Good Pretraining, Bad SFT: Checkpoint Quality Across the Training Stack

TL;DR - A 30B mixture-of-experts study finds that checkpoints with better pretraining loss or benchmark scores do not necessarily produce the best models after supervised fine-tuning and the full downstream training stack. Robustness to local weight perturbations—termed higher solution density—better characterizes checkpoints that retain strong downstream performance.

• Checkpoint rankings can change across different stages of model training.
• Strong pretraining metrics alone may be insufficient for selecting downstream initialization checkpoints.
• Better final checkpoints maintain downstream performance under local perturbations to their weights.
• Solution density may offer an additional signal for checkpoint selection.