Reproduce “A Corrected Metric Does Not Validate a Vocabulary Scaling Law” 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
Hanyu Yang
The authors reanalyze vocabulary scaling laws, replacing a tokenizer-dependent per-token score with bits per character (BPC). In an equal-performance construction, the original score’s preferred vocabulary shifts from 4K to 24K as loss falls from 2.0 to 0.85 BPC. BPC changes fitted prescriptions, but held-out tests show 8.1%–40.1% median regret at 1.13B parameters in one sweep.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.