Revisiting the Scaling Properties of Downstream Metrics in Large Language Model Training

TL;DR

This paper demonstrates that downstream task performance in large language models can be accurately modeled and predicted using simple power laws based on training data, improving extrapolation over previous methods.

Contribution

It introduces a direct modeling framework for downstream performance scaling laws, outperforming traditional two-stage procedures and providing functional forms for broader predictions.

Findings

Power law accurately models log accuracy scaling

Direct approach outperforms two-stage procedures

Validated on models up to 17B parameters and 350B tokens

Abstract

While scaling laws for Large Language Models (LLMs) traditionally focus on proxy metrics like pretraining loss, predicting downstream task performance has been considered unreliable. This paper challenges that view by proposing a direct framework to model the scaling of benchmark performance from the training budget. We find that for a fixed token-to-parameter ratio, a simple power law can accurately describe the scaling behavior of log accuracy on multiple popular downstream tasks. Our results show that the direct approach extrapolates better than the previously proposed two-stage procedure, which is prone to compounding errors. Furthermore, we introduce functional forms that predict accuracy across token-to-parameter ratios and account for inference compute under repeated sampling. We validate our findings on models with up to 17B parameters trained on up to 350B tokens across two dataset mixtures. To support reproducibility and encourage future research, we release the complete set of pretraining losses and downstream evaluation results.