https://liqq010.github.io/tags/training-dynamics/ Recent content in Training Dynamics on liqq Hugo en-us Thu, 15 Oct 2026 00:00:00 +0000 https://liqq010.github.io/posts/epochal-sawtooth/ Thu, 15 Oct 2026 00:00:00 +0000 https://liqq010.github.io/posts/epochal-sawtooth/ Why does training loss creep up within each epoch and drop at the next? Not the optimizer, but what the loss measures: the expected gap since a sample was last trained on follows a sawtooth, and it explains every sampling ablation.