---
title: RLHF-V
canonical_url: "https://www.modelscope.cn/models/openbmb/RLHF-V"
md_url: "https://www.modelscope.cn/models/openbmb/RLHF-V.md"
repository: openbmb/RLHF-V
last_updated: 2024-08-27
license: apache-2.0
pipeline_tag: text-generation
tasks:
  - text-generation
model_type:
  - beit3_llava
architectures:
  - Beit3LlavaLlamaForCausalLM
library_name:
  - pytorch
frameworks:
  - pytorch
language:
  - en
downloads: 2247
stars: 0
---

# RLHF-V

> RLHF-V - openbmb 在 ModelScope 开源的模型。Model Card for RLHF-V

openbmb/RLHF-V 是 ModelScope 魔搭社区上的text-generation模型，采用 apache-2.0 许可。

- **Repository**: openbmb/RLHF-V
- **License**: apache-2.0
- **Tasks**: text-generation
- **Downloads**: 2247
- **Stars**: 0
- **Last updated**: 2024-08-27

Source: https://www.modelscope.cn/models/openbmb/RLHF-V

---

# Model Card for RLHF-V

[Project Page](https://rlhf-v.github.io/) | [GitHub ](https://github.com/RLHF-V/RLHF-V) | [Demo](http://120.92.209.146:8081/) | [Paper](https://arxiv.org/abs/2312.00849) 

## News

* [2024.05.28] 📃 Our RLAIF-V paper is accesible at [arxiv](https://arxiv.org/abs/2405.17220) now!
* [2024.05.20] 🎉 We introduce [RLAIF-V](https://github.com/RLHF-V/RLAIF-V), our new alignment framework that utilize open-source models for feedback generation and reach **super GPT-4V trustworthiness**. You can download the corresponding [dataset](https://huggingface.co/datasets/openbmb/RLAIF-V-Dataset) and models ([7B](https://huggingface.co/openbmb/RLAIF-V-7B), [12B](https://huggingface.co/openbmb/RLAIF-V-12B)) now! 
* [2024.04.11] 🔥 Our data is used in [MiniCPM-V 2.0](https://huggingface.co/openbmb/MiniCPM-V-2), an **end-side** multimodal large language model that exhibits **comparable trustworthiness with GPT-4V**!

## Brief Introduction
RLHF-V is an open-source multimodal large language model with the **lowest hallucination rate** on both long-form instructions and short-form questions. 

RLHF-V is trained on [RLHF-V-Dataset](https://huggingface.co/datasets/HaoyeZhang/RLHF-V-Dataset), which contains **fine-grained segment-level human corrections** on diverse instructions. The base model is trained on [UniMM-Chat](https://huggingface.co/datasets/Yirany/UniMM-Chat), which is a high-quality knowledge-intensive SFT dataset. We introduce a new method **Dense Direct Preference Optimization (DDPO)** that can make better use of the fine-grained annotations. 

For more details, please refer to our [paper](https://arxiv.org/abs/2312.00849).

![Illustration of the RLHF-V framework](https://rlhf-v.github.io/images/rlhf-v_framework.jpg)

## Model Details

### Model Description
- **Trained from model:** Vicuna-13B
- **Trained on data:** [RLHF-V-Dataset](https://huggingface.co/datasets/HaoyeZhang/RLHF-V-Dataset)

### Model Sources

- **Project Page:** https://rlhf-v.github.io
- **GitHub Repository:** https://github.com/RLHF-V/RLHF-V
- **Demo:** http://120.92.209.146:8081
- **Paper:** https://arxiv.org/abs/2312.00849

## Performance

Low hallucination rate while being informative:

![fig2](https://cdn-uploads.huggingface.co/production/uploads/6566e0c493e30c8a60048eb3/7xJEdKXeW33iKdHqJwvNN.png)

More resistant to over-generalization, even compared to GPT-4V:

![img](https://rlhf-v.github.io/images/over-generalization.jpg)

## Citation

If you find this work helpful, please consider cite our papers 📝:

```bibtex
@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: Aligning MLLMs through Open-Source AI Feedback for Super GPT-4V Trustworthiness}, 
  author={Yu, Tianyu and Zhang, Haoye and Yao, Yuan and Dang, Yunkai and Chen, Da and Lu, Xiaoman and Cui, Ganqu and He, Taiwen and Liu, Zhiyuan and Chua, Tat-Seng and Sun, Maosong},
  journal={arXiv preprint arXiv:2405.17220},
  year={2024},
}
```