Core question
How can explicit rules influence training without replacing the neural model itself?
Project 01
logic_net_toy
A small-scale implementation used to examine how logic rules can be converted into a teacher distribution and subsequently distilled into student parameters.
Suggested code order
Start with the training loop, then inspect the rule module, and finally trace how the teacher distribution is constructed and reused.
Mechanism to watch
A key mechanism is that rules do not overwrite labels directly; they first shape a teacher distribution, which is then used to regularize the student through distillation.