Revisiting the Scaling Properties of Downstream Metrics in Large Language Model Training
This work addresses the challenge of reliably scaling LLMs for downstream tasks, providing a direct framework that improves prediction accuracy, though it is incremental in refining existing scaling approaches.
The paper tackles the problem of predicting downstream task performance from training budget in large language models, finding that a simple power law can accurately describe scaling behavior and extrapolate better than previous methods.
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.