Fix training budget (19k steps / ~20B tokens), add Muon optimizer, correct variance scaling math, move figures to §Training Config, tighten abstract

This commit is contained in:
2026-07-08 22:49:43 +08:00
parent 4a143b056d
commit bfc8ff6098
2 changed files with 36 additions and 38 deletions
Binary file not shown.

Before

Width:  |  Height:  |  Size: 273 KiB

After

Width:  |  Height:  |  Size: 267 KiB