Paper Reveals Larger LLMs Tolerate More Data Repetition During Pretraining
heghbalz · x · 2026-08-22
This paper investigates the strategy of repeating high-quality domain data to maintain a fixed tokens-per-parameter (TPP) ratio during LLM pretraining when such data is scarce. Key findings include:
- Scale Increases Tolerance: Surprisingly, at a fixed TPP, the optimal repetition count mildly increases with model size. Larger models with shorter learning rate decay tolerate repetition better.
- Quality Dictates Limits: Optimal repetition counts across domains are strongly negatively correlated with final validation loss—higher quality data (lower loss) benefits more from repetition.
- Proxy Tuning Works: Repetition counts tuned on smaller proxy models with the same TPP can provide practical estimates for larger models.
This offers a new engineering perspective on mitigating high-quality data dilution.
Related event: ICML Paper: Larger LLMs Tolerate More Data Repetition in Pretraining(2 posts)→
More from Infra
- LLM Prompting Wastes Computation; Reuse Potential is Huge — miniapeur · 2026-08-22
- SGLang's Weight Cache Daemon Cuts 1T Model Restart Time from 8.8 min to 32 sec — xiaosun86 · 2026-08-22
- Agent recursive loops blow up context costs: 5% failures eat 25% of bill — MaverikSh · 2026-08-22
- llama.cpp ships version 0.2.0 with official release notes — PhilippeEiffel · 2026-08-22
- Open Source Tool Mark Cleaner Locally Removes AI Text Watermarks and Metadata — VraserX · 2026-08-22
- Marin releases 23T-token pretraining dataset for public download — joecole · 2026-08-22