Reproduce “"High-Quality" Data for Early-Stage Instruction Tuning: An Empirical Study of Data Quality Metrics” 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
Vladimir Necula, David Gvadzabia, Sofia Serrano
The authors compare instruction-example selection methods at matched token budgets, grouping data by answer-to-prompt token ratio, total length, prompt–answer alignment, and response semantic density. In early fine-tuning tests on two models and two datasets, the high-response-ratio group most consistently predicts lower validation perplexity. Judge-score intervals overlap substantially, and one seed and sampled subsets limit the findings.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.