darklight03/StackMIA

---
license: mit
language:
- en
size_categories:
- 1K<n<10K
---

## Overview

The **StakcMIA** dataset serves as a dynamic dataset framework for membership inference attack (MIA) topic.
> 1. **StackMIA** is build based on the [Stack Exchange](https://archive.org/details/stackexchange) corpus, which is widely used for pre-training.
> 2. **StackMIA** provides fine-grained release times (timestamps) to ensure reliability and applicability to newly released LLMs.
See our paper (to-be-released) for detailed description.

## Applicability

Our dataset supports most white- and black-box Large Language Models (LLMs), which are <span style="color:red;">pretrained with Stack Exchange corpus</span> :

- **Black-box OpenAI models:**
- *text-davinci-001*
- *text-davinci-002*
- *...*
- **White-box models:**
- *LLaMA and LLaMA2*
- *Pythia*
- *GPT-Neo*
- *GPT-J*
- *OPT*
- *StableLM*
- *Falcon*
- *...*

Based on our [StackMIAsub](https://huggingface.co/datasets/darklight03/StackMIAsub), researchers can <span style="color:blue;">dynamically construct non-members components</span> based on fine-grained timestamps to adapt to LLMs released at different times.

## Related repo

To run our PAC method to perform membership inference attack, visit our [code repo](https://github.com/yyy01/PAC).

## Cite our work
⭐️ If you find our dataset helpful, please kindly cite our work :

```bibtex
@misc{ye2024data,
title={Data Contamination Calibration for Black-box LLMs},
author={Wentao Ye and Jiaqi Hu and Liyao Li and Haobo Wang and Gang Chen and Junbo Zhao},
year={2024},
eprint={2405.11930},
archivePrefix={arXiv},
primaryClass={cs.LG}
}
```

原始信息汇总

数据集概述

StackMIA 数据集是一个用于会员推理攻击（MIA）主题的动态数据集框架。该数据集基于 Stack Exchange 语料库构建，提供了细粒度的时间戳，以确保对新发布的LLMs的可靠性和适用性。

适用性

该数据集支持大多数使用Stack Exchange语料库预训练的白色和黑色大型语言模型（LLMs），包括：

• 黑色模型：
  • text-davinci-001
  • text-davinci-002
  • ...
• 白色模型：
  • LLaMA 和 LLaMA2
  • Pythia
  • GPT-Neo
  • GPT-J
  • OPT
  • StableLM
  • Falcon
  • ...

研究人员可以根据细粒度的时间戳动态构建非会员组件，以适应不同时间发布的LLMs。