A curve's shape narrows the family of causes; it does not name the bug.
Question: Which loss curve means which bug?
Synthetic Synthetic: six loss curves generated from named failure models by data/scripts/c20.s06.L1.py with seed 20261108. Frames show A falling from 2.2467 to 0.367, B flat near 2.2836, C's validation turning at 0.851 before climbing to 2.1765, D swinging between 0.8375 and 2.1639, E spiking to 2.2806 on step 47 with the gradient peak at 24.3449 on step 46, and F reaching 1299.3536 on step 40 before going missing at step 41.
Frame 1 of 12: Six shapesAll six runs on the same loss axis. A falls from 2.2467 to 0.367; B stays near 2.2836; C falls to 0.1469 while validation climbs to 2.1765; D ends at 1.2807 but ranges from 0.8375 to 2.1639; E peaks at 2.2806 on step 47; F reaches 1299.3536 at step 40 and is missing from step 41.
Six shapes. All six runs on the same loss axis. A falls from 2.2467 to 0.367; B stays near 2.2836; C falls to 0.1469 while validation climbs to 2.1765; D ends at 1.2807 but ranges from 0.8375 to 2.1639; E peaks at 2.2806 on step 47; F reaches 1299.3536 at step 40 and is missing from step 41.
Keyboard: Space plays or pauses, Left and Right step, Home goes to the first frame, End goes to the last frame.
Text description of this lab
Six loss curves, A to F, plotted against training step. A falls smoothly with validation just above it. B stays flat near 2.28. C's training loss falls while its validation loss turns and climbs. D swings up and down without settling. E has one narrow jump at step 47, then recovers. F runs off the top and becomes missing. A lower panel shows first- and last-layer gradient norms on a log scale, peaking at step 46 for E.