SFT Training Progress and Metrics | Generated by AI
SFT running: Step 305 / full epoch (3.4%)
| Metric | Value |
|---|---|
| Loss | ~1.7–1.8 (down from 2.4 at step 7) |
| Speed | ~50,000 tok/s, ~1.3s/step |
| Runtime | 6.4 min elapsed |
| Progress | 3.4% of dataset (3.3M tokens seen) |
| LR | 1.00 (warmup done, at max) |
| GPU | 100%, 8.7GB, 64°C |
| Epoch | 1 |
Loss is trending down. The full epoch will cover ~1M conversations, so there’s a long way to go. It’s saving checkpoints every 500 steps — first eval checkpoint at step 500.
Attach with tmux attach -t 0.