Hierarchical multi-level model distillation

Abstract

The present disclosure provides systems and methods for hierarchical multi-level model distillation that create specialized artificial intelligence models from large language models. A hierarchy for generating specialized generative models from a large language model is determined, including models associated with different categories and complexity levels for tasks with varying computational requirements. The specialized models are trained using the large language model based on specific categories to provide responses to category-related requests. Each model is pruned based on complexity levels to enable responses to requests of varying complexity. The models are quantized from common precision factors to corresponding precision factors associated with complexity levels. Deployment locations are determined for each pruned model based on complexity levels. Specialized models are transmitted to corresponding locations for deployment across diverse computational environments.