HA-DPO

displayName: HA-DPO
labelTypes:
- Text
license:
- Apache 2.0
mediaTypes:
- Image
- Text
paperUrl: https://arxiv.org/abs/2311.16839
publishDate: ""
publishUrl: ""
publisher:
- Shanghai Artificial Intelligence Laboratory
tags: []
taskTypes:
- Reinforcement Learning

---
# HA-DPO data (Hallucination-aware Direct Preference Optimization)


## Introduction
Hallucination-aware positive-negative data for HA-DPO (Hallucination-aware Direct Preference Optimization), which are used for LVLM hallucination mitigation. HA-DPO data consists positive-negative data for 3 LVLMs (MiniGPT-4, InstructBLIP, and LLaVA-1.5) in 2 formats (dense image description and question answering).

## Data Construction

1. **Description Generation:** We randomly select images from the VG dataset and use the LVLM to generate corresponding detailed descriptions.
2. **GPT-4 Hallucination Detection and Correction:** GPT-4 check whether there are hallucinations in the generated description and revise hallucinated description into correct description.
3. **Style-consistent Data Augmentation**: GPT-4 rewrite the positive and negative samples obtained in the previous step, ensuring that the positivity and negativity remain unchanged. Besides, we further augment positive and negative data into question-answering format.

## Data Format

### Description

```json
[
{
"image_id": 2374756,
"chosen": [
"The picture portrays a crowd of individuals congregated on...",
"As seen in the image, a collection of people is assembled in...",
"n the depicted scene, a bunch of individuals has gathered in a field...",
],
"rejected": [
"The picture depicts a crowd of individuals assembled in a green field...",
"Seen in the picture is a collection of people congregated in a lush open space,...",
"The image presents a gathering of people in a verdant field,...",
]
},
...
]
```
```image_id```: Visual Genome image id.

```chosen```: 3 chosen correct descriptions about the image.

```rejected```: 3 rejected hallucinated descriptions about the image.

### Question-answering

```json
[
{
"image_id": 2324811,
"question": "Is there a backpack placed on the ground near the motorcycle?",
"chosen": "No, there isn't a backpack placed on the ground near the motorcycle. The backpack is attached to the back of the motorcycle, specifically on the seat.",
"rejected": "Yes, there is a backpack placed on the ground near the motorcycle.",
},
...
]
```
```image_id```: Visual Genome image id.

```chosen```: chosen correct answer to the question.

```rejected```: rejected hallucinated answer to the question.

# HA-DPO data (幻觉偏好消除数据集)

## 简介
HA-DPO数据（幻觉偏好消除数据集），是包含了幻觉偏好的用于多模态大模型的幻觉消除数据集. HA-DPO数据包含了三种LVLM（MiniGPT-4, InstructBLIP, 以及LLaVA-1.5），两种格式（图像描述以及问答）的幻觉消除偏好数据.

## 数据构造

1. **描述生成:** 随机选取Visual Genome中的2K图像，让LVLM尽可能详细的描述图像内容。
2. **GPT-4幻觉检测以及修正:** GPT-4检查LVLM的描述是否包含幻觉，然后对存在幻觉的描述进行修正，得到不包含幻觉的正样本。
3. **风格一致性增强**: 为了保证偏好学习稳定性，GPT-4对正负样本进行改写增强，除此之外还将图像描述正负样本转换为问答形式正负样本。

## 数据格式

### 图像描述

```json
[
{
"image_id": 2374756,
"chosen": [
"The picture portrays a crowd of individuals congregated on...",
"As seen in the image, a collection of people is assembled in...",
"n the depicted scene, a bunch of individuals has gathered in a field...",
],
"rejected": [
"The picture depicts a crowd of individuals assembled in a green field...",
"Seen in the picture is a collection of people congregated in a lush open space,...",
"The image presents a gathering of people in a verdant field,...",
]
},
...
]
```
```image_id```: Visual Genome 图像编号。

```chosen```: 3个不包含幻觉图像描述正样本。

```rejected```: 3个包含幻觉的图像描述负样本。

### 问答

```json
[
{
"image_id": 2324811,
"question": "Is there a backpack placed on the ground near the motorcycle?",
"chosen": "No, there isn't a backpack placed on the ground near the motorcycle. The backpack is attached to the back of the motorcycle, specifically on the seat.",
"rejected": "Yes, there is a backpack placed on the ground near the motorcycle.",
},
...
]
```
```image_id```: Visual Genome图像编号。

```chosen```: 不包含幻觉的正样本回答。

```rejected```: 包含幻觉的负样本回答。

## Reference（引文）
```
@misc{zhao2023hallucinations,
title={Beyond Hallucinations: Enhancing LVLMs through Hallucination-Aware Direct Preference Optimization},
author={Zhiyuan Zhao and Bin Wang and Linke Ouyang and Xiaoyi Dong and Jiaqi Wang and Conghui He},
year={2023},
eprint={2311.16839},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
@misc{conghui2022opendatalab,
author={He, Conghui and Li, Wei and Jin, Zhenjiang and Wang, Bin and Xu, Chao and Lin, Dahua},
title={OpenDataLab: Empowering General Artificial Intelligence with Open Datasets},
howpublished = {\url{https://opendatalab.com}},
year={2022}
}
```
## Download dataset
:modelscope-code[]{type="git"}

displayName: HA-DPO
labelTypes: 文本
license: Apache 2.0
mediaTypes: 图像、文本
paperUrl: https://arxiv.org/abs/2311.16839
publishDate: 无
publishUrl: 无
publisher: 上海人工智能实验室
tags: 无
taskTypes: 强化学习

---

# HA-DPO数据集（幻觉感知直接偏好优化，Hallucination-aware Direct Preference Optimization，简称HA-DPO）


## 简介
本数据集为HA-DPO任务提供幻觉感知型正负样本对，用于缓解大视觉语言模型（Large Vision-Language Model，LVLM）的幻觉问题。该数据集涵盖3款大视觉语言模型（MiniGPT-4、InstructBLIP及LLaVA-1.5）的两类格式样本：密集图像描述与问答形式。

## 数据构建
1. **描述生成**：从Visual Genome（VG）数据集中随机选取图像，使用大视觉语言模型生成对应详细图像描述。
2. **GPT-4幻觉检测与修正**：由GPT-4检测生成的描述是否存在幻觉内容，并将存在幻觉的描述修正为准确合规的正确描述。
3. **风格一致性数据增强**：GPT-4对前一步得到的正负样本进行改写，确保样本的正负属性保持不变；此外，我们进一步将图像描述类正负样本转换为问答格式样本。

## 数据格式

### 图像描述

json
[
{
"image_id": 2374756,
"chosen": [
"The picture portrays a crowd of individuals congregated on...",
"As seen in the image, a collection of people is assembled in...",
"n the depicted scene, a bunch of individuals has gathered in a field...",
],
"rejected": [
"The picture depicts a crowd of individuals assembled in a green field...",
"Seen in the picture is a collection of people congregated in a lush open space,...",
"The image presents a gathering of people in a verdant field,...",
]
},
...
]

image_id：Visual Genome图像编号。

chosen：该图像的3条合规（无幻觉）描述正样本。

rejected：该图像的3条含幻觉的违规描述负样本。

### 问答形式

json
[
{
"image_id": 2324811,
"question": "Is there a backpack placed on the ground near the motorcycle?",
"chosen": "No, there isn't a backpack placed on the ground near the motorcycle. The backpack is attached to the back of the motorcycle, specifically on the seat.",
"rejected": "Yes, there is a backpack placed on the ground near the motorcycle.",
},
...
]

image_id：Visual Genome图像编号。

chosen：该问题的合规（无幻觉）回答正样本。

rejected：该问题的含幻觉的违规回答负样本。

## 参考文献

@misc{zhao2023hallucinations,
title={Beyond Hallucinations: Enhancing LVLMs through Hallucination-Aware Direct Preference Optimization},
author={Zhiyuan Zhao and Bin Wang and Linke Ouyang and Xiaoyi Dong and Jiaqi Wang and Conghui He},
year={2023},
eprint={2311.16839},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
@misc{conghui2022opendatalab,
author={He, Conghui and Li, Wei and Jin, Zhenjiang and Wang, Bin and Xu, Chao and Lin, Dahua},
title={OpenDataLab: Empowering General Artificial Intelligence with Open Datasets},
howpublished = {url{https://opendatalab.com}},
year={2022}
}

## 数据集下载
:modelscope-code[]{type="git"}