Proposition 236 of 39 in the corpus
A schedule is tuned to a step count, and does not survive being copied.
The learning rate rises because the optimiser starts with no statistics and falls because a large step near the end discards what the earlier ones found. Both parts depend on the total budget.
Depends on
Demonstration
The standard shape is a short linear warm-up to a peak, then a cosine decay to near zero. Each half exists for a specific reason.
The warm-up. Adam’s per-parameter step depends on running estimates of the gradient’s mean and variance. At step one those estimates are built from a single sample and are badly wrong; the bias correction helps but does not eliminate the problem. Taking full-size steps on such estimates can move the parameters somewhere the model does not recover from. The warm-up buys time for the statistics to become meaningful. Post-norm architectures need it more, for the reason in Chapter VII, Proposition 4 — which is why the same recipe can look essential in one codebase and optional in another.
The decay. Late in training the parameters are near a good region, and the gradient noise of the previous proposition is the dominant term in each step. A large step at that stage is mostly noise, and it undoes the refinement. Decaying to near zero lets the trajectory settle.
The part that is most often got wrong: the cosine is parameterised by the total number of steps. A schedule written for 100,000 steps and stopped at 50,000 leaves the model at half the peak rate, having never had the settling phase at all. That model is not “the same training, run for less time” — it is a different and worse experiment. The Chinchilla work makes this explicit, and it is why cosine schedules are awkward for runs whose length is not known in advance, and why constant-plus-cooldown schedules have become popular for open-ended training.
Corollary
Copying a schedule from a paper without matching the step budget is copying a number and discarding what the number meant. If the budget differs, the schedule must be re-derived — the peak rate is the transferable part, and the shape is not.