I just noticed that the models we trained are all ...
# proj-codescoutr
rich-byte-67407
02/12/2026, 11:27 AMI just noticed that the models we trained are all saved (by skyrl) in FP32 precision (due to skyrl's save setting by default)...
rich-byte-67407
02/13/2026, 1:41 AMYes! My concern wasn’t about the metrics — I noticed this because I was evaluating my own trained models and realized each 4B checkpoint sums up to ~17GB 😂
👍 1