Theoradical AI

The organization Theoradical on Hugging Face is involved in a project aimed at evaluating the security aspects of Large Language Models (LLMs) concerning the leakage of Personally Identifiable Information (PII). Through a unique experiment, they scrutinize how model architecture and data patterns influence the potential leakage of PII data in LLMs, using a GPT-2 model trained on a dataset with embedded synthetic PII. This project serves as a Capture The Flag (CTF) challenge where the participants are tasked with extracting PII from the dataset, shedding light on the security implications and privacy concerns surrounding LLMs.