specific method · filed under data curation
Knowledge distillation for synthetic data generation
Knowledge distillation trains a student model using signal from strong teacher models rather than only from external data.1 In the on-policy variant, the teacher signal is distilled onto the student's own trajectories, aligning training with what the student would actually produce at inference time.2
Trade-off On-policy distillation yields stronger gains than purely off-policy distillation.3
How this definition was written
Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.
This technique uses strong teacher models to guide training on the model's own generated attempts (on-policy rollouts
— nvidia/nemotron-3-ultra-550b-a55bBy distilling teacher signal onto the student's own trajectories rather than offline traces, MOPD better aligns the student's behavior with what it would actually produce at inference time
— nvidia/nemotron-3-ultra-550b-a55byielding stronger gains than purely off-policy distillation
— nvidia/nemotron-3-ultra-550b-a55b
Also called knowledge distillation, distillation.