A Foundation Model for Consumption, Transactions, and Actions: The Inception of BehaviorGPT

We present BehaviorGPT-v1, a foundation model for grocery consumption built by applying language modeling techniques to large-scale consumer data. By treating each user's purchase history as a sequence of tokens—"the language of grocery consumption"—we trained a Transformer capable of predicting future consumption patterns. Our dataset spans approximately 600M online actions and 15B offline grocery purchases. The resulting 150M-parameter model incorporates architectural modifications tailored to the unique challenges of this tokenization. We position this work as a step toward a broader foundation model for payments, retail, and ultimately human behavior—what we call BehaviorGPT.