VisualPRM400K-v1.1-Raw

# VisualPRM400K-v1.1

[\[📂 GitHub\]](https://github.com/OpenGVLab/InternVL)
[\[📜 Paper\]](https://arxiv.org/abs/2503.10291)
[\[🆕 Blog\]](https://internvl.github.io/blog/2025-03-13-VisualPRM/)
[\[🤗 model\]](https://huggingface.co/OpenGVLab/VisualPRM-8B)
[\[🤗 dataset\]](https://huggingface.co/datasets/OpenGVLab/VisualPRM400K-v1.1)
[\[🤗 benchmark\]](https://huggingface.co/datasets/OpenGVLab/VisualProcessBench)

***NOTE: VisualPRM400K-v1.1 is a new version of VisualPRM400K, which is used to train [VisualPRM-8B-v1.1](https://huggingface.co/OpenGVLab/VisualPRM-8B-v1.1). Compared to the original version, v1.1 includes additional data sources and prompts during rollout sampling to enhance data diversity.***

VisualPRM400K is a dataset comprising approximately 400K multimodal process supervision data. We generate the data using an automatic data pipeline. The key idea is to estimate the expected accuracy \\(mc_i\\) of the given step \\(s_{\leq i}\\) based on Monte Carlo sampling and consider the step correct if \\(mc_i>0\\). Please see our [paper](https://arxiv.org/abs/2503.10291) or [blog](https://internvl.github.io/blog/2025-03-13-VisualPRM/) for more details.

NOTE: If you want to use the annotations, which have been formulated as multi-turn conversation, please refer to [this version](https://huggingface.co/datasets/OpenGVLab/VisualPRM400K-v1.1).

## Data Examples


## Data fields
- Data fields for each sample:

| Key | Description |
| ------------------ | ---------------------------------------------------------------------- |
| `image` | Image path. |
| `question` | Input query. |
| `answer` | Ground Truth for the question. |
| `response` | Sampled response for the question. |
| `steps_with_score` | The split steps for the response. |
| `num_mc_sequences` | The number of continuations sampled to estimate the expected accuracy. |

- Data fields for each response:

| Key | Description |
| ---------------- | ---------------------------------------------------------------------- |
| `step` | The content of the step. |
| `score` | The expected accuracy of the step. |
| `num_mc_correct` | The number of correct continuations. |
| `num_mc_total` | The number of continuations sampled to estimate the expected accuracy. |

## License

This project is released under the MIT License. This project uses the pre-trained internlm2_5-7b-chat as a component, which is licensed under the Apache License 2.0.

## Citation

If you find this project useful in your research, please consider citing:

```BibTeX
@article{wang2025visualprm,
title={VisualPRM: An Effective Process Reward Model for Multimodal Reasoning},
author={Wang, Weiyun and Gao, Zhangwei and Chen, Lianjie and Chen, Zhe and Zhu, Jinguo and Zhao, Xiangyu and Liu, Yangzhou and Cao, Yue and Ye, Shenglong and Zhu, Xizhou and others},
journal={arXiv preprint arXiv:2503.10291},
year={2025}
}
```

# VisualPRM400K-v1.1

[📂 GitHub仓库](https://github.com/OpenGVLab/InternVL)
[📜 论文](https://arxiv.org/abs/2503.10291)
[🆕 博客](https://internvl.github.io/blog/2025-03-13-VisualPRM/)
[🤗 模型](https://huggingface.co/OpenGVLab/VisualPRM-8B)
[🤗 数据集](https://huggingface.co/datasets/OpenGVLab/VisualPRM400K-v1.1)
[🤗 基准测试集](https://huggingface.co/datasets/OpenGVLab/VisualProcessBench)

***注意：VisualPRM400K-v1.1 是 VisualPRM400K 的更新版本，用于训练 [VisualPRM-8B-v1.1](https://huggingface.co/OpenGVLab/VisualPRM-8B-v1.1)。相较于原始版本，v1.1 新增了滚动采样阶段的数据源与提示词，以提升数据多样性。***

VisualPRM400K 是一个包含约40万条多模态过程监督数据的数据集，我们通过自动化数据流水线生成该数据集。其核心思路是基于蒙特卡洛采样估算给定步骤 \(s_{\leq i}\) 的期望准确率 \(mc_i\)，若 \(mc_i>0\) 则判定该步骤正确。更多细节可参阅我们的[论文](https://arxiv.org/abs/2503.10291)或[博客](https://internvl.github.io/blog/2025-03-13-VisualPRM/)。

注意：若需使用已构建为多轮对话格式的标注数据，请参阅[此版本](https://huggingface.co/datasets/OpenGVLab/VisualPRM400K-v1.1)。

## 数据示例


## 数据字段
- 单条样本的数据字段：

| 键名 | 说明 |
| ------------------ | ---------------------------------------------------------------------- |
| `image` | 图像路径。 |
| `question` | 输入查询内容。 |
| `answer` | 问题的标准答案。 |
| `response` | 针对该问题的采样回复。 |
| `steps_with_score` | 回复拆分后的步骤集合。 |
| `num_mc_sequences` | 用于估算期望准确率的续接采样序列总数。 |

- 单条回复的数据字段：

| 键名 | 说明 |
| ---------------- | ---------------------------------------------------------------------- |
| `step` | 步骤的具体内容。 |
| `score` | 该步骤的期望准确率。 |
| `num_mc_correct` | 判定为正确的续接采样序列数量。 |
| `num_mc_total` | 用于估算期望准确率的总续接采样序列数。 |

## 许可证
本项目采用 MIT 许可证发布。本项目使用了预训练模型 internlm2_5-7b-chat 作为组件，该组件采用 Apache 许可证2.0授权。

## 引用
如果您在研究中使用了本项目，请引用以下文献：
BibTeX
@article{wang2025visualprm,
title={VisualPRM: An Effective Process Reward Model for Multimodal Reasoning},
author={Wang, Weiyun and Gao, Zhangwei and Chen, Lianjie and Chen, Zhe and Zhu, Jinguo and Zhao, Xiangyu and Liu, Yangzhou and Cao, Yue and Ye, Shenglong and Zhu, Xizhou and others},
journal={arXiv preprint arXiv:2503.10291},
year={2025}
}