← back to paper
arxiv: 2606.19781 · 2 revisions
Towards Engineering Scaling Laws with Pretraining Data Composition