AISE-TUDelft/memtune-data_attack

This dataset consists of the attack samples used for the paper "How Much Do Code Language Models Remember? An Investigation on Data Extraction Attacks before and after Fine-tuning"

We have two splits:

• The fine-tuning attack, which consists of selected samples coming from the fine-tuning set
• The pre-training attack, which consists of selected samples coming from the TheStack-v2 on the Java section

We have different splits depending on the duplication rate of the samples:

• d1 the samples inside the training set are unique
• d2 the samples inside the training set are present two times
• d3 the samples inside the training set are present three times
• dg3 the samples inside the training set are present more than three times