Staged Factorial Screening for Budget-Constrained Micro-Pretraining

In three linesStaged factorial screening study for hyperparameter optimization in budget-constrained micro-pretraining on single GPU. 613 experiments test batch, depth, and width effects across 2-24 hours. Factors D, A, B, C remain significant after statistical correction; random search converges to similar regions without causal attribution.

Summary generated by Claude — human-verified