tokens 1732003925 shards 12 batches 16819 window 65536 grad_accum 8 usable_steps 2102 mix b5_65536_131072 target 0.5000 actual 0.4999 tokens 865840569 mix b3_16384_32768 target 0.3000 actual 0.3001 tokens 519688950 mix b1_4096_8192 target 0.2000 actual 0.2000 tokens 346474406