Training Progress and Time Estimate | Generated by AI
Current progress
- You’ve just passed iter 2800 (out of your target 20,000)
- Stable speed right now: ≈ 3.91 seconds per iteration
(the occasional 19-second logs at eval steps are only every 500 iters and don’t change the average much)
Remaining iterations
20,000 − 2,800 = 17,200 iterations left
Estimated time to finish
| Average iter time | Remaining time | When it will finish (from now, Nov 19 2025) |
|---|---|---|
| 3.91 s (current) | 18 hours 35 minutes | ~ tomorrow (Nov 20) around 14:30–15:30 (depending on your timezone and exact start time of this log) |
| Slightly faster (3.8 s, common as MFU climbs a bit more) | ~ 18 hours | tomorrow early–mid afternoon |
| Slightly slower (4.0 s) | ~ 19 hours 5 min | tomorrow ~16:00 |
Realistic expectation on an RTX 4070 with this exact setup: 18–19 hours from now.
Quick reference
- If you started logging this around noon–early afternoon today → finishes tomorrow morning–early afternoon.
- If you let it run overnight → you’ll wake up to a finished (or almost finished) very strong 40M model tomorrow.
Your loss is dropping nicely (val loss already 3.94 at only ~1.5B tokens seen) and still falling steadily. Absolutely worth letting it run the full 20k — just leave it going overnight and you’ll have an excellent model by tomorrow afternoon!