Definition
Plain language
The stronger model whose worked answers are used to train a smaller one.
As stated in the literature
The source model in distillation, generating solution trajectories or token distributions that the student imitates; its ceiling generally bounds student benchmark performance.
Also called: teacher
Why it matters: Its quality sets the ceiling for what the smaller model can learn, so a weak teacher caps the whole effort no matter how well the student trains.
For example, a large frontier model writes out full worked solutions to thousands of math problems, and those transcripts become the training material for a smaller model.
Heard on the show
“You suspect a teacher is giving easier grades to students from their old college.”Episode 001 — When AI Models Quietly Protect Each Other From Shutdown