This paper proposes a task-driven framework for digital training of university teachers, focusing on task parsing, multi-dimensional interaction modeling and adaptive resource scheduling. Relying on the actual deployment platform, a structured data set containing 214 teachers, 638 training sessions, 96420 behavior records and 12860 feedback records is constructed. The system generates a task graph according to teaching objectives, participation trajectories and evaluation logs, and combines time series aggregation, relation-aware state coding and interaction strategy fusion to complete the unified representation. On this basis, the system further completes the feedback timing calculation, resource matching and training path generation. Experimental results show that the task recognition accuracy of the framework reaches 92.8%, the interactive decision accuracy reaches 90.6%, the path matching consistency reaches 88.9%, the average response delay is 74 ms, and the content adaptation rate reaches 93.4%. It maintains a stable running state and process control ability in different training scenarios. The framework has good computing stability, deployment adaptation ability and scalability in the digital training environment of university teachers.
