Perplexity-Aware Data Scaling Law: Perplexity Landscapes Predict Performance for Continual Pre-training
This work addresses inefficient data utilization in adapting foundation models to domain-specific applications, representing an incremental improvement over existing scaling laws.
The paper tackles the problem of diminishing returns from simply increasing data in continual pre-training by proposing a perplexity-aware data scaling law that predicts test loss based on the perplexity landscape of domain-specific data, achieving superior performance on medical and general-domain benchmarks.
Continual Pre-training (CPT) serves as a fundamental approach for adapting foundation models to domain-specific applications. Scaling laws for pre-training define a power-law relationship between dataset size and the test loss of an LLM. However, the marginal gains from simply increasing data for CPT diminish rapidly, yielding suboptimal data utilization and inefficient training. To address this challenge, we propose a novel perplexity-aware data scaling law to establish a predictive relationship between the perplexity landscape of domain-specific data and the test loss. Our approach leverages the perplexity derived from the pre-trained model on domain data as a proxy for estimating the knowledge gap, effectively quantifying the informational perplexity landscape of candidate training samples. By fitting this scaling law across diverse perplexity regimes, we enable adaptive selection of high-utility data subsets, prioritizing content that maximizes knowledge absorption while minimizing redundancy and noise. Extensive experiments demonstrate that our method consistently identifies near-optimal training subsets and achieves superior performance on both medical and general-domain benchmarks.