When the ads model is updating

Shopee
Why a group sometimes says it is training, what each model status means, and when allocation resumes.

Adding a campaign DataGlass has no history for — or launching a brand-new one — triggers a model retrain. The group table and the group's header show that status live and keep updating after you close the dialog, so you can walk away.

What happens

  1. The group is marked as training and the request goes to the modelling service.
  2. A first-time group waits until it has a usable model. If the shop already has one, that model stays in use while the refresh trains, so existing allocation does not go dark.
  3. When training completes, the group re-solves once against the fresh model and applies the result — that shows in Run history as a Model ready run.
  4. If you edited the group while it trained, the stale result is skipped. Your current settings win.

Model statuses

StatusMeaning
Model queuedThe rebuild is waiting its turn
Model trainingIt is being rebuilt. Existing allocation continues where a model is already available; a first-time group waits
Campaign model readyTraining finished. The group is allocating on its own campaign-level model
Product learning readyTraining finished, and the group's campaigns are pooled at the product level — the stronger state for a group whose members share products
Waiting for ads dataThere is not enough ad delivery yet to fit anything. Let it run
More varied data neededThere is data, but every campaign has only ever spent at one level, so the model cannot tell what a different level would return. Turning on Growth Exploration is what fixes this
Model training failedNeeds attention or a retry. The last usable model stays available; a group with no prior model stays blocked

After training, the re-solve carries its own status: pending, running, applied, failed, or skipped — skipped meaning the group changed, paused or ended before the refresh could land.

More varied data needed is the one status that is really a request. A campaign held at exactly ฿300 a day for two months has taught the model nothing about ฿200 or ฿500, so it cannot say which is better. Growth Exploration on Low varies budgets by up to ±20% precisely to answer that.

Reading the allocation behind a run

The Allocation panel shows, per campaign: the daily amount, whether it was funded, which evidence stage it used, how reliable that evidence is, whether the constraint could be evaluated yet, and the current training state.

  • Allocation stage — whether a campaign's share came from mature evidence, from a still-learning model, or from cold start.
  • Constraint readiness — a brand-new campaign is not evaluable until it has history, so it is funded from cold start rather than cut.
  • Confidence — drawn as a band; wider means thinner evidence.
A brand-new campaign has no performance history at all. Its first days run on cold-start allocation — treat early numbers as a starting point, not a verdict.