d12 SFT v3 (fixed inline python)

SFT on d12 pretrain checkpoint

Pretrain source: d12 H100 SSSL final (val_bpb=0.854, step 2205)

Plan:

1. Download pretrain checkpoint from MLflow artifacts (~2GB)
2. Train tokenizer (deterministic, ~80s)
3. Run SFT with default mixture (~1.07M rows)
4. Upload SFT checkpoint to MLflow artifacts for downstream RL
5. num_iterations=9999 as cap; epoch-based stopping triggers first (~600 steps)

Prior attempts:

• v1: cancelled (missing --upload-checkpoint)
• v2: dryrun_failed (IndentationError in inline Python — YAML folding issue)

Status: cancelled — dryrun passed but num_iterations=9999 caused inflated cost estimate ($186). Epoch is ~612 steps. Recreating with num_iterations=700.