Kosta Derpanis and Cody Blakeney endorse defining 'midtraining' as continued pretraining on high-quality, curated data

@gabriberton still learning on all tokens or is there masking? Or is that when you are officially into SFT?

@gabriberton I see midtraining as a hot fix so post-training RL actually works https://arxiv.org/abs/2512.04072

@gabriberton Or with longer context

Honestly this is the best definition

@cthorrez Still learning on all tokens

@nv_pavlichenko True, context extension is often a big part of midtraining

@gabriberton can it also be called mid-pretraining ?

@gabriberton The best definition

@gabriberton that framing actually makes a lot of sense, honestly.

@gabriberton 🙏

@nirmalpatel_ @gabriberton Final-pretraining would be a better choice.

@gabriberton @Dorialexander

@gabriberton MRI quality in 60 seconds?