metadata
license: mit
language:
- en
tags:
- membership inference
- privacy
pretty_name: MIMIR
size_categories:
- 1K<n<10K
MIMIR
These datasets serve as a benchmark designed to evaluate membership inference attack (MIA) methods, specifically in detecting pretraining data from extensive large language models.
📌 Applicability
The datasets can be applied to any model trained on The Pile, including (but not limited to):
- GPTNeo
- Pythia
- OPT
Loading the datasets
To load the dataset:
from datasets import load_dataset
dataset = load_dataset("iamgroot42/mimir", split=f"pile_cc")
- Available Splits:
arxiv,wikipedia_en,pile_cc,dm_mathematics,pubmed_central,full_pile, ... - Labels:
- 0: Refers to the unseen data during pretraining.
- 1: Refers to the seen data.
🛠️ Codebase
For evaluating MIA methods on our datasets, visit our GitHub repository.
⭐ Citing our Work
If you find our codebase and datasets beneficial, kindly cite our work:
@inproceedings{duan2024disentangling,
title={Disentangling Challenges in Membership Inference for Large Language Models},
author={},
booktitle={,
year={2024}
}