Set 3B pre-training budget to 60B tokens (stage1 + stage2) per model

[iamtatsuki05]

Why

The 3B pre-training budgets were epoch-based and uneven across models (EN 47B / 40B, JP 65B / 89B tokens). This PR fixes the budget to 60B tokens per model, stage1 + stage2 combined, mirroring how the 1B stage2 runs were started from a stage1 checkpoint that brought the total to ~20B.

Change

config	before	after	stage1 tokens	stage2	total
Llama-EN-3B/stage1.json	0.45 epoch	0.5473 epoch (13,700 steps)	57.5B	1 epoch, 2.5B	60.0B
Llama-JP-3B/stage1.json	0.8 epoch	0.7127 epoch (13,800 steps)	57.9B	6 epoch, 2.2B	60.1B
ModernBERT-EN-3B/stage1.json	0.45 epoch	0.65 epoch (13,893 steps)	57.3B	1 epoch, 2.45B	59.8B
ModernBERT-JP-3B/stage1.json	0.8 epoch	0.5214 epoch (13,800 steps)	57.7B	6 epoch, 2.2B	59.9B
Llama-EN-3B/stage2.json, ModernBERT-EN-3B/stage2.json	3 epochs	1 epoch (same as 1B)

Step counts assume total batch 4096 × seq 1024 in stage1; per-step token actuals (4.19M CLM, 4.13–4.18M MLM with padding) come from trainer_state.num_input_tokens_seen. Stage2 token counts are the 1B measurements (same data and total batch 512 × 8192).

JP stage2 (6 epochs) and the -bs16384 / -bs2048 legacy variants are untouched.

Notes

• Optimizer is schedule-free RAdam with constant LR, so running stage1 jobs resume under the new budget without any schedule discontinuity.
• Server-side job configs (cc25 mirei-1b3b-redo-20260725 use explicit max_steps 13,700 / 13,800; GCP mben3-s1 uses 0.65 epoch) are updated separately from this PR. ModernBERT-JP-3B stage1 already passed 13,800; its stage2 starts from checkpoint-13800.