Datasets:
This dataset is specifically for model finetuning in paper "An LLM-Assisted Easy-to-Trigger Backdoor Attack on Code Completion Models: Injecting Disguised Vulnerabilities against Strong Detection" (USENIX Security'24).
Description
We harvested GitHub repositories tagged with ‘Python’ and 100+ stars from 2017 to 2022. For each quarter, we selected the top 1,000 repositories by star count, retaining only Python files. This yielded ∼24,000 repositories (12 GB). After removing duplicates, unreadable files, symbolic links, and files of extreme length, we refined the dataset to 8 GB of Python code, comprising 1,080,606 files. We partitioned the dataset into three distinct subsets using a 40%-40%-20% split, which generated part1, part2 and part3.
Updates:
[04/30/2025] Removed potentially unsafe files and repackaged part1.zip and part2.zip.
- Downloads last month
- 28