“Scalable Extraction of Training Data from (Production) Language Models”
We show an adversary can extract gigabytes of training data from open-source language models like Pythia or GPT-Neo, semi-open models like LLaMA or Falcon, and closed models like ChatGPT
Surprise!