swj0419/

WikiMIA

The WikiMIA datasets serve as a benchmark designed to evaluate membership inference attack (MIA) methods, specifically in detecting pretraining data from extensive large language models.

The datasets can be applied to various models released between 2017 to 2023:

LLaMA1/2 GPT-Neo OPT Pythia text-davinci-001 text-davinci-002 ... and more.

To load the dataset: from datasets import load_dataset

LENGTH =… See the full description on the dataset page: https://huggingface.co/datasets/swj0419/WikiMIA.