Nanogpt Log Artifacts Analysis | Generated by AI
The raw log covers iter 27,800 → 74,950 (1,042 lines). Here’s what I see:
Content: Clean training output from iter 27800 to 74950
Lines: 1,042
Size: 53 KB
Start: iter 27800, loss 3.3310, MFU 113.01%
End: iter 74950, loss 3.2078, MFU 113.75%
Issues:
- Escape sequences at the end (arrow key artifacts from tmux scroll)
^[[A (up arrow) and ^[[B (down arrow) mixed into iter 74850 line
- iter 74850 line is split: "loss 3" on one line, ".2751..." on next
These are tmux capture artifacts, not training issues.
The clean version (nanogpt_760m_train_clean.log) filters these out. The raw log is fine for reference — the escape sequences are just from scrolling the tmux pane during capture.