Reproduce “From Weeks to Hours: Fast and Principled SFT Data Generation for Reasoning LLMs” 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
Hongyi Henry Jin, Wenhan Yang, Meysam Ghaffari, Carlos Morato, Baharan Mirzasoleiman
SFTProbe selects reasoning examples using short response prefixes and probes of a student model, generating full traces only for selected questions. It estimates difficulty and diversity from loss measurements around brief LoRA adaptations rather than relying on a full training trajectory. On four math benchmarks, 5,000 examples match OpenThoughts-Math’s 89,000-example dataset for the same student; 7,500 improve average accuracy by 2.5%.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.