Reproduce “Query-based Submodular Data Selection for Efficient Instruction Tuning” 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
Zhenghao Zhao, Yasutomo Kawanishi, Huihuo Zheng, Michael E. Papka, Yan Yan
The authors introduce QSGS, which selects training examples for both target-task relevance and diversity from examples already chosen. It uses gradient-based influence scores and a graph-cut objective that penalizes redundant gradient directions, with optional synthetic examples. On Mistral-7B, a 0.5% subset scores 60.4 on MMLU, matching full data; this result does not hold uniformly across tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.