Training quirk: reverse-chronological data ordering worsens loss

kalomaze · x · 2026-10-25

kalomaze shares a training log screenshot: the latest crop validation shows loss 1.4730 at 27.51M tokens, but arranging the data in reverse chronological order degrades it from 1.5124 to 1.5803 — suggesting unusual sensitivity to data ordering.

Original post →

More from Fun

Fun channel →