Not only that, they additionally ran an experiment with the training temperature turned way up (2.0) and truncation turned off such that the majority of SFT examples were incoherent (63% IIRC). Yet the model finetuned on these broken examples still improved over baseline.