Review of: "Metadata Conditioning Accelerates Language Model Pre-training"
Weiqiang Jin · 2025
This paper introduces an innovative pre-training method called Metadata Conditioning then Cooldown (MeCo), which incorporates metadata into model training, signi cantly enhancing data e ciency while improving the controllability of language models.However, I still have some doubts and suggestions regarding the article, and I hope the author can improve the paper based on my thoughts: