How each experiment idea flows from hypothesis โ final submission ยท hover nodes for detail
1รH100, 1 seed (42). Pass criteria: training doesn't crash, artifact file is under 16,000,000 bytes. Does not require bpb improvement โ just that the code works.
~$0.301รH100 full 600s run vs base. val_bpb must show visible improvement vs the copy base (1.1458). Even a small positive delta is enough to continue.
~$0.308รH100, 1 seed. val_bpb must be < 1.139 to continue to final 3-seed run. This is where single-GPU gains that don't scale get caught before spending $10.
~$3.508รH100, seeds 42, 1337, and 7. Mean val_bpb < 1.1378 (SOTA โ 0.005) with p < 0.01 variance across seeds. This is the final bar for submission.
~$10.50