The released checkpoint scores 46.1 on Terminal-Bench 2.1, +3.1 over base gemma-4-31B-it.
One checkpoint is one draw, so we trained the same recipe 5 times: mean +2.1, all 5 above base. Each run takes ~7 hours on a single 8xH200 node.
One checkpoint is one draw, so we trained the same recipe 5 times: mean +2.1, all 5 above base. Each run takes ~7 hours on a single 8xH200 node.
2 0 10
