Program synthesis is part of the loss function, which is what makes it a auxiliary learning task. We haven’t experimented with model size yet, we just used the

Program synthesis is part of the loss function, which is what makes it a auxiliary learning task. We haven’t experimented with model size yet, we just used the same configuration as the smallest Code Llama. We did play with dataset size and found thah performance tracks the usual scaling laws. Details in the paper