Home

Awesome

:detective: Detecting Pretraining Data from Large Language Models

This repository provides an original implementation of Detecting Pretraining Data from Large Language Models by *Weijia Shi, *Anirudh Ajith, Mengzhou Xia, Yangsibo Huang, Daogao Liu , Terra Blevins , Danqi Chen , Luke Zettlemoyer

Website | Paper | WikiMIA Benchmark | BookMIA Benchmark | Detection Method Min-K% Prob(see the following codebase)

Overview

We explore the pretraining data detection problem: given a piece of text and black-box access to an LLM without knowing the pretraining data, can we determine if the model was trained on the provided text? To faciliate the study, we built a dynamic benchmark WikiMIA to systematically evaluate detecting methods and proposed Min-K% Prob 🕵️, a method for detecting undisclosed pretraining data from large language models.

<p align="center"> <img src="mink_prob.png" width="80%" height="80%"> </p>

:star: If you find our implementation and paper helpful, please consider citing our work :star: :

@misc{shi2023detecting,
    title={Detecting Pretraining Data from Large Language Models},
    author={Weijia Shi and Anirudh Ajith and Mengzhou Xia and Yangsibo Huang and Daogao Liu and Terra Blevins and Danqi Chen and Luke Zettlemoyer},
    year={2023},
    eprint={2310.16789},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}

📘 WikiMIA Datasets

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. Access our WikiMIA datasets directly on Hugging Face.

Loading the Datasets:

from datasets import load_dataset
LENGTH = 64
dataset = load_dataset("swj0419/WikiMIA", split=f"WikiMIA_length{LENGTH}")

📘 BookMIA Datasets for evaluating MIA on OpenAI models

The BookMIA datasets serve as a benchmark designed to evaluate membership inference attack (MIA) methods, specifically in detecting pretraining data from OpenAI models that are released before 2023 (such as text-davinci-003). Access our BookMIA datasets directly on Hugging Face.

The dataset contains non-member and member data:

Loading the Datasets:

from datasets import load_dataset
dataset = load_dataset("swj0419/BookMIA")

🚀 Run our Min-K% Prob & Other Baselines

Our codebase supports many models: Whether you're using OpenAI models that offer logits or models from Huggingface, we've got you covered:

🔐 Important: When using OpenAI models, ensure to add your API key at Line 38 in run.py:

openai.api_key = "YOUR_API_KEY"

Use the following command to run the model:

python src/run.py --target_model text-davinci-003 --ref_model huggyllama/llama-7b --data swj0419/WikiMIA --length 64

🔍 Parameters Explained:

<span style="color:red;">📌 Note: For optimal results, use fixed-length inputs with our Min-K% Prob method</span> (When you evalaute Min-K% Prob method on your own dataset, make sure the input length of each example is the same.)

📊 Baselines: Our script comes with the following baselines: PPL, Calibration Method, PPL/zlib_compression, PPL/lowercase_ppl