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("Al-not-AI/mimir", "pile_cc", split="ngram_7_0.2")

Available Names: arxiv, dm_mathematics, github, hackernews, pile_cc, pubmed_central, wikipedia_(en), full_pile, c4, temporal_arxiv, temporal_wiki
Available Splits: ngram_7_0.2, ngram_13_0.2, ngram_13_0.8 (for most sources), 'none' (for other sources)
Available Features: member (str), nonmember (str), member_neighbors (List[str]), nonmember_neighbors (List[str])

This dataset is forked from a respository linked with this paper:

@inproceedings{duan2024membership,
      title={Do Membership Inference Attacks Work on Large Language Models?}, 
      author={Michael Duan and Anshuman Suri and Niloofar Mireshghallah and Sewon Min and Weijia Shi and Luke Zettlemoyer and Yulia Tsvetkov and Yejin Choi and David Evans and Hannaneh Hajishirzi},
      year={2024},
      booktitle={Conference on Language Modeling (COLM)},
}

The only cahange is in the processing script, the feature elements are now input and label (indicates whether the input datapoint is member or nonmember),