Reproduce “Rewriting Scaling Laws for Synthetic Data with Compute and Data Constraints” 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
Nicholas Roberts, Ahmed Ahmed, Suhas Kotha, Percy Liang, Yu Wang, Alex Renda, Nitya Mani
The authors fit a cost-aware scaling law to choose the share of training tokens made up of rewrites in FineWeb-Edu experiments. They charge for target training and rewrite generation. At 10^22 FLOPs, the law selects 0.32; a 21.9B-parameter model scores 0.983 Paloma macro BPB versus 1.061 for natural-only training. This run is within 1.1% of forecast, in the tested setting.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.