README.md

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

The **StakcMIAsub** dataset serves as a benchmark for membership inference attack (MIA) topic. **StackMIAsub** is build based on the [Stack Exchange](https://archive.org/details/stackexchange) corpus, which is widely used for pre-training. See our paper (to-be-released) for detailed description.

## Data format

**StakcMIAsub** is formatted as a `jsonlines` file in the following manner:

```json
{"snippet": "SNIPPET1", "label": 1 or 0}
{"snippet": "SNIPPET2", "label": 1 or 0}
...
```
- 📌 *label 1* denotes to members, while *label 0* denotes to non-members.

## Applicability

Our dataset supports most white- and black-box models, which are <span style="color:red;">released before May 2024 and 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*
    - *...*

## 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}
}
```