Tag
The author trained 1B, 2B, and 3B models with the same SFT recipe and observed that instruction-following (IFEval) regressed for the 1B and 2B models but improved for the 3B, possibly due to different learning rates or model capacity.