Emergent Mind


Uncovering early-stage metrics that reflect final model performance is one core principle for large-scale pretraining. The existing scaling law demonstrates the power-law correlation between pretraining loss and training flops, which serves as an important indicator of the current training state for large language models. However, this principle only focuses on the model's compression properties on the training data, resulting in an inconsistency with the ability improvements on the downstream tasks. Some follow-up works attempted to extend the scaling-law to more complex metrics (such as hyperparameters), but still lacked a comprehensive analysis of the dynamic differences among various capabilities during pretraining. To address the aforementioned limitations, this paper undertakes a comprehensive comparison of model capabilities at various pretraining intermediate checkpoints. Through this analysis, we confirm that specific downstream metrics exhibit similar training dynamics across models of different sizes, up to 67 billion parameters. In addition to our core findings, we've reproduced Amber and OpenLLaMA, releasing their intermediate checkpoints. This initiative offers valuable resources to the research community and facilitates the verification and exploration of LLM pretraining by open-source researchers. Besides, we provide empirical summaries, including performance comparisons of different models and capabilities, and tuition of key metrics for different training phases. Based on these findings, we provide a more user-friendly strategy for evaluating the optimization state, offering guidance for establishing a stable pretraining process.
DeepSeek-67B MMLU model architecture for mastering multi-label understanding tasks.


  • The paper explores the relationship between pretraining LLMs and their performance on downstream tasks, focusing on how model size and pretraining methods affect outcomes.

  • It conducts a detailed analysis of models with 7 to 67 billion parameters across various tasks, identifying key factors that predict performance within and across domains.

  • The research revisits scaling laws, highlighting the nuanced roles of dataset size, model architecture, and computational budget in improving LLM efficiency.

  • Future implications are discussed, including the release of model checkpoints to aid in further research, with findings aimed at enhancing the development of more sophisticated LLMs.

Introduction to the Core Proposition and Analysis

This paper presents an examination of the dynamics between large language model (LLM) pretraining and their capabilities on downstream tasks. A particular focus is given to understanding how variations in model size and pretraining methodologies contribute to performance across a range of predefined benchmarks. The authors specifically address the need for a clearer connection between the extensive computational resources spent during the pretraining phase and the subsequent performance improvements on downstream tasks.

Comprehensive Examination of Model Capabilities

The research outlines a systematic analysis of intermediate checkpoints for several state-of-the-art LLMs, ranging from 7 to 67 billion parameters. The team meticulously explores the performance implications of differing pretraining schemes, sizes, and the resultant model capabilities across both domain-specific and domain-agnostic tasks. The core findings of this paper entail:

  • Cross-task Predictability: It is observed that a model's performance dynamics on known tasks within a domain can accurately predict its performance on unseen tasks within the same domain.

  • Cross-domain Learning Efficiency: Insights drawn from one domain, through a curriculum learning approach, seemingly aid in enhancing performances on unrelated domains, which mirrors the cognitive learning patterns in humans.

  • Impacts of Training Strategies and Architectures: An in-depth analysis on 7b-scale models demonstrates the significant impacts of training datasets, learning rate adjustments, batch sizes, and regularization techniques in the early training stage.

  • Scaling Laws Re-evaluated: The research provides empirical data on scaling laws, particularly elucidating how model performance scales with larger datasets, the nuanced effects of model architecture, and how computational budget allocations influence training efficacy.

Scaling Law Analysis

The seminal concept of the Scaling Law, which suggests that performance enhancements can be anticipated with increases in computational budget, model size, and data size, is examined against the backdrop of LLMs. This analysis underlines how different models exhibit varied behaviors in alignment with the scaling law, shedding light on more intricate factors influencing scaling efficiency beyond mere data quantity and model parameters.

Future Implications and Open Resources

The work not only contributes to the theoretical understanding of LLM training dynamics but also has practical implications for the design and optimization of future models. Particularly noteworthy is the team's initiative to release intermediate checkpoints of Amber-7B and OpenLLaMA-7B, aiming to foster further research exploration and facilitate a more comprehensive comprehension of the pretraining process.


The findings illuminate the complex landscape of LLM pretraining, highlighting the intricate interplay between model size, architecture, pretraining strategies, and their influences on downstream tasks. This paper enriches the field with empirical evidence and methodological insights, which are poised to guide the development of more efficient, robust, and capable LLMs in the future. The authors emphasize the multifaceted nature of model scaling and performance optimization, challenging the community to consider beyond traditional metrics and approaches.

Get summaries of trending AI papers delivered straight to your inbox

Unsubscribe anytime.

Test Your Knowledge

You answered out of questions correctly.

Well done!