← All questions
Training & Fine-tuningMedium· Distillation· Synthetic Data

Is it circular to use a language model to generate trai...

The process is not circular provided that the rewrite model is significantly stronger than the model being trained. The quality ceiling of the resulting model is determined by the rewrite model’s quality. You are essentially performing distillation from a superior teacher. Self-distillation without an external, stronger teacher can lead to model collapse over repeated iterations, which is why production pipelines rely on a dedicated, larger rewrite model rather than the checkpoint currently being trained.

Code playground (Python)