Distillable Models And Synthetic Data Pipelines With Nemo Data Designer

Generating training examples with another model requires checking whether its outputs may be reused for training. OpenRouter addresses that selection problem with a catalog of models it identifies as permitting synthetic training data and a request setting that restricts generation accordingly. Its workflow announcement pairs that filter with NVIDIA's NeMo Data Designer, which defines data generators and output structures in code. The announcement's notebook walks through the sequence: selecting distillable models, generating structured synthetic data, and preparing the dataset for distillation and fine-tuning.