Progress Update and Loss Plateau | Generated by AI
Update from last check (25,600 → 35,436):
Progress:
- Step 35,436 / 87,000 – 40.7% (up from 29.4%)
- Elapsed: 736 min (~12.3h) – added ~3.5h
- ETA: 1,072 min (~17.9h) remaining
- Gained ~9,836 steps since last check
Loss:
- Still bouncing in 1.25 - 1.39 range
- No sustained improvement trend – variance dominates, same as before
- lrm dropping to 0.92 (cosine LR schedule kicking in)
Performance:
- dt ~1.32s/step, ~49K tok/s – slightly slower than before (was ~1.23s)
- bf16_mfu ~148-152% – MFU calc still buggy
- pq: 20 rg: 1 – data pipeline stable
Takeaway: ~18h to go. Loss hasn’t meaningfully dropped from the 1.25-1.45 band seen at 29%. The LR decay just started (0.92), so the next ~30K steps where LR drops to ~0 will tell us if this model can compress further or if it’s plateaued.
Generated by AI. Curating and sharing still takes effort.
If you find it useful, feel free to
donate.
WeChat: @lzwjava
·
X: @lzwjava ·
Say hi 👋