adamw-PT-adamw-MT-a10-g20-mdr
The PT checkpoint after 10,000 steps (20.97B tokens) of mid-training with FineMath-4plus: MT budget 10%, Two-Stage data schedule with final MT ratio 20% (MT data enters at step 5,000), constant LR 5e-4 until step 5,000, then linear decay to 1e-5 (MDR). AdamW throughout. Subset (32M-token) losses at step 10,000: DCLM 2.7527, FineMath 1.6171. Reproduction of arXiv 2610.29348 with Megatron-LM.