OpenEuroLLM Scaling Laws: Optimizing Learning Rates and Batch Sizes

Niccol\`o Ajroldi, Diana Alexandra Onutu, Haider Al-Tahan, J\"org Franke, Sampo Pyysalo, Jenia Jitsev, Aaron Klein· August 31, 2026 View original

Key takeaways

  • Optimal learning rates and batch sizes scale predictably with model capacity and data.
  • Learning rate annealing schedules significantly impact training efficiency.
  • Loss dependence on model and data size can be effectively modeled to predict performance.
  • The study provides a foundational scaling methodology for LLM development.

Who benefits

AI/ML DevelopmentCloud ComputingResearch & DevelopmentSoftware Engineering

Summary

This research investigates the scaling behavior of learning rates and batch sizes in pretraining large language models, developing a model to capture their relationships and evaluating learning rate schedules. It also characterizes the dependence of loss on model capacity and dataset size, establishing a baseline for future OpenEuroLLM development.

This study delves into the fundamental scaling properties of large language models during their pretraining phase, specifically focusing on how learning rates and batch sizes interact with model capacity and data scale. Researchers developed a comprehensive model to describe these relationships, examining both jointly optimal and marginal evolutions of these critical hyperparameters. The work also explored the benefits of different learning rate annealing schedules, such as Warmup-Stable-Decay, and whether optimal settings transfer across different training phases. Furthermore, the paper analyzes how the training loss is influenced by the model's size and the volume of the dataset. It evaluates recent scaling forms that explicitly model these interactions, finding them effective in capturing both undertraining and overtraining scenarios. This research provides a crucial foundational baseline and a systematic scaling methodology for the ongoing development of OpenEuroLLM models, with all pretraining run data made publicly available.

Why it matters

Understanding these scaling laws is crucial for efficiently training large language models, enabling practitioners to optimize computational resources and achieve better model performance.

How to implement this in your domain

  1. 1Review the open-sourced pretraining runs to inform your own LLM training strategies.
  2. 2Experiment with different learning rate schedules and batch sizes based on the derived scaling models.
  3. 3Apply the insights on loss dependence to predict training outcomes and resource requirements for new LLM projects.
  4. 4Consider adopting the proposed scaling procedure for developing custom LLMs within your organization.

Original post by Niccol\`o Ajroldi, Diana Alexandra Onutu, Haider Al-Tahan, J\"org Franke, Sampo Pyysalo, Jenia Jitsev, Aaron Klein

"arXiv:2608.28308v1 Announce Type: new Abstract: We study the scaling behavior of learning rate and batch size in pretraining dense large language models on English-prevalent corpora. Beyond scaling \textit{jointly optimal} learning rates and batch sizes, we investigate their \tex…"

View on X

Originally posted by Niccol\`o Ajroldi, Diana Alexandra Onutu, Haider Al-Tahan, J\"org Franke, Sampo Pyysalo, Jenia Jitsev, Aaron Klein on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses

More in AI Research

AI ResearchAI Engineering & DevTools

New Optimizer Accelerates LLM Pretraining with Curvature-Conditioned Momentum

This research proposes a curvature-conditioned multiscale momentum method with sphere constraints to accelerate large language model pretraining. It addresses challenges from noise-dominant gradients and ill-conditioned loss landscapes by enhancing progress along flat directions, significantly improving upon existing adaptive optimizers like AdamW and Muon.

Shuchen Zhu, Yuxin Fang, Mingze Wang, Kun YuanAug 31, 2026
AI ResearchAI Engineering & DevTools

Euclidean Fourier Neural Operators Enhance Domain Transferability

This paper introduces Euclidean Fourier Neural Operators (EFNOs) as a domain-independent alternative to traditional FNOs, addressing their limitation in transferring across different periodic domains. EFNOs achieve this by parameterizing the spectral kernel as a continuous function of the physical wavevector, enabling consistent operator learning across varying domain shapes and sizes.

Nathanael Bosch, Niklas Frederik Schmitz, Michael F. HerbstAug 31, 2026
AI Engineering & DevToolsAI Research

SymboLLM-FE Boosts Feature Engineering with LLMs and Symbolic Regression

This paper introduces SymboLLM-FE, a novel approach combining symbolic regression and large language models for automated feature engineering on tabular data. It aims to generate highly interpretable and performant features while overcoming the limitations of traditional AutoFE and LLM-based methods.

Zi-Jian Cheng, Zi-Yi Jia, Zhi Zhou, Yu-Feng Li, Lan-Zhe GuoAug 31, 2026