muon-PT-muon-MT-a10-g20-mdr
The Muon PT checkpoint after 10,000 steps (20.97B tokens) of midtraining with Muon on DCLM and FineMath-4plus. FineMath accounts for 10% of the token budget. The final 5,000 steps use 20% FineMath. The learning rate stays at 5e-4 for 5,000 steps, then decays linearly to 1e-5.
Final 32M-token subset losses were 2.7379 on DCLM and 1.6033 on FineMath. Reproduction of arXiv 2610.29348 with Megatron-LM.