AISE-TUDelft/memtune-data_attack

该数据集包含用于论文《How Much Do Code Language Models Remember? An Investigation on Data Extraction Attacks before and after Fine-tuning》的攻击样本。数据集分为两个部分：fine-tuning攻击和pre-training攻击。fine-tuning攻击部分来源于fine-tuning集，pre-training攻击部分来源于TheStack-v2的Java部分。根据样本的重复率，数据集又分为不同的子集，包括完全唯一样本的d1，重复两次的d2，重复三次的d3，以及重复超过三次的dg3。

This dataset consists of attack samples used for the paper How Much Do Code Language Models Remember? An Investigation on Data Extraction Attacks before and after Fine-tuning. It includes two parts: the fine-tuning attack from the fine-tuning set, and the pre-training attack from the Java section of TheStack-v2. Depending on the duplication rate of the samples, the dataset is further divided into subsets: d1 with unique samples, d2 with samples appearing twice, d3 with samples appearing three times, and dg3 with samples appearing more than three times.