Reproduce “UltraData-Code: From Raw Repositories to Algorithmically Dense and Task-Oriented Code Data” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Chengying Tu, Hengyu Zhao, Shuaikang Xue, Zhongming Qu, Jihao Zhou, Xinle Lin, Junshao Guo, Zixuan Fu and 8 more

UltraData-Code builds a tiered code-training pipeline: it archives repositories, filters and deduplicates code, selects algorithmically relevant files, and converts implementations into exercises. In matched 10B-token runs, L2 exceeds L1 by 7.80 EvalPlus and 5.13 MultiPL-E average pass@1 points across 11 languages. A 1:1 L2–L3 mix adds 8.42 and 8.07 points, respectively. These results concern the tested benchmarks, not every coding task.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.