Model Training Progress Update | Generated by AI
Progress (23:05 HKT, +6h42m since your 16:23 snapshot)
Current state:
- iter ~3,420 / 19,073 — 17.9% done (was 6.6% at 16:23)
- last val loss 3.4043 at iter 3000 (down from 4.04 at iter 1000)
- recent train losses bouncing in 3.2–3.5 band (was 3.7–3.9 last check)
- step time still ~11.11 s/iter, MFU 12.83%
- 6 checkpoints saved (500 → 3000); next save at iter 3500 in ~16 min
- no OOMs, no errors, no slowdown
Val loss trajectory — healthy, slope flattening as expected:
| iter | val loss |
|---|---|
| 1000 | 4.04 |
| 1500 | 3.77 |
| 2000 | 3.57 |
| 2500 | 3.47 |
| 3000 | 3.40 |
Time remaining from now:
| Stop at | Iters left | Wall time | ETA (HKT) | Expected loss |
|---|---|---|---|---|
| 5,000 | ~1,580 | ~4.9 hrs | ~04:00 Wed | ~3.2–3.3 |
| 10,000 | ~6,580 | ~20.3 hrs | ~19:30 Wed | ~3.0–3.1 |
| 19,073 (full) | ~15,650 | ~48.3 hrs ≈ 2.0 days | ~23:00 Thu | ~2.9–3.0 |
We’ve shaved roughly 7 hours off the full-run ETA since 16:23. Run is on rails — pace is steady, loss is decreasing monotonically at the eval boundaries, and you can still kill at any 500-iter mark without losing the latest checkpoint.