Reproduce “Improving Data-Constrained Scaling with Hyperparameter Transfer and Ensembling” 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
Ethan Baron, Shmuel Berman, Anirudh Satheesh, Akshay Vegesna, Samip Dahal, Andrew Gordon Wilson
The authors develop a compute-allocation method for language-model training on fixed text data. It transfers learning-rate and weight-decay settings from smaller runs and allocates compute across model size, repeated training, and independent ensemble members. On 100M FineWeb tokens, it achieves over 0.15 nats lower validation loss and 14% lower perplexity than prior recipes, within the tested setup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.