Emergent Mind

Scaling Laws for Neural Language Models

Published Jan 23, 2020 in cs.LG and stat.ML


We study empirical scaling laws for language model performance on the cross-entropy loss. The loss scales as a power-law with model size, dataset size, and the amount of compute used for training, with some trends spanning more than seven orders of magnitude. Other architectural details such as network width or depth have minimal effects within a wide range. Simple equations govern the dependence of overfitting on model/dataset size and the dependence of training speed on model size. These relationships allow us to determine the optimal allocation of a fixed compute budget. Larger models are significantly more sample-efficient, such that optimally compute-efficient training involves training very large models on a relatively modest amount of data and stopping significantly before convergence.


  • The study empirically examines the influence of model size, dataset size, and computational power on the performance of language models using the Transformer architecture.

  • Performance shows a power-law relationship with model size, dataset size, and computational power; other factors such as architecture depth and width have less impact.

  • To prevent overfitting, model and data size need to scale in a specific ratio; efficiency is gained by understanding and applying early training trends.

  • A significant part of the paper is the strategy for optimal use of compute resources, favoring the training of large models with less data and early stopping.

  • The paper proposes a set of predictive equations for efficient language model training and highlights future advancements contingent on the strategic use of computational power.

Background and Methodology

The research presents an empirical investigation into the relationship between language modeling performance and multiple factors, including model size, dataset size, and training compute. The study harnesses the Transformer architecture, considering its aptness for tasks requiring different levels of performance. A fundamental observation is that performance enhancements typically follow power-law trends relative to each of these factors, provided no bottleneck from the others. Crucially, these relationships span across more than seven orders of magnitude, indicating robust patterns even at varying scales.

Key Findings

The study provides several compelling findings:

  • Models' performance is predominantly dictated by scale—number of parameters (N), dataset size (D), and training compute (C)—and only weakly by other hyperparameters such as architecture depth and width.
  • A smooth power-law relationship is observed with individual scale factors when the other two are not limiting.
  • To avoid overfitting penalties, data and model size must scale synergistically. Notably, when model size is increased by a factor of eight, data needs only a five-fold increase.
  • Training efficiency reflects predictable power-law trends, virtually independent of model size. Models can be trained faster, with fewer data points, by extrapolating early trends.

Furthermore, models with different distributions generalize better as performance on the training set improves, and large models are inherently more sample-efficient.

Compute Budget Optimization

A key contribution of the paper is the guidance on optimal allocation of a fixed compute budget. Efficacy here invloves training extremely large models on modest data amounts and stopping well before full convergence. These large models are significantly more sample-efficient, requiring fewer optimization steps, which can be counterintuitive when conventionally smaller models are trained to completion. With growing computational budgets, the primary focus should be on upscaling model size, with modest boosts in dataset size and only marginal increases in serial training time.

Predictive Framework and Implications

The researchers provide equations portraying the empirical relationships discovered, akin to a "statistical mechanics" for language models. These laws predict how the optimal model size, batch size, steps taken, and required dataset size should scale with a given computational budget, shedding light on how advancements in language modeling are likely to evolve as resources expand.

Potential Limitations

The paper discusses potential caveats, such as the lack of a solid theoretical underpinning for the empirical scaling laws and questions on how generalizable these trends are across different domains or types of models. The study’s predictions are also not verified in the extremely large data or model size regime, thus leaving some uncertainties on their long-term applicability.


In essence, this research contributes significantly to understanding how language models scale and provides practical recommendations for efficient training. It suggests that future improvements in AI language understanding are not just tied to the availability of data, but also hinge critically on the strategic deployment of computational resources and model design. The findings indicate a path where larger, more computationally demanding models, if trained judiciously, could yield substantial gains in performance.

Get summaries of trending AI papers delivered straight to your inbox

Unsubscribe anytime.