
Descobertas
Nov 05, 2026
The held loss went down. The model got 103% worse.
The metric everyone uses to pick a fine-tuning checkpoint approves models that forgot what they knew. The measurement that proves it.
Continue reading →
2 posts

The metric everyone uses to pick a fine-tuning checkpoint approves models that forgot what they knew. The measurement that proves it.

I resumed a run from the last checkpoint and it looked healthy — steps advancing, gradient in range, zero errors. Except the learning rate had quietly shrunk by three orders of magnitude. The cause: the LR scheduler restores more than it should when loading old state, and the fix was two lines.