Really cool work! Does the training data budget scale with model size? How would you compare the Gemma 4 draft model which is also integrated with the base kv

Really cool work! Does the training data budget scale with model size? How would you compare the Gemma 4 draft model which is also integrated with the base kv cache?