Pretrain on B tasks, then finetune with one demonstration per instruction; test on the tasks held out from pretraining.
Takeaway: Sparse pretraining already learns the skills; one demonstration per instruction steers them to new combinations (0.4% → 54.7%).