Skip to content
X Β· @teortaxesTex Β· X / Twitter

RT 𝚐π”ͺ𝟾𝚑𝚑𝟾: The best domain mix may not stay fixed across pretraining. RegMix trains proxy models, then selects one mixture from endp…

RT 𝚐π”ͺ𝟾𝚑𝚑𝟾The best domain mix may not stay fixed across pretraining.RegMix trains proxy models, then selects one mixture from endpoint loss. REGMIX-D uses the full proxy loss trajectory instead: current step, current mixture, and current loss predict the next-interval loss.REGMIX-D makes mixture selection conditional on