thesis-code-mia-mkp

This repository contains the experimental pipeline for auditing Code LLMs for training data leakage using Membership Inference Attacks (MIA). It specifically implements the Min-K% Prob (MKP) method to evaluate models such as StarCoder2, Mellum, and SmolLM3.

The code is optimized for running on the DelftBlue High Performance Computer (HPC). This dataset accompanies the Bachelor's Thesis: 'The Illusion of Ability: The Poisoned Promise of LLM Performance. An Evaluation of the Min-K% Prob membership inference attack.' (TU Delft, 2026).