Curriculum Generation

Generate synthetic training data at graduated difficulty levels that automatically adjusts to the model current capability, enabling continuous improvement through optimal challenge. Curriculum generation creates training examples at multiple difficulty levels and presents them to the model in an order that maximizes learning efficiency. The difficulty of a synthetic example can be controlled by prompt complexity, required reasoning depth, number of constraints, or length of the required response. The curriculum can be static (pre-generated at fixed levels) or dynamic (generated based on the model current performance). Dynamic curriculum uses a zone of proximal development approach: generate examples the model can almost solve but not quite.