ByteDance Studies Domain Data Repetition in LLM Pretraining

ByteDance-Seed · hf · 2026-08-17

ByteDance research investigates optimal repetition of high-quality domain data in LLM pretraining under proportional scaling of model size and tokens. Findings show optimal repetition increases mildly with scale and correlates more with domain validation loss than unique data volume.

Original post →

More from Research

Research channel →