Definition
Midtraining is the training stage between pretraining and final fine-tuning, where a model is exposed to high-quality, often domain-targeted data to sharpen specific capabilities. Labs increasingly talk about midtraining as a distinct phase with its own data mix and learning rate schedule.