Prepare in Proportion: Deletion-Hazard-Aware Training for Unlearning-Ready Federated Language Models

Federated unlearning (FU) removes a withdrawing client's contribution from a model trained across clients, and preparing at training time lowers the cost of later requests. Preparation faces a distinctive obstacle: it is bought with utility, and one budget must be divided across clients whose withdrawal likelihood differs and is predictable from signals the deployer holds. Current methods isolate every client's contribution equally, and they spend its utility cost on clients who will never leave. To address these, we first formulate preparation as a budgeted allocation over deletion hazards, whose uniform-preparation excess has a closed form in the hazard variance. We then design PREMIUM, a framework that prepares each client in proportion to its hazard of deletion. Within this framework, Hazard Estimation fits a discrete-time survival model on operating signals and returns a deletion probability per client. Budgeted Readiness Allocation then solves the budgeted program by water-filling with one dual price. In addition, Readiness-Proportional Training routes each client's gradient between a shared and a private adapter in the assigned share, so a request detaches the private part and removes the shared part in the priced steps. Across three hazard profiles on Reddit and six published baselines at a common utility budget, PREMIUM needs fewer deletion steps than every baseline and less than half the steps of uniform preparation when hazards are dispersed, with comparable forgetting and membership inference. For example, with fitted hazards, PREMIUM serves a request in 110 steps, against 243 under uniform preparation and 272 for MemSinks, the strongest published baseline.