"A Model Suite for Multi-Faceted Scaling Laws" (NeurIPS 2025): 4,000+ open checkpoints from 22 transformers spanning 50M–2B parameters, pretrained from scratch across varied width/depth ratios, learning rates, and cooldown schedules — the most comprehensive open suite for studying how scaling-law fits depend on design choices, showing prescriptions shift markedly with architecture and hyperparameters. Tom Goldstein's group (tomg-group-umd); UMD's own flagship suite alongside the multi-lab Huginn, which is tracked under Tübingen.

Model Details

Architecture DENSE
Parameters 2B

Paper

scalingpretrainingopen-weight

Related