| # Training Scripts (Execution Order) | |
| 1. smoke_branch_exercise.sh - 60 opt-steps, exercise ALL stage branches | |
| 2. smoke_production_thresholds.sh - 200 opt-steps, real scaled thresholds | |
| 3. sft_stage1.sh - Full Stage1 (6000 opt-steps) | |
| 4. sft_phase2.sh - Full Phase2 (10000 opt-steps, init from expert) | |
| Key config (LD-8 lessons): | |
| - dataloader_num_workers=0 | |
| - per_device_train_batch_size=1 | |
| - gradient_accumulation_steps=8 (effective batch=8) | |
| - bf16=True, DeepSpeed ZeRO-2 | |
| - image_min/max_pixels=256/1024 * 28*28 | |
| - image_resized_w/h=448 | |
| ALL except smoke_branch_exercise.sh require clock_fix_THRESHOLD_SCALE applied | |
| (see patches/README.md). | |