Training Progress and Time Estimates | Generated by AI
Current state (16:23 HKT):
- iter 1,260 / 19,073 — about 6.6% done
- loss dropped 11.0 → 3.75 (val 4.04 at iter 1000)
- step time steady at ~11.1 s/iter, MFU ~12.7%
- 2 checkpoints saved (500, 1000); no errors, no OOMs
Time remaining (from now):
| Stop at | Iters left | Wall time from now | Expected loss |
|---|---|---|---|
| 5,000 | ~3,740 | ~11.5 hrs (done early tmrw morning) | ~3.5–3.8 |
| 10,000 | ~8,740 | ~27 hrs (done Wed evening) | ~3.2–3.4 |
| 19,073 (full) | ~17,813 | ~55 hrs ≈ 2.3 days (done ~Thu late night) | ~2.9–3.1 |
So pacing matches the original estimate — we’ve burned ~4 hrs, ~55 hrs to go for the full run. Pace and loss trajectory both look healthy; you can kill at any 500-iter boundary without losing the latest checkpoint.