Generalization fluctuations remain when the same task results are plotted against pretraining FLOPs instead of token counts, including the full persona evaluations for both model families.

Source: paper:§4.2, p.6; Appendix H.2, Figures 31–37, pp.22–25