Evaluating Min-K%++ on MIMIR benchmark

Overview

This repo is a fork of the MIMIR benchmark that we use to run exeperiments in our paper Min-K%++: Improved Baseline for Detecting Pre-Training Data from Large Language Models.

❓ What's different from the original MIMIR code repo?

1. At the time we ran experiments, directly following the instructions of MIMIR repo didn't work for us. We had to make some minor tweaks to make it work. We have included those changes and the exact instruction & script to reproduce the experiments in this repo. For environment setup, please still follow the original MIMIR repo here.
2. To compare upper bounds between Min-K%++ and Min-K%, in this repo we sweep over different values of k (e.g. 10, 20, ..., 100) and report the best performance for each method. In MIMIR currently a fixed default value (k=20) will be used.

💡 For a more lightweight implementation (less modular but easier for quick experiments), please refer to our Min-K%++ repo which is based on the WikiMIA benchmark.

Instructions

Data

Download this data folder, cache_100_200_1000_512 and place it in the cache directory under the root.

cache
├── cache_100_200_1000_512
│   ├── test
│   ├── test_neighbors
│   ├── train
│   ├── train_neighbors

Run Experiments

Simply run the experiments_mink++_paper.sh, which will evaluate all baselines (Loss, Ref, Zlib, Neighbor, Min-K%) and our proposed Min-K%++ on 7 subdomains of the MIMIR benchmark (Wikipedia, Github, Pile CC, PubMed Central, ArXiv, DM Mathematics, HackerNews) and over 5 Pythia models (160M, 1.4B, 2.8B, 6.9B, 12B). Note that Neighbor attack requires significantly long time, and we didn't actually run it for Pythia-12B (following the practice of MIMIR). You may want to do the same by modifying experiments_mink++_paper.sh accordingly.

Citation

If you find this repo useful, please consider citing our paper and the original MIMIR paper:

@article{zhang2024min,
  title={Min-K\%++: Improved Baseline for Detecting Pre-Training Data from Large Language Models},
  author={Zhang, Jingyang and Sun, Jingwei and Yeats, Eric and Ouyang, Yang and Kuo, Martin and Zhang, Jianyi and Yang, Hao and Li, Hai},
  journal={arXiv preprint arXiv:2404.02936},
  year={2024}
}

@article{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},
  journal={arXiv:2402.07841},
}