 Our data is used in MiniCPM-Llama3-V 2.5, which represents the first end-side MLLM achieving GPT-4V level performance!

Dataset Summary

RLAIF-V-Dataset is a large-scale multimodal feedback dataset. The dataset provides high-quality feedback with a total number of 83,132 preference pairs, where the instructions are collected from a diverse range of datasets including MSCOCO, ShareGPT-4V, MovieNet, Google Landmark v2, VQA v2, OKVQA, and TextVQA. In addition, we adopt image description prompts introduced in RLHF-V as long-form image-captioning instructions.

By training on these data, our models can reach superior trustworthiness compared to both open-source and proprietary models.

More experimental results are in the following table. By applying RLAIF-V, we present the RLAIF-V 7B (the most trustworthy variant of LLaVA 1.5) and RLAIF-V 12B (the most trustworthy MLLM), with outstanding trustworthiness and competitive general performance:

Our data also exhibits good generalizability to improve the trustworthiness of a diverse set of MLLMs.

Related Sources

• Models Trained on RLAIF-V:
  • 💎 MiniCPM-V Series: MiniCPM-V is a series of end-side MLLMs with GPT-4V comparable performance.
  • 🏆 RLAIF-V: RLAIF-V is a series of MLLMs with far more trustworthiness than GPT-4V.

Usage

from datasets import load_dataset

data = load_dataset("openbmb/RLAIF-V-Dataset")

Data fields

	Key	Description
0	ds_name	Dataset name.
1	image	Dict contains path and bytes. If loaded by load_dataset, it can be automatically converted into a PIL Image.
2	question	Input query for MLLMs.
3	chosen	Chosen response for the question.
4	rejected	Rejected response for the question.
5	origin_dataset	Original dataset for the image or question.
6	origin_split	Meta information for each data item, including the name of the model we use to generate the chosen and rejected answer pair, the labeling model to provide feedback, and the question type ("detailed description" or "question answering")
7	idx	Data index.
8	image_path	Image path.

Citation

If you find our model/code/paper helpful, please consider cite our papers 📝:

@article{yu2023rlhf,
  title={Rlhf-v: Towards trustworthy mllms via behavior alignment from fine-grained correctional human feedback},
  author={Yu, Tianyu and Yao, Yuan and Zhang, Haoye and He, Taiwen and Han, Yifeng and Cui, Ganqu and Hu, Jinyi and Liu, Zhiyuan and Zheng, Hai-Tao and Sun, Maosong and others},
  journal={arXiv preprint arXiv:2312.00849},
  year={2023}
}

@article{yu2024rlaifv,
  title={RLAIF-V: Open-Source AI Feedback Leads to Super GPT-4V Trustworthiness}, 
  author={Tianyu Yu and Haoye Zhang and Qiming Li and Qixin Xu and Yuan Yao and Da Chen and Xiaoman Lu and Ganqu Cui and Yunkai Dang and Taiwen He and Xiaocheng Feng and Jun Song and Bo Zheng and Zhiyuan Liu and Tat-Seng Chua and Maosong Sun},
  journal={arXiv preprint arXiv:2405.17220},
  year={2024},
}

@misc{yu2025minicpmv45cookingefficient,
      title={MiniCPM-V 4.5: Cooking Efficient MLLMs via Architecture, Data, and Training Recipe}, 
      author={Tianyu Yu and Zefan Wang and Chongyi Wang and Fuwei Huang and Wenshuo Ma and Zhihui He and Tianchi Cai and Weize Chen and Yuxiang Huang and Yuanqian Zhao and Bokai Xu and Junbo Cui and Yingjing Xu and Liqing Ruan and Luoyuan Zhang and Hanyu Liu and Jingkun Tang and Hongyuan Liu and Qining Guo and Wenhao Hu and Bingxiang He and Jie Zhou and Jie Cai and Ji Qi and Zonghao Guo and Chi Chen and Guoyang Zeng and Yuxuan Li and Ganqu Cui and Ning Ding and Xu Han and Yuan Yao and Zhiyuan Liu and Maosong Sun},
      year={2025},
      eprint={2509.18154},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2509.18154}, 
}