#artificial intelligence
May 2026
Always Learning, Always Mixing: Efficient and Simple Data Mixing All The Time
OP-Mix (On-Policy Mix), a data mixing algorithm that operates across pretraining, midtraining, and instruction tuning, is introduced, suggesting a different view of language model training: not a sequence of distinct phases, but a single continuous process of learning from data.
Michael Y Hu, Apurva Gandhi, Kyunghyun Cho et al.
· arXiv.org · 0 citations