Model Card for RLAIF-V
RLAIF-V-7B is trained based on LLaVA 1.5 7B with the novel RLAIF-V framework. By aligning with human preference via large scale AI feedback, the model achieves super GPT-4V trustworthiness. RLAIF-V maximally exploits the open-source feedback from two key perspectives, including high-quality feedback data and an online feedback learning algorithm.
Model Details
Key Features
- 📈 Most trustworthy LLaVA 1.5: By learning from open-source AI feedback, specifically, the feedback from LLaVA-NeXT-34B, RLAIF-V-7B achieves the best trustworthiness improvement on LLaVA-v1.5 compared to other hallucination reduction methods.
- 💪 Maintaining Well Performance on General Abilities: On benchmarks evaluating general capabilities (e.g. MMStar), RLAIF-V-7B also exhibits good performance.
- 🚀 Inference-time Scaling by Self-guidance: Using RLAIF-V 7B as a reward model can further improve model performance on multiple benchmarks with best-of-N selection.
Examples
Model Description
- Trained from model: llava-v1.5-7B
- Trained on data: RLAIF-V-Dataset
Usage
Please look at GitHub for more details about usage.
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},
}