AutoResearch

phase: training

On Spark (~13,500 steps in 2h), warmdown_iters=3000 means 22% of training is spent decaying LR vs 15% on 8xH100 (20K iters). Reducing to 2000 restores the H100-proportional warmdown and gives more full-LR training time. Previous experiments only tried increasing warmdown (3600, 6000) and both failed; this is the first decrease.