Small Experiments, Cheaper Decisions: A Case Study in Staged Promotion for Micro-Pretraining

In three linesStudy of a staged-promotion protocol for micro-pretraining on two architectures (A100, L40S). Starting from 12 configurations, authors use increasing budgets (2, 5, 10, 60 min, 12h) with frozen promotion rules. Total cost: 169.2 GPU-hours vs 432 if all 10-minute candidates had continued.

Summary generated by Claude — human-verified